fix(gl): thread caller-identity repo resolution through MCP webhook tools (#159 )

Merge remote-tracking branch 'origin/main' into pr-159
ci: run PR checks and CodeQL only for the surfaces a PR touches (#519 )
2026-11-10 05:01:48 -10:00 · 2026-11-10 08:42:58 -06:00 · 2026-11-10 08:38:53 -06:00 · 2026-11-10 00:41:03 +00:00 · 2026-11-09 23:40:38 +00:00 · 2026-11-09 23:02:05 +00:00
181 changed files with 128187 additions and 4971 deletions
@@ -8,24 +8,9 @@
 # The audit job stays hard-fail: only the advisories explicitly listed below are
 # suppressed, so any NEW advisory still breaks CI. Each entry below has no
 # available upstream fix; everything fixable is fixed in Cargo.lock instead.
-#
-# Tracking: the hickory ignores are reachable accepted risk. A scheduled re-check
-# (.github/workflows/audit-schedule.yml) runs without these ignores and fails if
-# the lockfile moves off the pinned vulnerable versions while the ignores remain.

 [advisories]
 ignore = [
-    # hickory-proto 0.25.2 (pulled by libp2p-dns 0.44 -> hickory-resolver 0.25).
-    # REACHABLE: the node wraps its QUIC transport in the system DNS resolver
-    # (crates/gitlawb-node/src/p2p/mod.rs:238), so a malicious DNS response can
-    # trigger these DoS paths. No fix is available: the latest libp2p-dns on
-    # crates.io is 0.44 and pins hickory 0.25; the fix requires hickory >=0.26.1.
-    # REMOVE both once libp2p-dns ships a release using hickory 0.26 (verified
-    # 2026-06-21: cargo update -p hickory-proto locks 0 packages; still 0.25.2).
-    # Tracking: https://github.com/Gitlawb/node/issues/76
-    "RUSTSEC-2026-0118", # hickory NSEC3 closest-encloser unbounded loop
-    "RUSTSEC-2026-0119", # hickory O(n^2) name-compression CPU exhaustion
-
    # rsa 0.9.10 (Marvin timing sidechannel). Present in Cargo.lock because sqlx
    # resolves sqlx-mysql unconditionally, and sqlx-mysql depends on rsa. cargo
    # audit scans the lockfile, so it flags rsa even though it is NOT linked into
@@ -35,4 +20,54 @@ ignore = [
    # ever built with the `mysql` feature, or any other consumer of rsa enters
    # the build, at which point it becomes a real reachable advisory.
    "RUSTSEC-2023-0071", # rsa Marvin attack (no fix; not linked in our build)
+
+    # lru 0.16.3 (use-after-free in pop()). Reachable via alloy. The
+    # lockfile holds only 0.16.3 — the aws-sdk-s3 path that once pulled
+    # 0.12.5 no longer does. The RUSTSEC-2026-0253 advisory covers this
+    # range, so the ignore stays load-bearing until a fixed 0.16.x lands;
+    # REMOVE when the locked version is fixed or lru leaves the tree.
+    # Retirement owner: node maintainers, enforced by the weekly
+    # audit-schedule drift guard ("Fail if lru moved while its ignore
+    # remains"), which fails if either locked version moves or lru
+    # leaves the tree. No separate tracking issue exists; file one
+    # against the lru upgrade if this waiver needs discussion outside
+    # the schedule job.
+    "RUSTSEC-2026-0253",
+    #
+    # Round-3 P2 (reviewer 2): the previous version of this file
+    # ignored RUSTSEC-2026-0258 for h2 0.4.13, but the lockfile
+    # already has h2 0.4.18 (verified 2026-08-31: cargo tree -p h2
+    # shows 0.4.18) — the fix is in. The ignore was stale and would
+    # have silently re-accepted the advisory if h2 dropped back to
+    # 0.4.13. Removed.
+    # crypto-bigint 0.5.5 (secret-dependent branches in operations meant
+    # to be constant-time). REACHABLE accepted risk, on two paths:
+    # - alloy-signer -> elliptic-curve 0.13 -> crypto-bigint 0.5: the k256
+    #   PrivateKeySigner that signs transactions in `gl name`
+    #   (crates/gl/src/name.rs) and the node operator
+    #   (crates/gitlawb-node/src/operator.rs).
+    # - aws-sigv4 (`sigv4a` feature, crates/gitlawb-node/Cargo.toml) ->
+    #   crypto-bigint 0.5 (directly and via p256 0.13): SigV4a key
+    #   derivation, used only for requests to endpoints that require SigV4a
+    #   (e.g. S3 Multi-Region Access Points).
+    # Exploiting it needs local timing observation of those signing operations.
+    # The fix is crypto-bigint >=0.7.6, and no compatible dependent uses it:
+    # every alloy-signer release (through 2.5.0) requires elliptic-curve ^0.13,
+    # and aws-sigv4 1.6.4 is the first release on crypto-bigint 0.7 (1.5.x -
+    # 1.6.3 still require ^0.5.4), but it needs Rust 1.94.1, above our 1.91 MSRV
+    # (verified 2026-11-03). REMOVE only when Cargo.lock has no crypto-bigint
+    # below 0.7.6: `grep -A1 'name = "crypto-bigint"' Cargo.lock`.
+    "RUSTSEC-2026-0387", # crypto-bigint timing variability (no compatible fix)
+
+    # libsqlite3-sys 0.30.1 (bundled SQLite WAL uninitialized-memory disclosure,
+    # CVE-2026-119375). Present in Cargo.lock because sqlx resolves sqlx-sqlite
+    # unconditionally; sqlx is built postgres-only
+    # (crates/gitlawb-node/Cargo.toml), so sqlx-sqlite and its bundled SQLite C
+    # source are never compiled or linked (`cargo tree -i sqlx-sqlite` prints
+    # nothing) and there is no reachable WAL open to attack. The fix requires
+    # libsqlite3-sys >= 0.39, but sqlx-sqlite 0.8.6 pins ^0.30.1; a compatible
+    # bump needs a sqlx upgrade, not a lockfile refresh. REMOVE this ignore if
+    # sqlx is ever built with the `sqlite` feature or any other libsqlite3-sys
+    # consumer enters the build.
+    "RUSTSEC-2026-0459", # libsqlite3-sys bundled SQLite CVE (not linked in our build)
 ]
@@ -1,15 +1,50 @@
-# gitlawb node — environment variables
+# Twigpine node — environment variables
 # Copy to .env and fill in your values.
 # All variables are optional unless marked REQUIRED.

 # ── Node identity ─────────────────────────────────────────────────────────
-# Path to the node's Ed25519 keypair PEM file.
+# Path to the node's Ed25519 keypair PEM file. Must name a key file, not a
+# directory: a trailing slash, a final `.` or `..`, and `..` traversal are
+# refused so the path cannot retarget a parent. A `~/` value must stay under
+# home after expansion. Bare `identity.pem` is still legal.
 # Generate with: gl identity new
 GITLAWB_KEY=/data/keys/identity.pem

+# Path to the node's persistent libp2p identity key file. Must include a
+# directory, because the node will not keep its p2p identity key in the working
+# directory. A value that cannot name a securable key file (a bare filename,
+# `..` traversal, a trailing slash, the filesystem root, or a `~/` path that
+# escapes home) is refused before the node binds, so the process exits.
+# Anything wrong with the storage itself (a symlinked or non-directory parent,
+# an unsafe cwd or ancestor, an unreadable or malformed key) leaves p2p off
+# while HTTP keeps serving, logged at error as p2p_identity_key_load_failed and
+# counted by the gitlawb_p2p_identity_key_load_failed metric. /health still
+# reports healthy in that state, so alert on the event or the metric. Fixing
+# the storage and restarting restores the same PeerId.
+# On Unix it is created 0600 inside a 0700 directory. A key directory that
+# grants access beyond its owner is tightened to 0700 on start; one closed too
+# far is refused with a chmod 700 remedy rather than widened. On other
+# platforms no permissions are enforced.
+# If the node logs that it tightened a key directory that was GROUP OR WORLD
+# WRITABLE, treat the key that was sitting there as possibly exposed: delete
+# it so a fresh identity is generated on the next start. An ordinary 0755
+# directory is tightened too, but others could not read a 0600 key or replace
+# its directory entry, so do not rotate for that case. A directory that was
+# merely closed too far is refused, not tightened, and no key in it was ever
+# exposed.
+# Keep it on a persistent volume so the PeerId survives redeploys.
+# Default: ~/.gitlawb/p2p.key
+#GITLAWB_P2P_KEY=/data/keys/p2p.key
+
 # Publicly reachable URL of this node (used in peer announcements)
 GITLAWB_PUBLIC_URL=https://your-node.example.com

+# Base URL for the web view of repos on this node (used by `gl` to print a
+# working View: link after repo creation). Omit for nodes with no web front-end.
+# Distinct from GITLAWB_PUBLIC_URL: that is API reachability for peers;
+# this is browser reachability for humans.
+# GITLAWB_WEB_URL=https://gitlawb.com
+
 # ── Server ────────────────────────────────────────────────────────────────
 GITLAWB_HOST=0.0.0.0
 GITLAWB_PORT=7545
@@ -24,11 +59,26 @@ DATABASE_URL=postgresql://gitlawb:changeme@localhost:5432/gitlawb
 # ── Database pool & startup resilience ────────────────────────────────────
 # Maximum connections in the PostgreSQL pool. A cap, not a floor —
 # connections open lazily. Size against the DB server's max_connections,
-# remembering admin tooling opens its own pool. Each concurrent write pins one
-# connection for its whole duration (the connection-affine advisory lock), so the
-# node REJECTS at boot any value below GITLAWB_MAX_CONCURRENT_GIT_PUSHES + 8
-# headroom — keep this comfortably above that (default 48 for pushes 32).
+# remembering admin tooling opens its own pool.
 GITLAWB_DB_MAX_CONNECTIONS=48
+# Maximum connections in the DEDICATED advisory-lock pool, separate from the
+# pool above. Every in-flight repo write pins one connection here for its whole
+# duration, so this is a hard ceiling on simultaneous writes node-wide: size it
+# to expected peak concurrent writers, not small. Keeping it separate is what
+# stops a push burst from starving ordinary request handlers. Budget
+# (GITLAWB_DB_MAX_CONNECTIONS + this) per node against the server's
+# max_connections, times node count, plus admin tooling.
+GITLAWB_DB_LOCK_POOL_MAX_CONNECTIONS=40
+# Upper bound, in seconds, on any object-storage transfer that runs while a
+# per-repo write lock is HELD (the archive download inside acquire_write and the
+# upload inside release). These were free before the lock's connection was
+# pinned to the guard; now an unbounded stall holds a lock-pool slot, and enough
+# stalls deny every write on the node. The bound applies PER SPAN and there are
+# two (the acquire-side refresh, which covers the existence check and download
+# together, and the release-side upload), so worst-case slot occupancy is about
+# twice this value plus the git work between them. Read it together with the pool
+# size above and with GITLAWB_GIT_SERVICE_TIMEOUT_SECS.
+GITLAWB_LOCK_HELD_TRANSFER_TIMEOUT_SECS=300
 # Seconds a request waits for a pool connection before failing with 503.
 GITLAWB_DB_ACQUIRE_TIMEOUT_SECS=5
 # Upper bound on each startup connect+migrate attempt, in seconds. Keep it
@@ -88,23 +138,32 @@ GITLAWB_P2P_PORT=7546
 # GITLAWB_BOOTSTRAP_DISABLE_SEEDS=true for isolated dev/test networks.
 GITLAWB_BOOTSTRAP_PEERS=
 GITLAWB_BOOTSTRAP_DISABLE_SEEDS=false
-# Require signatures on peer announce/sync write routes. Keep false during
-# rolling upgrades so existing live nodes can still communicate.
+# Require a peer write to prove its DID on every transport: RFC 9421 signatures
+# on the announce/sync routes, and an Ed25519 payload signature on gossip
+# ref-update events. Keep false during rolling upgrades so existing live nodes
+# can still communicate. Upgrade every gossip-publishing peer to a build that
+# signs events before setting this true, or their updates are dropped on
+# arrival with no error visible on their side.
 GITLAWB_REQUIRE_SIGNED_PEER_WRITES=false

 # Require the authenticated pusher to be the repo owner on git-receive-pack.
 # A valid did:key signature is authentication, not authorization: anyone can
-# sign as their own DID. When true, pushes from a non-owner DID are rejected.
-# Keep false until the repo owner is ready for owner-only writes.
-GITLAWB_ENFORCE_OWNER_PUSH=false
+# mint a key, derive its DID and sign, so with this off every signed caller may
+# push to every repository, private ones included. On by default.
+#
+# Set to false only for a rolling upgrade whose pushers are not yet the repo
+# owner. Note that delegated and CI keys count as non-owners: UCAN git/push is
+# verified but not yet honored for authorization, so they cannot push while this
+# is on. See docs/RUN-A-NODE.md.
+GITLAWB_ENFORCE_OWNER_PUSH=true

 # Comma-separated libp2p multiaddrs.
 # Example: /ip4/1.2.3.4/udp/7546/quic-v1/p2p/12D3KooW...
 GITLAWB_P2P_BOOTSTRAP=

 # ── Access control ────────────────────────────────────────────────────────
-# Reserved for private-read mode. Public/private repo read enforcement is not
-# wired in the current live release; do not rely on this for private repositories.
+# Reserved and currently inert. Repo reads are gated per repository by
+# is_public and path-scoped visibility rules; this flag changes nothing.
 GITLAWB_PUBLIC_READ=true

 # Maximum git smart-HTTP pack request size, in bytes.
@@ -191,8 +250,9 @@ GITLAWB_MAX_CONCURRENT_READS_PER_CALLER=16
 #     many WRITE-pool slots. This one is load-bearing for push availability, since
 #     it is acquired before the global write permit: without it, one host minting
 #     disposable did:key identities could open enough slow pushes to monopolize the
-#     write pool and 503 every other source (owner enforcement defaults off, and
-#     the push rate limiter caps arrival rate, not in-flight concurrency).
+#     write pool and 503 every other source (the push rate limiter caps arrival
+#     rate, not in-flight concurrency, and owner enforcement can be turned off for
+#     a rolling upgrade).
 # Keyed on the resolved source IP, never the DID, so a DID farm does not defeat
 # them; keying granularity follows GITLAWB_TRUSTED_PROXY like the read cap above.

@@ -213,10 +273,26 @@ GITLAWB_MAX_CONCURRENT_IPFS_WALKS=32
 # per-caller caps via GITLAWB_TRUSTED_PROXY; reject-before-insert bounded map).
 # Default 4.
 GITLAWB_IPFS_WALK_PER_SOURCE=4
+# Max legacy (NULL-provenance) repos probed per single /ipfs request, bounding the
+# scan-fallback fan-out (git cat-file per candidate repo) for an anonymous caller. A
+# truncated scan sheds a retryable 503, never a false 404. Default 256.
+GITLAWB_IPFS_MAX_LEGACY_PROBES=256
+# Max repo ROWS one /ipfs request's legacy scan may read from the database. The probe
+# ceiling above only counts once a probe runs, and quarantined or private repos are
+# denied before that, so without this an all-denying inventory paged the whole repo
+# table for one anonymous request. A truncated scan sheds a retryable 503 carrying a
+# sealed `continuation` token; echo it as ?scan= to resume, so a holder buried past the
+# ceiling is reachable in ceil(repos / ceiling) + 1 requests. Raising this also raises
+# the per-caller /ipfs work allowance, since each page is charged to it. Lowering it
+# sharpens a coarse oracle: laddering to the end reveals the node's total repo count to
+# within one ceiling. Default 2048.
+GITLAWB_IPFS_MAX_LEGACY_SCAN_ROWS=2048
 # Max EXPENSIVE path-scope visibility walks per single /ipfs request (only a
 # blob in a path-scoped repo costs a full-history walk). Over-cap repos are
 # skipped without a verdict and the scan continues; if the object is then found
-# nowhere the request sheds a retryable 503 instead of a false 404. Default 64.
+# nowhere the request sheds a retryable 503 instead of a false 404. The effective
+# cap is the tighter of this value and the node's internal per-request ceiling of
+# 17, so values above 17 have no effect; lower values do tighten it. Default 64.
 GITLAWB_IPFS_MAX_REPOS_WALKED=64
 # Ceiling on repos one /ipfs request may VISIT past the visibility gate. Each
 # visit costs a repo acquire — on a Tigris cache miss a full archive download,
@@ -237,10 +313,32 @@ GITLAWB_IPFS_MAX_REPO_VISITS=1024
 # Must be 1..=3153600000 (100 years): the node derives an Instant deadline from
 # this value, and a larger one cannot be represented. Default 600.
 GITLAWB_IPFS_REQUEST_BUDGET_SECS=600
+# Shorter budget (seconds) for the pre-walk CID resolve: the lookup that maps a
+# requested CID to its git oid(s), which runs while the scarce walk admission is
+# already held. A well-formed CID with no pin row does no probe and no walk work,
+# so without this a stalled lookup could hold a walk slot for the whole request
+# budget while nothing walked. The effective deadline is the lesser of this and
+# the remaining request budget; walk and probe work stay on the request budget,
+# so a slow but progressing scan is never shed by it.
+# Must be 1..=3153600000 (100 years). Default 10.
+GITLAWB_IPFS_RESOLVE_BUDGET_SECS=10
 # Max /ipfs/{cid} requests per client IP per hour (route flood brake, distinct
 # from the concurrency caps above). 0 disables. Default 600.
 GITLAWB_IPFS_RATE_LIMIT=600

+# Per-client-IP rate limit for the anonymous task read routes
+# (GET /api/v1/tasks, GET /api/v1/tasks/{id}), in requests per hour.
+# Both are publicly reachable (optional_signature). GET /api/v1/tasks runs
+# collect_visible_tasks and returns a visibility-filtered page;
+# GET /api/v1/tasks/{id} runs get_visible_task and returns an opaque 404 when
+# the task is hidden or missing. These reads can require task, repository,
+# and visibility-rule queries even when no task is returned, so the brake
+# bounds the cost of anonymous probes. Keyed on the resolved client IP via
+# GITLAWB_TRUSTED_PROXY. 0 disables. Default: 1200 (a list page
+# followed by per-task reads is a normal client pattern, so this sits above
+# the /ipfs budget). The GraphQL/WS task-read brake shares this budget.
+GITLAWB_TASK_READ_RATE_LIMIT=1200
+
 # ── Creation rate limiting (repo/agent/issue/PR flood brake) ──────────────
 # Max creation requests (POST /api/v1/repos, /api/register, fork, issues,
 # pulls) per client IP per hour, in addition to the per-DID limit. The per-DID
@@ -249,6 +347,14 @@ GITLAWB_IPFS_RATE_LIMIT=600
 # brake caps a single-source flood. Uses GITLAWB_TRUSTED_PROXY below to resolve
 # the client IP. 0 disables. Default 120.
 GITLAWB_CREATE_RATE_LIMIT=120
+# Per-DID creation requests allowed inside the window below (repo create,
+# issue create, PR create, register, fork). 0 is rejected at startup: it would
+# silently disable the limiter. Default 10.
+GITLAWB_RATE_LIMIT_MAX=10
+# Window the per-DID creation count above applies over, in seconds. 0 is
+# rejected at startup: every timestamp would prune immediately, disabling the
+# limiter. Default 3600.
+GITLAWB_RATE_LIMIT_WINDOW_SECS=3600

 # ── Peer-sync rate limiting (per client IP, uses GITLAWB_TRUSTED_PROXY below) ─
 # /api/v1/peers/announce and /api/v1/sync/notify accept unsigned requests from
@@ -274,6 +380,35 @@ GITLAWB_TRUSTED_PROXY=
 # Enable automatic background sync from known peers
 GITLAWB_AUTO_SYNC=false

+# ── External upstream mirroring ───────────────────────────────────────────
+# Default-off scheduled fetch worker for repositories configured as INBOUND.
+# Enabling requires GITLAWB_ENFORCE_OWNER_PUSH=true; the node refuses to start
+# otherwise. The worker updates branches and tags atomically while preserving
+# Gitlawb-owned review/internal refs. Enabling also requires Git >= 2.37 on
+# the host for http.curloptResolve DNS pinning; the node refuses to start
+# when the installed Git is older.
+GITLAWB_UPSTREAM_MIRROR_ENABLED=false
+# Pause after each complete scan. Must be 5..=86400 seconds.
+GITLAWB_UPSTREAM_MIRROR_INTERVAL_SECS=300
+# Deadline for the external fetch child only; the preflight probes run on
+# their own short bound. Must be 1..=3153600000 seconds.
+GITLAWB_UPSTREAM_MIRROR_FETCH_TIMEOUT_SECS=600
+# Keyset page size; every page is processed before the next scan delay.
+GITLAWB_UPSTREAM_MIRROR_PAGE_SIZE=100
+# ── Reconciliation sweep ─────────────────────────────────────────────────
+# Periodic durability sweep: re-derives the public pin set and the withheld-blob
+# recovery set each hour and fills gaps so a dropped replication job never means
+# data loss. Defaults to true; set to false to disable the sweep even when a pin
+# backend (IPFS/Pinata) is configured.
+#
+# Phase-capability matrix:
+#   - Public pin repair: IPFS-only, Pinata-only, or both (requires the
+#     respective backend to be configured).
+#   - Encrypted recovery repair: requires local IPFS (GITLAWB_IPFS_API).
+#     Pinata-only nodes reconcile public pins only; encrypted recovery
+#     reconciliation is not performed.
+GITLAWB_RECONCILIATION_SWEEP=true
+
 # ── iCaptcha proof-of-intelligence gate ───────────────────────────────────
 # Optional gate on create_repo + register: require callers to present an
 # iCaptcha proof (X-ICaptcha-Proof header) earned at icaptcha.gitlawb.com.
@@ -1,5 +1,5 @@
 name: Bug report
-description: Report a defect in a Gitlawb crate or node behavior
+description: Report a defect in a Twigpine crate or node behavior
 labels: ["bug"]
 body:
  - type: markdown
@@ -1,8 +1,8 @@
 blank_issues_enabled: false
 contact_links:
  - name: Security vulnerability
-    url: https://github.com/Gitlawb/node/blob/main/SECURITY.md
+    url: https://github.com/Twigpine/node/blob/main/SECURITY.md
    about: Please report security issues privately per our security policy, not as a public issue.
  - name: Questions & discussion
-    url: https://github.com/Gitlawb/node/discussions
+    url: https://github.com/Twigpine/node/discussions
    about: For usage questions and open-ended discussion.
@@ -1,5 +1,5 @@
 <!--
-  Thanks for contributing to Gitlawb. Keep PRs focused on one change.
+  Thanks for contributing to Twigpine. Keep PRs focused on one change.
  Protocol-level changes (identity, signatures, wire format) should start
  as an issue before code — see CONTRIBUTING.md.
 -->
@@ -86,3 +86,27 @@ jobs:
            exit 1
          fi
          echo "No drift: ignore list is consistent with the pinned hickory-proto version."
+
+      # Drift guard: the RUSTSEC-2026-0253 (lru) ignore is only valid while
+      # the lockfile holds exactly the version it was written for (0.16.3
+      # via alloy). Any move — an upgrade that may carry the upstream fix,
+      # or removal of the dependency — must revisit the ignore. Retirement
+      # owner: node maintainers, via this weekly job (no separate tracking
+      # issue; see the waiver comment in .cargo/audit.toml).
+      - name: Fail if lru moved while its ignore remains
+        run: |
+          set -euo pipefail
+          versions="$(grep -A1 'name = "lru"' Cargo.lock | grep '^version' | cut -d'"' -f2 | sort -u | tr '\n' ' ')"
+          echo "lru versions in Cargo.lock: ${versions:-not present}"
+          lru_ignore_present=false
+          if grep -q 'RUSTSEC-2026-0253' .cargo/audit.toml; then
+            lru_ignore_present=true
+          fi
+          # Drift = the lockfile no longer matches what the ignore assumes:
+          # the version moved (possible fix) or lru left the tree
+          # (dead ignore). Both must fail for explicit re-triage.
+          if [ "$lru_ignore_present" = true ] && [ "${versions}" != "0.16.3 " ]; then
+            echo "::error::lru versions are [${versions:-absent from Cargo.lock}] but the RUSTSEC-2026-0253 ignore is still in .cargo/audit.toml. Re-triage the waiver against the advisory's fixed range and either drop the ignore or update the pinned set."
+            exit 1
+          fi
+          echo "No drift: lru ignore is consistent with the pinned versions."
@@ -0,0 +1,100 @@
+name: CodeQL
+
+# Advanced-setup replacement for CodeQL default setup. Default setup analyzes
+# every detected language on every PR, including Swift (macos-app/), which
+# autobuilds on a macOS runner for ~13 minutes per PR and has hung for 55 on
+# main; that single job was the longest check on most PRs. Here a PR analyzes
+# only the languages it touches (scripts/ci-surfaces.sh), while push to main
+# and the weekly schedule analyze every language, so main's code-scanning
+# baseline stays complete.
+#
+# Requires CodeQL default setup to be DISABLED (Settings -> Code security ->
+# Code scanning): GitHub rejects advanced-setup uploads while it is enabled.
+#
+# A PR that skips a language may show "configuration not found" on the CodeQL
+# check for that language; that is informational and expected.
+
+on:
+  pull_request:
+    branches: [main]
+  push:
+    branches: [main]
+  schedule:
+    - cron: "17 4 * * 1"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: codeql-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  changes:
+    name: changed languages
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    outputs:
+      codeql: ${{ steps.surfaces.outputs.codeql }}
+      matrix: ${{ steps.surfaces.outputs.codeql_matrix }}
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+          fetch-depth: 2
+
+      - name: Map changed paths to languages
+        id: surfaces
+        env:
+          EVENT_NAME: ${{ github.event_name }}
+        run: |
+          set -euo pipefail
+          if [ "$EVENT_NAME" = "pull_request" ]; then
+            # -z: literal paths; plain --name-only C-quotes unusual names.
+            git diff -z --name-only --no-renames HEAD^1 HEAD > changed-paths
+            tr '\0' '\n' < changed-paths >&2
+            scripts/ci-surfaces.sh < changed-paths > surfaces.out
+          else
+            scripts/ci-surfaces.sh --all > surfaces.out
+          fi
+          cat surfaces.out
+          grep -E '^(codeql|codeql_matrix)=' surfaces.out >> "$GITHUB_OUTPUT"
+
+  analyze:
+    name: Analyze (${{ matrix.language }})
+    needs: changes
+    if: needs.changes.outputs.codeql == 'true'
+    runs-on: ${{ matrix.runner }}
+    # Normal runs take <=14 min (Swift autobuild is the longest); cap a hung
+    # autobuild instead of letting it hold a runner for an hour.
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      security-events: write
+      actions: read
+      packages: read
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.changes.outputs.matrix) }}
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: Initialize CodeQL
+        uses: github/codeql-action/init@9b3254085494b55a0f7c4c2b3d2d52cc712ce370 # v4.39.1
+        with:
+          languages: ${{ matrix.language }}
+          build-mode: ${{ matrix.build-mode }}
+
+      - name: Autobuild
+        if: matrix.build-mode == 'autobuild'
+        uses: github/codeql-action/autobuild@9b3254085494b55a0f7c4c2b3d2d52cc712ce370 # v4.39.1
+
+      - name: Perform CodeQL Analysis
+        uses: github/codeql-action/analyze@9b3254085494b55a0f7c4c2b3d2d52cc712ce370 # v4.39.1
+        with:
+          category: "/language:${{ matrix.language }}"
@@ -30,8 +30,62 @@ env:
  RUST_BACKTRACE: 1

 jobs:
+  # Decide which checks this event needs. On pull_request, each job below runs
+  # only when the PR touches the surface it validates, and the cargo test /
+  # release-build jobs are scoped to the touched crates plus their reverse
+  # dependencies (scripts/ci-surfaces.sh). Every other event (push to main,
+  # merge_group, the release flow's workflow_dispatch) runs the full suite, so
+  # main and the merge queue keep whole-workspace coverage and main-red-alert
+  # still sees a complete run. A skipped job reports as "skipped", not green.
+  changes:
+    name: changed surfaces
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    outputs:
+      rust: ${{ steps.surfaces.outputs.rust }}
+      rust_packages: ${{ steps.surfaces.outputs.rust_packages }}
+      windows: ${{ steps.surfaces.outputs.windows }}
+      windows_packages: ${{ steps.surfaces.outputs.windows_packages }}
+      docker: ${{ steps.surfaces.outputs.docker }}
+      audit: ${{ steps.surfaces.outputs.audit }}
+      core_deps: ${{ steps.surfaces.outputs.core_deps }}
+      release_versioning: ${{ steps.surfaces.outputs.release_versioning }}
+      release_tag: ${{ steps.surfaces.outputs.release_tag }}
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+          # The pull_request checkout is the test-merge commit; depth 2 brings
+          # its first parent (the base tip) so the diff is exactly this PR.
+          fetch-depth: 2
+
+      # Every job's gate comes from the mapper, so pin its mapping first; a
+      # failure here fails this job and the run is red, not silently narrowed.
+      - name: Test the surface mapper
+        run: scripts/test-ci-surfaces.sh
+
+      - name: Map changed paths to surfaces
+        id: surfaces
+        env:
+          EVENT_NAME: ${{ github.event_name }}
+        run: |
+          set -euo pipefail
+          if [ "$EVENT_NAME" = "pull_request" ]; then
+            # -z: literal paths; plain --name-only C-quotes unusual names.
+            git diff -z --name-only --no-renames HEAD^1 HEAD > changed-paths
+            tr '\0' '\n' < changed-paths >&2
+            scripts/ci-surfaces.sh < changed-paths > surfaces.out
+          else
+            scripts/ci-surfaces.sh --all > surfaces.out
+          fi
+          cat surfaces.out
+          cat surfaces.out >> "$GITHUB_OUTPUT"
+
  fmt-clippy:
    name: fmt + clippy
+    needs: changes
+    if: needs.changes.outputs.rust == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 30
    steps:
@@ -57,6 +111,8 @@ jobs:

  release-tag-validation:
    name: release tag validation
+    needs: changes
+    if: needs.changes.outputs.release_tag == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 5
    steps:
@@ -68,8 +124,26 @@ jobs:
      - name: Test release tag resolver
        run: scripts/test-resolve-release-tag.sh

+  triage-detector:
+    name: triage detector matrix
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+
+      # pr-triage.yml runs on pull_request_target and never checks out PR
+      # code, so its inline detector can only be exercised here, where the
+      # proposed workflow body IS the checked-out one.
+      - name: Test inline-test detector
+        run: node scripts/test-pr-triage-detect.mjs
+
  test:
    name: test (${{ matrix.toolchain }})
+    needs: changes
+    if: needs.changes.outputs.rust == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 45
    strategy:
@@ -114,10 +188,17 @@ jobs:
      - name: cargo test
        env:
          DATABASE_URL: postgres://postgres:postgres@localhost:5432/gitlawb_test
-        run: cargo test --locked --workspace
+          # `--workspace`, or `-p <crate>...` for the crates a PR touches plus
+          # every crate that depends on them.
+          PACKAGES: ${{ needs.changes.outputs.rust_packages }}
+        run: |
+          # shellcheck disable=SC2086 # PACKAGES is a word list by design
+          cargo test --locked $PACKAGES

  build-release:
    name: build --release
+    needs: changes
+    if: needs.changes.outputs.rust == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 45
    steps:
@@ -137,10 +218,20 @@ jobs:
          key: release

      - name: cargo build --release
-        run: cargo build --locked --release --workspace
+        env:
+          PACKAGES: ${{ needs.changes.outputs.rust_packages }}
+        run: |
+          # shellcheck disable=SC2086 # PACKAGES is a word list by design
+          cargo build --locked --release $PACKAGES

  audit:
    name: cargo audit
+    # cargo audit reads only Cargo.lock (+ .cargo/audit.toml), so a PR that
+    # leaves both unchanged cannot change its verdict; a newly published
+    # advisory against main is caught by the push-to-main run and the weekly
+    # Scheduled Audit instead of turning every unrelated PR red.
+    needs: changes
+    if: needs.changes.outputs.audit == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 30
    steps:
@@ -200,6 +291,8 @@ jobs:

  msrv:
    name: MSRV (Rust 1.91)
+    needs: changes
+    if: needs.changes.outputs.rust == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 30
    steps:
@@ -223,6 +316,12 @@ jobs:

  docker-build:
    name: Docker build smoke test
+    # Runs when the image's own inputs change: the Dockerfile, the manifests
+    # its dependency layer copies, the lockfile, bootstrap-peers.json. A
+    # source-only Rust change is already compiled by `build --release` (stable)
+    # and `MSRV` (1.91, the image's toolchain); push to main still runs this.
+    needs: changes
+    if: needs.changes.outputs.docker == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 30
    steps:
@@ -239,6 +338,8 @@ jobs:

  test-windows:
    name: test (windows, non-blocking)
+    needs: changes
+    if: needs.changes.outputs.windows == 'true'
    runs-on: windows-latest
    timeout-minutes: 45
    # Non-blocking, mirroring the informational `beta` lane above. We ship `gl`
@@ -266,7 +367,13 @@ jobs:
          key: windows

      - name: cargo test (shipped Windows crates)
-        run: cargo test --locked -p gl -p git-remote-gitlawb
+        shell: bash
+        env:
+          # `-p gl -p git-remote-gitlawb`, or whichever of the two a PR affects.
+          PACKAGES: ${{ needs.changes.outputs.windows_packages }}
+        run: |
+          # shellcheck disable=SC2086 # PACKAGES is a word list by design
+          cargo test --locked $PACKAGES

  # gitlawb-core is embedded by every consumer (gl, git-remote-gitlawb, the node
  # daemon), so it must stay lean. This gate fails if gitlawb-core's normal
@@ -276,6 +383,8 @@ jobs:
  # cannot do. Regen instructions live in the allowlist header.
  core-deps-purity:
    name: gitlawb-core dependency purity
+    needs: changes
+    if: needs.changes.outputs.core_deps == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 15
    steps:
@@ -308,6 +417,8 @@ jobs:
  # not its author knew this gate existed.
  release-versioning:
    name: release versioning coverage
+    needs: changes
+    if: needs.changes.outputs.release_versioning == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 15
    steps:
@@ -76,9 +76,221 @@ jobs:
            // line after horizontal whitespace, covering bare, cfg(test), namespaced,
            // optional leading ::, and raw-identifier (r#test) forms. No head-file fetch
            // or lexer: needs-tests is advisory, not a merge gate.
-            const addsInlineTest = files.some(f =>
-              f.filename.endsWith(".rs") && f.patch &&
-              /^\+[ \t]*#\[[ \t]*(?:cfg[ \t]*\([ \t]*test[ \t]*\)|(?:::[ \t]*)?(?:[\w-]+[ \t]*::[ \t]*)*(?:r#)?test\b)/m.test(f.patch));
+            //
+            // The final path component may also carry `test` as an underscore-delimited
+            // segment, so third-party harnesses land without enumerating them one by one
+            // (#[test_case(1)], #[wasm_bindgen_test]). It is deliberately a segment and
+            // not a substring: a substring would also match #[contest] / #[latest], and a
+            // false positive here SUPPRESSES needs-tests rather than adding it. This
+            // heuristic has to fail loud, because a wrongly labeled PR gets corrected by
+            // the author while a wrongly cleared one is invisible. That trade costs the
+            // no-underscore harness names (#[rstest]), which stay unmatched on purpose.
+            //
+            // The segment classes are [A-Za-z0-9]+ rather than \w+ on purpose, so that `_`
+            // is only ever the literal delimiter. \w+ contains `_`, which makes the repeat
+            // ambiguous and backtracks exponentially: on this workflow's pull_request_target
+            // trigger f.patch is fork-controlled, so `#[test` followed by a long `_a` run is
+            // a permissionless stall of the triage job.
+            // Rust allows whitespace, line breaks, and comments between the
+            // attribute path and ] or (. A flat regex cannot decide that
+            // boundary: `//` runs to end of line, and block comments NEST, which
+            // is beyond regular languages — every enumeration of comment
+            // spellings so far has either dropped a legal separator (regressing
+            // real tests into needs-tests) or let unrelated patch records
+            // complete each other (clearing needs-tests from a testless PR, the
+            // invisible direction). So the closer is found by a token-aware
+            // scan instead: it consumes horizontal whitespace, `//`-to-newline,
+            // and depth-counted `/* */` runs, and accepts only when the first
+            // real token after the attribute path is `]` or `(`.
+            //
+            // Association is positional: the scan starts on the attribute's own
+            // added line and may continue ONLY through the immediately following
+            // added lines (a context line, hunk header, or removal breaks the
+            // chain). Unrelated `]`/`(` lines elsewhere in the patch can never
+            // complete a path they do not adjoin. The continuation is bounded;
+            // past the bound the scan gives up and reports NO inline test, which
+            // fails toward the loud label rather than the silent clearance.
+            //
+            // The detector body is fenced by the TRIAGE_DETECTOR markers so
+            // scripts/test-pr-triage-detect.mjs can extract and run it against
+            // the committed case matrix. Keep the fenced block self-contained.
+            // TRIAGE_DETECTOR_BEGIN
+            const TEST_ATTR_PATH =
+              "(?:cfg[ \\t]*\\([ \\t]*test[ \\t]*\\)|(?:::[ \\t]*)?(?:[\\w-]+[ \\t]*::[ \\t]*)*(?:r#)?(?:[A-Za-z0-9]+_)*test(?:_[A-Za-z0-9]+)*)";
+            // Anchored to the start of one added line; group 1 is whatever
+            // follows the path on that line, handed to the separator scan.
+            const TEST_ATTR_LINE = new RegExp(
+              "^[ \\t]*#\\[[ \\t]*" + TEST_ATTR_PATH + "([\\s\\S]*)$"
+            );
+            // An attribute split across more added lines than this is not a
+            // spelling anyone writes; refusing to scan further keeps the walk
+            // linear in the patch and fails toward the loud label.
+            const MAX_ATTR_CONTINUATION_LINES = 16;
+            // segs[0] is the remainder of the attribute's own line; each later
+            // entry is the text of one immediately following added line. True
+            // only when the first non-separator token across them is ] or (.
+            // Single forward pass, no backtracking: fork-controlled input.
+            function attrCloserFollows(segs) {
+              let depth = 0; // open block-comment nesting carried across lines
+              for (const seg of segs) {
+                let i = 0;
+                while (i < seg.length) {
+                  if (depth > 0) {
+                    if (seg.startsWith("/*", i)) { depth += 1; i += 2; }
+                    else if (seg.startsWith("*/", i)) { depth -= 1; i += 2; }
+                    else { i += 1; }
+                    continue;
+                  }
+                  const c = seg[i];
+                  if (c === " " || c === "\t") { i += 1; continue; }
+                  if (seg.startsWith("//", i)) break; // comment to end of line
+                  if (seg.startsWith("/*", i)) { depth = 1; i += 2; continue; }
+                  return c === "]" || c === "(";
+                }
+                // Line exhausted inside separators; the newline is itself a
+                // separator, so continue on the next added line.
+              }
+              return false; // out of adjoined lines (or bound hit): stay loud
+            }
+            // Before the candidate scan, a pre-pass walks added patch lines
+            // tracking Rust lexical state: raw strings (r#*", br#*", cr#*"),
+            // block comments (/* */ with nesting), regular strings ("...",
+            // which can span newlines), and line comments (//). An added line
+            // that starts inside any of these is skipped as a candidate,
+            // because a #[test_case] appearing there is lexical non-code, not
+            // a real test attribute. State is only tracked across consecutive
+            // added lines and reset at any break (context line, removal,
+            // hunk header): context lines carry a partial view of the file,
+            // so a " or r#" or /* there might be inside a string that opened
+            // before the hunk's visible window, and treating it as an opener
+            // would mark every later added line as non-code (a false negative
+            // on a real #[test]). The tradeoff is that a non-code region
+            // opened by a context line is invisible: an added #[test] inside
+            // an existing raw string or block comment is scanned as code and
+            // may match (a false positive, silent direction). This is accepted
+            // because (1) the old detector had the same behavior, (2) the
+            // pattern is rare, and (3) needs-tests is advisory, not a merge
+            // gate. Scanning context lines to close this gap would re-introduce
+            // the false negative, which affects the far more common case of a
+            // context line carrying a stray " from a string that opened before
+            // the visible window. Single forward pass, O(n) in patch length.
+            function patchAddsInlineTest(patch) {
+              const lines = patch.split("\n");
+              const nonCodeStart = new Set();
+              let rawHashes = -1; // -1 = not in raw string; >=0 = # count
+              let blockDepth = 0;
+              let inString = false;
+              for (let i = 0; i < lines.length; i++) {
+                const line = lines[i];
+                if (!line) continue;
+                const prefix = line[0];
+                // Only scan added lines. Context and removal lines carry
+                // a partial view of the file: a " or r#" or /* on a context
+                // line might be inside a string that opened before the
+                // hunk's visible window, and treating it as an opener would
+                // mark every later added line as non-code (a false negative
+                // on a real #[test]). So lexical state is only tracked across
+                // consecutive added lines and reset at any break. A non-code
+                // region opened by a context line is invisible; the candidate
+                // is still scanned, which fails toward the loud label.
+                if (prefix !== "+") {
+                  rawHashes = -1; blockDepth = 0; inString = false;
+                  continue;
+                }
+                const text = line.slice(1);
+                if (rawHashes >= 0 || blockDepth > 0 || inString)
+                  nonCodeStart.add(i);
+                let j = 0;
+                while (j < text.length) {
+                  if (rawHashes >= 0) {
+                    if (text[j] === '"') {
+                      let h = 1;
+                      while (h <= rawHashes && text[j + h] === "#") h++;
+                      if (h > rawHashes) { rawHashes = -1; j += h; continue; }
+                    }
+                    j += 1; continue;
+                  }
+                  if (blockDepth > 0) {
+                    if (text.startsWith("/*", j)) { blockDepth += 1; j += 2; }
+                    else if (text.startsWith("*/", j)) { blockDepth -= 1; j += 2; }
+                    else { j += 1; }
+                    continue;
+                  }
+                  if (inString) {
+                    if (text[j] === "\\") { j += 2; continue; }
+                    if (text[j] === '"') { inString = false; j += 1; continue; }
+                    j += 1; continue;
+                  }
+                  // Recognize Rust char and byte-char literals ('X', '\X',
+                  // '\u{...}', b'X') so their contents never open string or
+                  // comment state. A char literal has a closing ' within a
+                  // bounded distance; a lifetime or label ('a, 'static,
+                  // 'label:) does not, so it falls through as a regular char.
+                  // This is not a skip-to-next-' rule: the lookahead is 2-4
+                  // chars, or bounded to 20 for \u escapes (at most 6 hex
+                  // digits), never unbounded.
+                  if (text[j] === "'") {
+                    let k = j + 1;
+                    if (text[k] === "\\") {
+                      k += 1;
+                      if (text[k] === "u" && text[k + 1] === "{") {
+                        const uBound = k + 20;
+                        while (k < uBound && text[k] && text[k] !== "}") k++;
+                        if (text[k] === "}") k++;
+                      } else {
+                        k += 1;
+                      }
+                    } else {
+                      k += 1;
+                    }
+                    if (text[k] === "'") { j = k + 1; continue; }
+                    // lifetime or label: just skip the apostrophe
+                  }
+                  if (text[j] === '"') { inString = true; j += 1; continue; }
+                  if (text.startsWith("//", j)) break;
+                  if (text.startsWith("/*", j)) { blockDepth = 1; j += 2; continue; }
+                  const prev = j > 0 ? text[j - 1] : "";
+                  const isIdent = /[A-Za-z0-9_]/.test(prev);
+                  if (!isIdent && text[j] === "r") {
+                    let k = j + 1, h = 0;
+                    while (text[k + h] === "#") h++;
+                    if (text[k + h] === '"') {
+                      rawHashes = h; j = k + h + 1; continue;
+                    }
+                  }
+                  if (!isIdent && (text[j] === "b" || text[j] === "c") && text[j + 1] === "r") {
+                    let k = j + 2, h = 0;
+                    while (text[k + h] === "#") h++;
+                    if (text[k + h] === '"') {
+                      rawHashes = h; j = k + h + 1; continue;
+                    }
+                  }
+                  j += 1;
+                }
+              }
+              for (let i = 0; i < lines.length; i++) {
+                if (lines[i][0] !== "+") continue;
+                if (nonCodeStart.has(i)) continue;
+                const m = TEST_ATTR_LINE.exec(lines[i].slice(1));
+                if (m === null) continue;
+                const segs = [m[1]];
+                for (
+                  let j = i + 1;
+                  j < lines.length &&
+                  lines[j][0] === "+" &&
+                  segs.length <= MAX_ATTR_CONTINUATION_LINES;
+                  j++
+                ) {
+                  segs.push(lines[j].slice(1));
+                }
+                if (attrCloserFollows(segs)) return true;
+              }
+              return false;
+            }
+            // TRIAGE_DETECTOR_END
+            const addsInlineTest = files.some(
+              f => f.filename.endsWith(".rs") && !!f.patch && patchAddsInlineTest(f.patch)
+            );
            const touchedTests =
              names.some(n => n.includes("/tests/") || n.endsWith("_test.rs")) || addsInlineTest;
            if (changedRust && !touchedTests) want.add("needs-tests");
@@ -124,7 +336,7 @@ jobs:
              "",
              lines,
              "",
-              "See [CONTRIBUTING.md](https://github.com/Gitlawb/node/blob/main/CONTRIBUTING.md). Update the PR and these notes will clear automatically.",
+              "See [CONTRIBUTING.md](https://github.com/Twigpine/node/blob/main/CONTRIBUTING.md). Update the PR and these notes will clear automatically.",
            ].join("\n");

            await github.rest.issues.createComment({ owner, repo, issue_number: prNumber, body: comment });
@@ -20,7 +20,15 @@ on:
        type: string

 concurrency:
-  group: release-${{ github.ref }}
+  # Not keyed on the ref: two runs racing `imagetools create` on the same
+  # moving tags could interleave, so runs of this workflow serialize outright.
+  # GitHub keeps at most one running and one pending run per group; a newer
+  # queued run replaces the pending one, and cancel-in-progress: false
+  # protects only the run already executing. So a backfill dispatched while a
+  # release is running, followed by a push to main, is replaced silently: the
+  # remedy is to re-dispatch it once the running release completes, which the
+  # push-only gate on the moving-tag step makes safe in any ordering.
+  group: release
  cancel-in-progress: false

 # Least privilege by default; every job opts in to exactly what it needs.
@@ -31,12 +39,13 @@ jobs:
    # Skipped on manual dispatch: a backfill must never create or advance a
    # release PR as a side effect (npm-publish tolerates the skipped need via
    # !cancelled(), and every other job gates on release_created == 'true').
-    if: ${{ github.repository == 'Gitlawb/node' && github.event_name != 'workflow_dispatch' }}
+    if: ${{ github.repository == 'Twigpine/node' && github.event_name != 'workflow_dispatch' }}
    name: Release Please
    runs-on: ubuntu-latest
    permissions:
      contents: write
      pull-requests: write
+    environment: release
    outputs:
      release_created: ${{ steps.release.outputs.release_created }}
      tag_name: ${{ steps.release.outputs.tag_name }}
@@ -64,7 +73,7 @@ jobs:
  sync-release-lock:
    name: Sync Cargo.lock on the release PR
    needs: release-please
-    if: ${{ !cancelled() && github.repository == 'Gitlawb/node' && github.event_name != 'workflow_dispatch' }}
+    if: ${{ !cancelled() && github.repository == 'Twigpine/node' && github.event_name != 'workflow_dispatch' }}
    runs-on: ubuntu-latest
    permissions:
      contents: write
@@ -78,6 +87,7 @@ jobs:
      # dispatch the release PR's head carries no check runs and the --locked
      # gate never validates it.
      actions: write
+    environment: release
    steps:
      # Prefer the fresh `pr` output when this run just created/updated the
      # release PR; otherwise fall back to querying for an open release-please
@@ -215,6 +225,7 @@ jobs:
    permissions:
      contents: read
      packages: write
+    environment: release
    steps:
      - name: Check out workflow scripts
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
@@ -226,6 +237,7 @@ jobs:
        env:
          DISPATCH_TAG: ${{ inputs.docker_backfill_tag }}
          RELEASE_TAG: ${{ needs.release-please.outputs.tag_name }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          set -euo pipefail
          scripts/resolve-release-tag.sh "${DISPATCH_TAG:-$RELEASE_TAG}"
@@ -237,7 +249,9 @@ jobs:
      - name: Checkout release tag
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
-          ref: ${{ steps.rel.outputs.tag }}
+          # Pin the commit the resolver verified; a tag ref re-resolved here
+          # could have been moved between resolve and checkout.
+          ref: ${{ steps.rel.outputs.tag_commit || format('refs/tags/{0}', steps.rel.outputs.tag) }}
          persist-credentials: false

      - name: Set up Docker Buildx
@@ -312,6 +326,7 @@ jobs:
    permissions:
      contents: read
      packages: write
+    environment: release
    steps:
      - name: Check out workflow scripts
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
@@ -323,9 +338,14 @@ jobs:
        env:
          DISPATCH_TAG: ${{ inputs.docker_backfill_tag }}
          RELEASE_TAG: ${{ needs.release-please.outputs.tag_name }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          set -euo pipefail
          scripts/resolve-release-tag.sh "${DISPATCH_TAG:-$RELEASE_TAG}"
+          # ghcr requires a lowercase repository path, and unlike metadata-action,
+          # buildx's `--output name=` does no lowercasing, so a mixed-case owner
+          # makes the digest push fail with "invalid reference format".
+          echo "image=ghcr.io/${GITHUB_REPOSITORY,,}" >> "$GITHUB_OUTPUT"

      - name: Download digests
        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
@@ -345,24 +365,80 @@ jobs:
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Create and push multi-arch manifest
+        id: manifest
        env:
          VERSION: ${{ steps.rel.outputs.version }}
+          IMAGE: ${{ steps.rel.outputs.image }}
        run: |
          set -euo pipefail
-          # ghcr requires a lowercase repository path.
-          IMAGE="ghcr.io/${GITHUB_REPOSITORY,,}"
-          MAJOR_MINOR="${VERSION%.*}"
+          # The docker matrix pushes one digest per arch leg; anything else is
+          # a broken set, not a smaller multi-arch image.
+          count="$(find /tmp/digests -maxdepth 1 -type f | wc -l)"
+          if [ "$count" -ne 2 ]; then
+            echo "::error::expected 2 arch digests in /tmp/digests, found $count"
+            exit 1
+          fi
          digests=""
          for f in /tmp/digests/*; do
-            digests="$digests $IMAGE@sha256:$(basename "$f")"
+            d="$(basename "$f")"
+            case "$d" in
+              *[!0-9a-f]*)
+                echo "::error::digest filename is not lowercase sha256 hex: $d"
+                exit 1
+                ;;
+            esac
+            if [ "${#d}" -ne 64 ]; then
+              echo "::error::digest is not 64 hex chars: $d"
+              exit 1
+            fi
+            digests="$digests $IMAGE@sha256:$d"
          done
+          # The immutable tag always publishes. The moving tags are applied by
+          # the gated step below, which never runs on workflow_dispatch.
          # shellcheck disable=SC2086
-          docker buildx imagetools create \
-            -t "$IMAGE:$VERSION" \
-            -t "$IMAGE:$MAJOR_MINOR" \
-            -t "$IMAGE:latest" \
-            $digests
+          docker buildx imagetools create -t "$IMAGE:$VERSION" $digests
          docker buildx imagetools inspect "$IMAGE:$VERSION"
+          echo "digests=${digests# }" >> "$GITHUB_OUTPUT"
+
+      - name: Move floating tags
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' }}
+        env:
+          VERSION: ${{ steps.rel.outputs.version }}
+          IMAGE: ${{ steps.rel.outputs.image }}
+          DIGESTS: ${{ steps.manifest.outputs.digests }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          MAJOR_MINOR="${VERSION%.*}"
+          # :latest and :X.Y track the newest release. Only a push to main
+          # carries a release-please-cut version, and release-please versions
+          # are monotonic, so on this path the tags only ever advance. A
+          # workflow_dispatch backfill republishes an existing tag and must
+          # never move a pointer, so this step skips on dispatch. After a
+          # backfill that is genuinely newest in its minor line, advance :X.Y
+          # by hand: docker buildx imagetools create -t "$IMAGE:X.Y" <digests>.
+          # The event gate alone does not bound a stale run: re-running an
+          # older push run replays its stored needs outputs under the
+          # original event and ref, and would move the pointers backward to
+          # that run's version. Floor against live state instead: this run's
+          # release already exists by now, so VERSION must be the newest
+          # published release or the run is stale and the tags stay put. The
+          # releases list is ordered by creation, not semver, so read every
+          # page and take the semver maximum, not the first page's.
+          tags="$(gh api --paginate "repos/$GITHUB_REPOSITORY/releases?per_page=100" \
+            -q '.[] | select(.draft == false) | .tag_name')" || {
+            echo "::error::could not list releases to bound the moving tags"
+            exit 1
+          }
+          floor="$(printf '%s\n' "$tags" \
+            | { grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' || true; } \
+            | sed 's/^v//' | sort -V | tail -n 1)"
+          if [ -n "$floor" ] && [ "$floor" != "$VERSION" ]; then
+            echo "::error::release v$floor is newer than this run's v$VERSION; refusing to move the moving tags backward (stale re-run?)"
+            exit 1
+          fi
+          # shellcheck disable=SC2086
+          docker buildx imagetools create -t "$IMAGE:latest" -t "$IMAGE:$MAJOR_MINOR" $DIGESTS

      - name: Release summary
        env:
@@ -386,6 +462,7 @@ jobs:
    continue-on-error: ${{ startsWith(matrix.target, 'x86_64-pc-windows') }}
    permissions:
      contents: write
+    environment: release
    strategy:
      fail-fast: false
      matrix:
@@ -405,10 +482,27 @@ jobs:
          - target: x86_64-pc-windows-msvc
            os: windows-latest
    steps:
+      - name: Check out workflow scripts
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: Resolve release tag
+        id: rel
+        env:
+          RELEASE_TAG: ${{ needs.release-please.outputs.tag_name }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        shell: bash # matrix includes windows-latest, where run: defaults to pwsh
+        run: |
+          set -euo pipefail
+          scripts/resolve-release-tag.sh "$RELEASE_TAG"
+
      - name: Checkout release tag
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
-          ref: ${{ needs.release-please.outputs.tag_name }}
+          # Pin the commit the resolver verified; a tag ref re-resolved here
+          # could have been moved between resolve and checkout.
+          ref: ${{ steps.rel.outputs.tag_commit || format('refs/tags/{0}', steps.rel.outputs.tag) }}
          persist-credentials: false

      - name: Set up Rust toolchain
@@ -454,7 +548,7 @@ jobs:
        shell: bash
        run: |
          set -euo pipefail
-          VERSION="${{ needs.release-please.outputs.version }}"
+          VERSION="${{ steps.rel.outputs.version }}"
          BIN_DIR="target/${{ matrix.target }}/release"
          EXE=""
          case "${{ matrix.target }}" in *windows*) EXE=".exe" ;; esac
@@ -474,7 +568,7 @@ jobs:
        run: |
          set -euo pipefail
          TARGET="${{ matrix.target }}"
-          NAME="gitlawb-node-${{ needs.release-please.outputs.version }}-${TARGET}"
+          NAME="gitlawb-node-${{ steps.rel.outputs.version }}-${TARGET}"
          BIN_DIR="target/${TARGET}/release"
          EXE=""
          case "$TARGET" in *windows*) EXE=".exe" ;; esac
@@ -515,7 +609,7 @@ jobs:
      - name: Attach to release
        uses: softprops/action-gh-release@72f2c25fcb47643c292f7107632f7a47c1df5cd8 # v2.3.2
        with:
-          tag_name: ${{ needs.release-please.outputs.tag_name }}
+          tag_name: ${{ steps.rel.outputs.tag }}
          files: |
            dist/*.tar.gz
            dist/*.zip
@@ -528,7 +622,9 @@ jobs:
    # (release-binaries is skipped on dispatch, hence !cancelled()). Publishing
    # auth is npm Trusted Publishing (GitHub OIDC) — no NPM_TOKEN. Each
    # @gitlawb package must have this repo + workflow (release.yml) configured
-    # as its trusted publisher on npmjs.com, or publish fails loudly here —
+    # as its trusted publisher on npmjs.com, scoped to the release environment
+    # so a dispatched copy of this file on another ref cannot mint a token, or
+    # publish fails loudly here —
    # deliberately loud: the silent secret-guard skip is how 0.4.x–0.6.0
    # never reached npm.
    if: >-
@@ -540,6 +636,7 @@ jobs:
    permissions:
      contents: read
      id-token: write # npm trusted publishing (OIDC) + provenance
+    environment: release
    steps:
      - name: Check out workflow scripts
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
@@ -551,6 +648,7 @@ jobs:
        env:
          DISPATCH_TAG: ${{ inputs.npm_backfill_tag }}
          RELEASE_TAG: ${{ needs.release-please.outputs.tag_name }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          set -euo pipefail
          scripts/resolve-release-tag.sh "${DISPATCH_TAG:-$RELEASE_TAG}"
@@ -558,7 +656,9 @@ jobs:
      - name: Checkout release tag
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
-          ref: ${{ steps.rel.outputs.tag }}
+          # Pin the commit the resolver verified; a tag ref re-resolved here
+          # could have been moved between resolve and checkout.
+          ref: ${{ steps.rel.outputs.tag_commit || format('refs/tags/{0}', steps.rel.outputs.tag) }}
          persist-credentials: false

      # npm >= 11.5.1 performs the OIDC token exchange automatically when the
@@ -571,14 +671,23 @@ jobs:
      - name: Ensure npm supports trusted publishing
        run: |
          set -euo pipefail
-          npm install -g npm@^11.5.1
-          npm --version
+          npm install -g npm@11.19.1
+          # Assert the pin took, so an accidental unpin or a drift off 11.19.1
+          # fails loudly here instead of silently running a different npm on
+          # the OIDC path.
+          installed="$(npm --version)"
+          echo "npm installed: $installed"
+          if [ "$installed" != "11.19.1" ]; then
+            echo "::error::npm is not the pinned 11.19.1 (got: $installed)"
+            exit 1
+          fi

      - name: Lay in release binaries
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          VERSION: ${{ steps.rel.outputs.version }}
          TAG: ${{ steps.rel.outputs.tag }}
+          ASSETS: ${{ steps.rel.outputs.assets }}
        run: |
          set -euo pipefail
          # npm platform package -> Rust target triple (unix only; Windows is not
@@ -595,8 +704,18 @@ jobs:
            target="${entry#*:}"
            archive="gitlawb-node-${VERSION}-${target}.tar.gz"
            echo "==> $pkg <- $archive"
-            gh release download "$TAG" --repo "$GITHUB_REPOSITORY" \
-              --pattern "$archive" --dir _dl --clobber
+            # Download by the asset id the resolver captured and uploader-
+            # checked, not by name: release assets are mutable, and an id can
+            # only ever point at the exact blob captured at resolve time.
+            # The map keys are base64-encoded asset names.
+            want="$(printf '%s' "$archive" | base64 -w0)"
+            asset_id="$(printf '%s\n' "$ASSETS" | awk -v n="$want" '$1 == n {print $2; exit}')"
+            if [ -z "$asset_id" ]; then
+              echo "::error::release $TAG has no asset $archive captured at resolve time"
+              exit 1
+            fi
+            gh api "repos/$GITHUB_REPOSITORY/releases/assets/$asset_id" \
+              -H 'Accept: application/octet-stream' > "_dl/$archive"
            tar -xzf "_dl/$archive" -C _dl
            src="_dl/gitlawb-node-${VERSION}-${target}"
            cp "$src/gl" "npm/packages/$pkg/gl"
@@ -635,14 +754,39 @@ jobs:
          # Platform packages first, then the wrapper (so its optionalDependencies resolve).
          # Skip versions already on the registry so a rerun after a partial publish
          # is idempotent instead of erroring on the first existing package.
+          # npm publish moves the latest dist-tag to whatever it publishes, so a
+          # backfilled older version would hand :latest to stale code. Decide
+          # the dist-tag per package from that package's own registry latest:
+          # a partial publish can leave the platform packages ahead of the
+          # wrapper. E404 means the package has never been published, so this
+          # release gets latest; any other lookup failure must fail closed
+          # rather than guess at the dist-tag and risk moving latest backward.
+          # stderr goes to a file so a warning on a successful lookup cannot
+          # leak into registry_latest; it is read only when the lookup fails.
+          npm_err="$(mktemp)"
+          trap 'rm -f -- "$npm_err"' EXIT
          for pkg in gl-darwin-arm64 gl-darwin-x64 gl-linux-arm64 gl-linux-x64 gl; do
            name="@gitlawb/$pkg"
            if npm view "$name@$VERSION" version >/dev/null 2>&1; then
              echo "==> $name@$VERSION already published, skipping"
              continue
            fi
-            echo "==> npm publish $name@$VERSION"
-            npm publish "npm/packages/$pkg" --provenance --access public
+            registry_latest="$(npm view "$name" dist-tags.latest 2>"$npm_err")" || {
+              if ! grep -q E404 "$npm_err"; then
+                cat "$npm_err" >&2
+                echo "::error::npm dist-tags lookup failed for $name; not guessing the dist-tag"
+                exit 1
+              fi
+              registry_latest=""
+            }
+            dist_tag="latest"
+            if [ -n "$registry_latest" ] && \
+               [ "$VERSION" != "$(printf '%s\n%s\n' "$registry_latest" "$VERSION" | sort -V | tail -1)" ]; then
+              dist_tag="backfill"
+              echo "::notice::$VERSION is older than $name@$registry_latest; publishing under dist-tag backfill"
+            fi
+            echo "==> npm publish $name@$VERSION (dist-tag $dist_tag)"
+            npm publish "npm/packages/$pkg" --provenance --access public --tag "$dist_tag"
          done

  homebrew-bump:
@@ -652,6 +796,7 @@ jobs:
    runs-on: ubuntu-latest
    permissions:
      contents: read # GITHUB_TOKEN downloads release assets; the tap push uses HOMEBREW_TAP_PAT
+    environment: release
    steps:
      - name: Guard on secret
        id: guard
@@ -665,11 +810,27 @@ jobs:
            echo "enabled=true" >> "$GITHUB_OUTPUT"
          fi

+      - name: Check out workflow scripts
+        if: ${{ steps.guard.outputs.enabled == 'true' }}
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: Resolve release tag
+        if: ${{ steps.guard.outputs.enabled == 'true' }}
+        id: rel
+        env:
+          RELEASE_TAG: ${{ needs.release-please.outputs.tag_name }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          scripts/resolve-release-tag.sh "$RELEASE_TAG"
+
      - name: Checkout tap repo
        if: ${{ steps.guard.outputs.enabled == 'true' }}
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
-          repository: Gitlawb/homebrew-tap
+          repository: Twigpine/homebrew-tap
          token: ${{ secrets.HOMEBREW_TAP_PAT }}
          path: tap
          persist-credentials: false
@@ -678,15 +839,32 @@ jobs:
        if: ${{ steps.guard.outputs.enabled == 'true' }}
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          VERSION: ${{ needs.release-please.outputs.version }}
-          TAG: ${{ needs.release-please.outputs.tag_name }}
+          VERSION: ${{ steps.rel.outputs.version }}
+          TAG: ${{ steps.rel.outputs.tag }}
+          ASSETS: ${{ steps.rel.outputs.assets }}
        run: |
          set -euo pipefail
          base="https://github.com/${GITHUB_REPOSITORY}/releases/download/${TAG}"
          mkdir -p _sums
-          gh release download "$TAG" --repo "$GITHUB_REPOSITORY" \
-            --pattern '*.tar.gz.sha256' --dir _sums --clobber
-          sha() { awk '{print $1}' "_sums/gitlawb-node-${VERSION}-$1.tar.gz.sha256"; }
+          # Hash the tarballs themselves, fetched by the asset id the resolver
+          # captured and uploader-checked. The *.sha256 sidecars are mutable
+          # release assets too; trusting them would trust the same surface.
+          sha() {
+            archive="gitlawb-node-${VERSION}-$1.tar.gz"
+            want="$(printf '%s' "$archive" | base64 -w0)"
+            asset_id="$(printf '%s\n' "$ASSETS" | awk -v n="$want" '$1 == n {print $2; exit}')"
+            if [ -z "$asset_id" ]; then
+              echo "::error::release $TAG has no asset $archive captured at resolve time" >&2
+              exit 1
+            fi
+            # errexit is off inside the $(...) this runs under, so a failed
+            # download must exit the function itself instead of hashing a
+            # truncated file.
+            gh api "repos/$GITHUB_REPOSITORY/releases/assets/$asset_id" \
+              -H 'Accept: application/octet-stream' > "_sums/$archive" \
+              || exit 1
+            sha256sum "_sums/$archive" | awk '{print $1}'
+          }
          SHA_MAC_ARM="$(sha aarch64-apple-darwin)"
          SHA_MAC_X64="$(sha x86_64-apple-darwin)"
          SHA_LNX_ARM="$(sha aarch64-unknown-linux-musl)"
@@ -695,8 +873,8 @@ jobs:
          mkdir -p tap/Formula
          cat > tap/Formula/gl.rb <<EOF
          class Gl < Formula
-            desc "Gitlawb CLI — decentralized git for AI agents and developers"
-            homepage "https://gitlawb.com"
+            desc "Twigpine CLI — decentralized git for AI agents and developers"
+            homepage "https://twigpine.com"
            version "${VERSION}"
            license "MIT OR Apache-2.0"

@@ -747,7 +925,7 @@ jobs:
        working-directory: tap
        env:
          HOMEBREW_TAP_PAT: ${{ secrets.HOMEBREW_TAP_PAT }}
-          VERSION: ${{ needs.release-please.outputs.version }}
+          VERSION: ${{ steps.rel.outputs.version }}
        run: |
          set -euo pipefail
          git config user.name "github-actions[bot]"
@@ -760,7 +938,7 @@ jobs:
          fi
          git commit -m "gl ${VERSION}"
          # Credentials are not persisted in .git/config; supply the token only for the push.
-          git push "https://x-access-token:${HOMEBREW_TAP_PAT}@github.com/Gitlawb/homebrew-tap.git" HEAD:main
+          git push "https://x-access-token:${HOMEBREW_TAP_PAT}@github.com/Twigpine/homebrew-tap.git" HEAD:main

  web-sync:
    name: Sync web (install scripts + version)
@@ -769,6 +947,7 @@ jobs:
    runs-on: ubuntu-latest
    permissions:
      contents: read # GITHUB_TOKEN checks out node; the web PR uses WEB_SYNC_PAT
+    environment: release
    steps:
      - name: Guard on secret
        id: guard
@@ -782,11 +961,29 @@ jobs:
            echo "enabled=true" >> "$GITHUB_OUTPUT"
          fi

+      - name: Check out workflow scripts
+        if: ${{ steps.guard.outputs.enabled == 'true' }}
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: Resolve release tag
+        if: ${{ steps.guard.outputs.enabled == 'true' }}
+        id: rel
+        env:
+          RELEASE_TAG: ${{ needs.release-please.outputs.tag_name }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          scripts/resolve-release-tag.sh "$RELEASE_TAG"
+
      - name: Checkout node (release tag)
        if: ${{ steps.guard.outputs.enabled == 'true' }}
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
-          ref: ${{ needs.release-please.outputs.tag_name }}
+          # Pin the commit the resolver verified; a tag ref re-resolved here
+          # could have been moved between resolve and checkout.
+          ref: ${{ steps.rel.outputs.tag_commit || format('refs/tags/{0}', steps.rel.outputs.tag) }}
          path: node
          persist-credentials: false

@@ -794,7 +991,7 @@ jobs:
        if: ${{ steps.guard.outputs.enabled == 'true' }}
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
-          repository: Gitlawb/web
+          repository: Twigpine/web
          token: ${{ secrets.WEB_SYNC_PAT }}
          path: web
          persist-credentials: false
@@ -806,8 +1003,8 @@ jobs:
        if: ${{ steps.guard.outputs.enabled == 'true' }}
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          VERSION: ${{ needs.release-please.outputs.version }}
-          TAG: ${{ needs.release-please.outputs.tag_name }}
+          VERSION: ${{ steps.rel.outputs.version }}
+          TAG: ${{ steps.rel.outputs.tag }}
        run: |
          set -euo pipefail
          gh release view "$TAG" --repo "$GITHUB_REPOSITORY" --json assets \
@@ -825,8 +1022,8 @@ jobs:
        if: ${{ steps.guard.outputs.enabled == 'true' }}
        env:
          GH_TOKEN: ${{ secrets.WEB_SYNC_PAT }}
-          VERSION: ${{ needs.release-please.outputs.version }}
-          TAG: ${{ needs.release-please.outputs.tag_name }}
+          VERSION: ${{ steps.rel.outputs.version }}
+          TAG: ${{ steps.rel.outputs.tag }}
        run: |
          set -euo pipefail
          # Canonical install scripts (source of truth lives in the node repo).
@@ -837,7 +1034,7 @@ jobs:
          # One-time, idempotent drift fixes: drop orphaned binaries, fix stale link.
          rm -rf web/public/bin
          if [ -f web/public/skill.md ]; then
-            sed -i 's#github.com/gitlawb/releases#github.com/Gitlawb/node/releases#g' web/public/skill.md
+            sed -i 's#github.com/gitlawb/releases#github.com/Twigpine/node/releases#g' web/public/skill.md
          fi

          cd web
@@ -851,16 +1048,16 @@ jobs:
          git config user.name "github-actions[bot]"
          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
          git checkout -b "$branch"
-          git commit -m "chore: sync gitlawb ${TAG} (install scripts + version)"
+          git commit -m "chore: sync Twigpine ${TAG} (install scripts + version)"
          # Credentials are not persisted in .git/config; supply the token only for the push.
-          git push --force "https://x-access-token:${GH_TOKEN}@github.com/Gitlawb/web.git" "HEAD:$branch"
-          gh pr create --repo Gitlawb/web --head "$branch" \
-            --title "Sync gitlawb ${TAG}" \
-            --body "Automated sync from Gitlawb/node ${TAG}: install.sh, install.ps1, public/version.json, and one-time drift fixes (removed orphaned public/bin, fixed skill.md releases link). Review and merge to deploy via Vercel." \
+          git push --force "https://x-access-token:${GH_TOKEN}@github.com/Twigpine/web.git" "HEAD:$branch"
+          gh pr create --repo Twigpine/web --head "$branch" \
+            --title "Sync Twigpine ${TAG}" \
+            --body "Automated sync from Twigpine/node ${TAG}: install.sh, install.ps1, public/version.json, and one-time drift fixes (removed orphaned public/bin, fixed skill.md releases link). Review and merge to deploy via Vercel." \
            || {
              # Only swallow the "PR already exists" case; surface auth/API failures
              # (a failed gh pr list must not be misread as "PR exists").
-              pr_count="$(gh pr list --repo Gitlawb/web --head "$branch" --state open --json number --jq 'length')" || exit 1
+              pr_count="$(gh pr list --repo Twigpine/web --head "$branch" --state open --json number --jq 'length')" || exit 1
              if [ "$pr_count" != "0" ]; then
                echo "PR for $branch already exists; branch was updated."
              else
@@ -40,5 +40,5 @@ When a node denies a request, surface the denial to the user. Never render a den
 ## Toolchain and CI reality

 - MSRV is Rust 1.91 (`rust-version` in the workspace `Cargo.toml`). The MSRV CI job runs `cargo check --workspace --all-targets` on exactly 1.91, so check that an API you rely on is stable there; the full test suite runs on stable.
- The PR workflow runs these jobs on every PR: `cargo fmt --check`, `cargo clippy --workspace --all-targets -- -D warnings`, `cargo test --workspace` (against Postgres), a release build, `cargo audit`, the MSRV check, and a Docker build. The `main` branch ruleset currently requires one approving review but does not require these checks to pass before merge; a failing job should block review approval in practice, but nothing enforces that mechanically today.
+- The PR workflow's jobs are `cargo fmt --check`, `cargo clippy --workspace --all-targets -- -D warnings`, `cargo test` (against Postgres), a release build, `cargo audit`, the MSRV check, and a Docker build. On a PR, `scripts/ci-surfaces.sh` runs each job only when the PR touches its inputs, and scopes `cargo test` and the release build to the touched crates plus every crate that depends on them. Root `Cargo.toml`/`Cargo.lock`/`.cargo/` changes run the Rust jobs (fmt, clippy, tests, release build, MSRV) on the whole workspace; every other job still runs only when its own inputs change. Edits to the CI wiring itself (`pr-checks.yml`, `codeql.yml`, `scripts/ci-surfaces.sh`), push to `main`, and the merge queue run every job on the whole workspace. Edits to any other workflow run none of these jobs on a PR; CodeQL only analyzes them as Actions code. Run `cargo test --workspace` locally regardless. The same script limits CodeQL (`codeql.yml`) to the languages a PR touches. The `main` branch ruleset currently requires one approving review but does not require these checks to pass before merge; a failing job should block review approval in practice, but nothing enforces that mechanically today.
 - Conventional commit titles (`feat:`, `fix:`, `docs:`, ...) are required by convention and drive releases, but no CI job checks them; a bad title fails review, not the pipeline.
@@ -1,12 +1,12 @@
-# Contributing to gitlawb
+# Contributing to Twigpine

-Thanks for your interest in contributing. gitlawb is an open protocol — agents and humans welcome.
+Thanks for your interest in contributing. Twigpine is an open protocol — agents and humans welcome.

 ## Quick start for contributors

 ```sh
 # Clone and build
-git clone https://github.com/Gitlawb/node
+git clone https://github.com/Twigpine/node
 cd node

 # Build everything
@@ -38,7 +38,7 @@ docs/                   Operator guides
 scripts/                Build helpers
 ```

-Smart contracts live in a separate repo: [github.com/Gitlawb/contracts](https://github.com/Gitlawb/contracts).
+Smart contracts live in a separate repo: [github.com/Twigpine/contracts](https://github.com/Twigpine/contracts).

 ## Ways to contribute

@@ -81,7 +81,7 @@ time once updated.

 **Requirements:**
 - Rust stable (≥ 1.91) — install via [rustup](https://rustup.rs)
- PostgreSQL — required for the node. Use the bundled `docker-compose.yml` for local dev.
+- PostgreSQL 16+ — required for the node (it uses `pg_input_is_valid`, which exists only on PostgreSQL 16 and later). Use the bundled `docker-compose.yml` for local dev.
 - Docker (optional, for full-stack local testing)

 **Environment variables:**
@@ -20,8 +20,8 @@ members = [
 edition = "2021"
 rust-version = "1.91"
 license = "MIT OR Apache-2.0"
-authors = ["gitlawb contributors"]
-repository = "https://github.com/gitlawb/node"
+authors = ["Twigpine contributors"]
+repository = "https://github.com/Twigpine/node"

 [workspace.dependencies]
 # async
@@ -53,6 +53,11 @@ chrono = { version = "0.4", features = ["serde"] }
 uuid = { version = "1", features = ["v4"] }
 # http client
 reqwest = { version = "0.12", features = ["blocking", "json", "multipart", "rustls-tls"], default-features = false }
+# URL parsing (what reqwest::Url re-exports, so the shared redirect predicate can
+# take a parsed URL without pulling reqwest into gitlawb-core; also used for the
+# absolute-browser-url validation on GITLAWB_WEB_URL in the node and CLI)
+url = "2"
+percent-encoding = "2"
 # HMAC
 hmac = "0.12"

@@ -75,6 +75,7 @@ WORKDIR /data

 ENV GITLAWB_REPOS_DIR=/data/repos \
    GITLAWB_KEY=/data/keys/identity.pem \
+    GITLAWB_P2P_KEY=/data/keys/p2p.key \
    GITLAWB_HOST=0.0.0.0 \
    GITLAWB_PORT=7545 \
    GITLAWB_P2P_PORT=7546
@@ -169,7 +169,7 @@

   END OF TERMS AND CONDITIONS

-   Copyright 2026 gitlawb contributors
+   Copyright 2026 Twigpine contributors

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
@@ -1,6 +1,6 @@
 MIT License

-Copyright (c) 2026 gitlawb contributors
+Copyright (c) 2026 Twigpine contributors

 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
@@ -1,10 +1,10 @@
-# Gitlawb Node
+# Twigpine Node

 **Decentralized git infrastructure for developers, AI agents, and app delivery.**

-Gitlawb Node is the open-source node software behind the Gitlawb network. It lets anyone run a self-hosted node, publish repositories under a DID, sign writes with Ed25519 HTTP signatures, replicate git activity across peers, and move toward a resilient app-delivery network where code and build assets can be served closer to users.
+Twigpine Node is the open-source node software behind the Twigpine network. It lets anyone run a self-hosted node, publish repositories under a DID, sign writes with Ed25519 HTTP signatures, replicate git activity across peers, and move toward a resilient app-delivery network where code and build assets can be served closer to users.

-Gitlawb is not trying to be only "another git host." The long-term direction is:
+Twigpine is not trying to be only "another git host." The long-term direction is:

 ```txt
 Decentralized GitHub
@@ -24,17 +24,17 @@ This is a Rust workspace with six crates:
 | Crate | Purpose |
 |---|---|
 | `gitlawb-node` | The node daemon: Axum HTTP server, git smart-HTTP, Postgres metadata, libp2p gossip, optional S3/Tigris/IPFS/Arweave/Base PoS hooks. |
-| `gl` | The Gitlawb CLI for identity, repos, issues, PRs, bounties, tasks, peers, node status, MCP, and setup flows. |
-| `git-remote-gitlawb` | Git remote helper for `gitlawb://` URLs, so normal `git clone`, `git fetch`, and `git push` can talk to Gitlawb nodes. |
+| `gl` | The Twigpine CLI for identity, repos, issues, PRs, bounties, tasks, peers, node status, MCP, and setup flows. |
+| `git-remote-gitlawb` | Git remote helper for `gitlawb://` URLs, so normal `git clone`, `git fetch`, and `git push` can talk to Twigpine nodes. |
 | `gitlawb-core` | Shared primitives: Ed25519 identities, `did:key`, CIDs, RFC 9421 HTTP signatures, certificates, and UCAN tokens. |
 | `gitlawb-attest` | Pluggable external provenance attestations for ref-update certificates. |
 | `icaptcha-client` | Client for the iCaptcha proof-of-intelligence flow used to protect spam-prone writes. |

 ---

-## Why Gitlawb Nodes?
+## Why Twigpine Nodes?

-Most git hosting today depends on a small number of centralized platforms. Gitlawb Nodes are designed for a different model:
+Most git hosting today depends on a small number of centralized platforms. Twigpine Nodes are designed for a different model:

 - **Own your identity**: every user, agent, and node is an Ed25519 keypair represented as `did:key:z6Mk...`.
 - **Signed writes by default**: write requests use RFC 9421 HTTP Signatures instead of passwords.
@@ -47,12 +47,13 @@ Most git hosting today depends on a small number of centralized platforms. Gitla

 ## Current status

-Gitlawb Node is live early infrastructure. It is useful today, but some security and reliability features are intentionally staged for compatibility with existing nodes.
+Twigpine Node is live early infrastructure. It is useful today, but some security and reliability features are intentionally staged for compatibility with existing nodes.

 Good today:

 - Local or Docker node startup.
 - Postgres-backed repo metadata.
+- Bounded GraphQL queries with [repository pagination](docs/graphql-pagination.md).
 - Bare git repository storage.
 - Git smart-HTTP clone/fetch/push.
 - RFC 9421-signed writes.
@@ -66,12 +67,12 @@ Good today:

 Known limitations:

- Repository write authorization is not secure by default: `GITLAWB_ENFORCE_OWNER_PUSH` defaults to `false` for compatibility, so a valid HTTP Signature identifies a pusher but does not enforce owner-only pushes.
+- Repository write authorization is secure by default: `GITLAWB_ENFORCE_OWNER_PUSH` defaults to `true`, so a valid HTTP Signature identifies a pusher and the node additionally requires that pusher to be the repository owner. Set `GITLAWB_ENFORCE_OWNER_PUSH=false` only for the duration of a rolling upgrade whose pushers are not yet the repo owner. Authentication is not authorization on its own (anyone can mint a `did:key` and sign), so owner enforcement is the only thing that prevents every signed caller from pushing to every repository, private ones included.
 - UCAN proof chains are validated when supplied, but UCAN capabilities are not consulted by write authorization and the root issuer is not independently trust-anchored. UCANs therefore do not yet grant scoped collaborator access.
 - Agent lifecycle revocation is not enforced by HTTP Signature authorization; do not rely on removing or revoking an agent record to block a compromised signer.
- Read visibility is not a blanket data-classification boundary: task, IPFS-pin, and Arweave-anchor listings are not repository-gated; withheld path names can be visible to a root reader; and later visibility changes cannot retract content already announced or externally anchored.
+- Task list/get reads (REST and GraphQL) are repository/task-gated and omit `ucan_token`. Slash-form `repo_id` rows (mirror-hosted tasks) can appear on list for the delegator or trimmed assignee, but `GET /api/v1/tasks/{id}` and GraphQL `task` fail closed for every caller because there is no local repository backing to serve. Read visibility is not a blanket data-classification boundary: IPFS-pin and Arweave-anchor listings are not repository-gated; withheld path names can be visible to a root reader; and later visibility changes cannot retract content already announced or externally anchored.
 - Peer writes are signed by upgraded nodes, but strict signed-peer enforcement is opt-in during rolling upgrades.
- Current GraphQL mutations require an authenticated signer, but there is no mutation-specific guardrail that prevents a future mutation from omitting that check.
+- GraphQL mutations require a signed identity (a verified signer bound to the acting DID), are limited to the agent-task queue, and are covered by a CI completeness fence that requires signer acquisition and actor binding on every registered mutation; repo writes are not exposed over GraphQL and remain REST-only.
 - Pull-request review comments do not yet have threaded line-level anchors, and merges do not enforce approval requirements.

 See:
@@ -87,7 +88,7 @@ See:
 The fastest path is Docker Compose. It starts a node and Postgres.

 ```bash
-git clone https://github.com/Gitlawb/node.git
+git clone https://github.com/Twigpine/node.git
 cd node
 cp .env.example .env
 docker compose up -d
@@ -132,10 +133,10 @@ npm install -g @gitlawb/gl
 brew install gitlawb/tap/gl

 # curl (macOS / Linux)
-curl -fsSL https://gitlawb.com/install.sh | sh
+curl -fsSL https://twigpine.com/install.sh | sh

 # PowerShell (Windows)
-irm https://gitlawb.com/install.ps1 | iex
+irm https://twigpine.com/install.ps1 | iex
 ```

 Or build from source:
@@ -178,7 +179,7 @@ gl register --node http://localhost:7545
 Create a repo:

 ```bash
-gl repo create my-repo --description "My first Gitlawb repo" --node http://localhost:7545
+gl repo create my-repo --description "My first Twigpine repo" --node http://localhost:7545
 ```

 Use the git remote helper:
@@ -192,7 +193,7 @@ For public-network use, make sure `GITLAWB_NODE` points to the node you want. Th

 ### Full lifecycle against an iCaptcha-enforcing node

-Public nodes (e.g. `node.gitlawb.com`) require two things on writes:
+Public nodes (e.g. `node.twigpine.com`) require two things on writes:

 1. **RFC 9421 HTTP Signatures**: every write is signed by your identity key. `gl`
   and the `git-remote-gitlawb` helper do this automatically. An old/unsigned CLI
@@ -205,8 +206,8 @@ Public nodes (e.g. `node.gitlawb.com`) require two things on writes:

 ```bash
 gl identity new                                   # create did:key identity
-gl register      --node https://node.gitlawb.com  # signed + auto-solves iCaptcha
-gl repo create memlawb --node https://node.gitlawb.com   # signed + auto-solves iCaptcha
+gl register      --node https://node.twigpine.com  # signed + auto-solves iCaptcha
+gl repo create memlawb --node https://node.twigpine.com   # signed + auto-solves iCaptcha
 git push  origin2 main                            # origin2 = gitlawb://<your-did>/memlawb (signed)
 git clone gitlawb://<your-did>/memlawb            # public read, no proof needed
 gl doctor                                         # preflight: identity, node, version, iCaptcha
@@ -229,6 +230,53 @@ Notes:
  sends the bearer token **only** to your configured origin, never to a URL a
  node advertises, so a hostile node can't capture the key or redirect the solve.

+### Fetching an object by CID
+
+```bash
+gl ipfs list                                  # CIDs this node has pinned
+gl ipfs get bafkrei... > object.bin           # object bytes on stdout
+```
+
+Objects that were pinned before the node started recording which repo they came
+from are found by scanning its repo inventory, and that scan stops at the
+per-request ceilings in the [Configuration](#configuration) table. A stopped scan
+answers 503 with a resume token instead of a false "not found", and `gl ipfs get`
+follows the token automatically: up to 8 resumes after the first request, so at
+most 9 calls to the node, waiting between attempts for as long as the node's
+`Retry-After` asks and never longer than 5 seconds.
+
+The whole ladder runs under a 60 second wall-clock deadline. The deadline bounds
+the search, not the download: each attempt gets the time left on it to produce
+response headers, and once an object is found its bytes stream outside that
+deadline. They are not unbounded, though. The client's own 30 second HTTP timeout
+is a total request timeout, running from the start of a request until its body
+has finished, so a transfer still going 30 seconds after its request began is cut
+off. Waits between attempts never run past the deadline either, so a single run
+spends at most around 90 seconds on the network: the deadline plus the 30 second
+timeout covering the last attempt. Writing the object out sits outside both
+bounds, so piping into a reader that stops reading can hold the command open
+longer than that.
+
+Two node-side brakes end a ladder early and are reported rather than retried
+around. A 429 is terminal, because the node's rate-limit window is an hour and
+that wait cannot be honored inside one invocation; a transient overload (a 503
+carrying no incomplete-scan code) is retried on the token already held, under the
+same cap, clamp and deadline. The per-IP fanout brake can also stop a ladder well
+short of the 9 calls, so automatic resumption is not a guarantee of reaching the
+object.
+
+When a bound stops the ladder with a usable token in hand, the command prints the
+token and the invocation that continues from it before exiting nonzero:
+
+```txt
+resume from where this stopped: gl ipfs get bafkrei... --scan <token>
+```
+
+Run that to carry on from where the scan stopped. Re-running without `--scan`
+restarts at the first row, reproduces the same truncation and spends the node's
+per-IP budget again, so the token is the only thing that makes progress. Tokens
+are valid for an hour.
+
 ---

 ## Architecture
@@ -264,6 +312,7 @@ metadata        local disk / optional S3
 | DID | A user, agent, or node identity derived from an Ed25519 public key. |
 | HTTP Signature | RFC 9421 signature proving control of the DID key for write requests. |
 | Ref certificate | Signed record of a ref update. Useful for audit and replication. |
+| Ref certificate versioning | The wire format is versioned; new certificates are currently issued as v1. `gl cert show <repo> <id> --verify` requires `--expect-node <did>`, and the queried node's self-reported DID is **not** accepted as a trust anchor. v2 is reserved for a future payload shape that carries the version inside the signed bytes; this client refuses to verify v2+ rather than guessing. |
 | UCAN | Delegation token for future capability-based workflows. |
 | Peer announce | Node-to-node HTTP announcement of DID + public URL. |
 | Gossipsub | libp2p topic for ref-update events. |
@@ -322,6 +371,12 @@ GITLAWB_REQUIRE_SIGNED_PEER_WRITES=true

 `POST /api/v1/sync/trigger` is not part of the staged rollout: it always requires a signature in both config modes and returns 401 without one, because each call drives an O(peers) outbound fan-out.

+The flag is not HTTP-only. It also gates inbound gossip ref-update events on the libp2p mesh, which carry an Ed25519 payload signature rather than RFC 9421 headers because there are no HTTP headers to sign. Set to `true`, an unsigned gossip event is dropped. Set to `false`, an unsigned event is accepted with a warning during the rolling-upgrade window, and the write budget it consumes is charged to the forwarding peer rather than to the DID it claims, since an unsigned event's claimed DID is asserted and not proven. A present-but-invalid signature is refused in both modes; that is forgery, not an un-upgraded peer.
+
+Because the flag now spans both transports, the rollout order matters in one direction: **upgrade every gossip-publishing peer to a build that signs events before you set this to `true`.** Enabling it while an old publisher is still live drops that publisher's ref-updates on arrival, and the publisher sees no error, because gossip has no response to carry one. Upgrading the HTTP peers alone is not sufficient.
+
+There is a second precondition on gossip ingest that the flag does not control. A publisher's `node_did` must already exist in the receiving node's peers table, or the event is dropped as an unknown peer DID. This holds in both modes, signed and unsigned alike: a valid signature proves key possession, not membership. Rows reach that table over HTTP, either from `GITLAWB_BOOTSTRAP_PEERS` when this node contacts a bootstrap peer and records the DID it reports, or from a prior `POST /api/v1/peers/announce`. So a peer that only ever joined the libp2p mesh, with no HTTP announce and no bootstrap contact in either direction, will have its ref-updates dropped even with a good signature and the flag off. If gossip is silently not landing from a peer you can see on the mesh, check that its DID is in `GET /api/v1/peers` on the receiving side first.
+
 ---

 ## Configuration
@@ -344,9 +399,12 @@ Important node settings:
 | `GITLAWB_P2P_PORT` | libp2p QUIC/UDP port. Use `0` to disable. |
 | `GITLAWB_BOOTSTRAP_PEERS` | Comma-separated HTTP peer URLs. |
 | `GITLAWB_P2P_BOOTSTRAP` | Comma-separated libp2p multiaddrs. |
+| `GITLAWB_P2P_KEY` | Path to the node's persistent libp2p identity key file, which fixes the PeerId across restarts. Must name a key file inside a directory, such as `/data/keys/p2p.key` or `./keys/p2p.key`. Two kinds of problem are handled differently. A value that cannot name a securable key file at all (a bare filename, `..` traversal, a trailing `/`, the filesystem root, or a `~/` path that escapes home) is refused before the node binds, so the process exits. Anything wrong with the storage itself (a symlinked or non-directory parent, an unsafe ancestor or working directory, an unreadable or malformed key) leaves p2p off while HTTP keeps serving; the node logs it at error level as `p2p_identity_key_load_failed` and sets the `gitlawb_p2p_identity_key_load_failed` metric to 1. Note that `/health` still reports healthy in that state, so alert on the log event or the metric rather than on the health check. Correcting the storage and restarting restores the same PeerId. On Unix a new key is created `0600` inside a `0700` directory. A key directory that grants access beyond its owner is tightened to `0700` on start; one that is closed too far is refused with a `chmod 700` remedy rather than widened, since a directory locked down on purpose is an operator decision. Rotate the key only when the node logs that the directory was group- or world-writable (or the key file itself was group- or world-readable). An ordinary `0755` directory is tightened too, but others could not read a `0600` key or replace its directory entry. An existing key file is never modified: one with group or other bits set is rejected (`chmod 600`), and so is one this node cannot read. On other platforms no permissions are enforced. Default `~/.gitlawb/p2p.key`; point it at a persistent volume when running in a container. |
 | `GITLAWB_BOOTSTRAP_DISABLE_SEEDS` | Disable embedded seed peers for isolated dev/test networks. |
-| `GITLAWB_REQUIRE_SIGNED_PEER_WRITES` | Require signed peer announce/sync writes. |
+| `GITLAWB_REQUIRE_SIGNED_PEER_WRITES` | Require peer writes to prove their DID on every transport: RFC 9421 signatures on the announce/sync routes, and an Ed25519 payload signature on gossip ref-update events. Defaults to `false` during the staged rollout below. Upgrade every gossip publisher before enabling, or their updates are dropped with no error on their side. |
+| `GITLAWB_ENFORCE_OWNER_PUSH` | Require the authenticated pusher to be the repo owner on `git-receive-pack`. **Defaults to `true`.** A `did:key` signature is authentication, not authorization (anyone can mint a key and sign), so with this off every signed caller may push to every repository, private ones included. Delegated and CI keys count as non-owners: a UCAN `git/push` capability is verified but not yet honored for authorization, so they cannot push while this is on. Set `false` only for a rolling upgrade; see [`docs/RUN-A-NODE.md`](docs/RUN-A-NODE.md). |
 | `GITLAWB_AUTO_SYNC` | Enable automatic sync from known peers. |
+| `GITLAWB_RECONCILIATION_SWEEP` | Enable the hourly durability sweep that re-pins/backstops missing objects (default `true`; disabled when no IPFS/Pinata backend is configured). Public pin repair runs against any configured backend (IPFS, Pinata, or both). Encrypted recovery repair requires local IPFS (`GITLAWB_IPFS_API`); Pinata-only nodes reconcile public pins only. |
 | `GITLAWB_MAX_PACK_BYTES` | Max git pack body size for smart-HTTP routes. |
 | `GITLAWB_GIT_SERVICE_TIMEOUT_SECS` | Max seconds a served git upload-pack, receive-pack, or `info/refs` advertisement may run before it is aborted (504). Default 600. Also bounds the withheld-blob classification walk (on both the upload-pack serve and receive-pack replication paths) and the push-side pin-candidate discovery (`rev-list` / `cat-file`), each reaped via process-group teardown at the deadline. On the path-scoped upload-pack path the classification walk and the pack serve share ONE deadline, so this value bounds their combined duration rather than granting each stage a full budget: a walk that consumes it leaves the serve nothing and the clone gets a 504. Serving large path-scoped repos may therefore need a higher value than they did when each stage was budgeted separately. Accepted range is 1 to 3153600000 (100 years), since the node derives deadlines from this value and a larger one cannot be represented. |
 | `GITLAWB_GIT_ACQUIRE_TIMEOUT_SECS` | Max seconds the storage-acquisition phase (Tigris HEAD/GET, push advisory-lock) of a served git op may run before the request is shed with a 503, separate from the git-run timeout. The concurrency permit is released on expiry so a stalled backend cannot pin the pool. Default 30. |
@@ -357,21 +415,49 @@ Important node settings:
 | `GITLAWB_REPO_LEASE_MAX_WAITERS` | Max pushes parked at once waiting for the same repo's write lease. Each waiter pins its buffered pack body, so this bounds that memory for a hot repo; past the cap the newest push sheds a 503 + Retry-After instead of queueing. Pushes to other repos are unaffected, and the lease holder is not counted. Default 8. |
 | `GITLAWB_MAX_CONCURRENT_IPFS_WALKS` | Max concurrent `GET /ipfs/{cid}` visibility walks across all callers (own pool, disjoint from the served-git pools); over-cap sheds 503. Default 32. |
 | `GITLAWB_IPFS_WALK_PER_SOURCE` | Max concurrent `/ipfs` walks a single source IP may hold. Default 4. |
-| `GITLAWB_IPFS_MAX_REPOS_WALKED` | Max expensive path-scope visibility walks per `/ipfs/{cid}` request; over-cap repos are skipped and the scan continues, shedding a retryable 503 (not a false 404) if the object is then found nowhere. Default 64. |
+| `GITLAWB_IPFS_MAX_LEGACY_PROBES` | Max legacy (NULL-provenance) repos probed per `/ipfs/{cid}` request, bounding the scan-fallback fan-out. A truncated scan returns a retryable 503, not a false 404. Default 256. |
+| `GITLAWB_IPFS_MAX_LEGACY_SCAN_ROWS` | Max repo rows one `/ipfs/{cid}` request's legacy scan may read from the database. The probe ceiling above only starts counting once a probe runs, and quarantined or private repos are denied before that, so this is what bounds a scan over an inventory that denies the caller everywhere. A truncated scan sheds a retryable 503 carrying an opaque `continuation` token; echoing it as `?scan=` resumes the scan where it stopped. Every per-request ceiling on this path (rows, probes, visits, retained rule bytes) mints one, so each request advances the ladder by at least the rows it read and a holder buried past a ceiling is reached in a bounded number of requests, `ceil(repos / ceiling) + 1` when this row ceiling is the one that binds. No ceiling ever produces a 404. Every page is charged to the caller's `/ipfs` work allowance, so raising this raises that allowance too. Lowering it sharpens an oracle: because a truncation emits a token and a completed wrap does not, laddering to the end reveals the node's total repo count (private and quarantined included) to within one ceiling. Default 2048. |
+| `GITLAWB_IPFS_MAX_REPOS_WALKED` | Max expensive path-scope visibility walks a `/ipfs/{cid}` request may run per phase; over-cap repos are skipped and the scan continues, shedding a retryable 503 (not a false 404) if the object is then found nowhere. The effective cap is the tighter of this knob and the node's internal history-walk ceiling of 17 (`MAX_PIN_SOURCES + 1`), so a value above 17 has no effect while a value below it does tighten the cap. It is charged per phase: the provenance lookup and the legacy-scan fallback get separate equal budgets, so one request can run up to twice the cap in total. Default 64. |
 | `GITLAWB_IPFS_MAX_REPO_VISITS` | Ceiling on repos one `/ipfs/{cid}` request may visit (acquire + probe) past the visibility gate. Also the worst-case per-request Tigris fetch count. On exhaustion the scan stops with a retryable 503. Default 1024. |
 | `GITLAWB_IPFS_REQUEST_BUDGET_SECS` | Absolute wall-clock budget for one admitted `/ipfs/{cid}` request's acquire+walk lifetime. Per-stage clamps bound the acquire and walk stages to the remaining budget, and no stage starts once it is exhausted; the scan then stops with a retryable 503. The object-type probe and content-read `cat-file` subprocesses are budget-checked before starting and each also run under their own deadline (the lesser of `GITLAWB_GIT_SERVICE_TIMEOUT_SECS` and the remaining budget), reaped via process-group teardown, so a hung `cat-file` cannot hold the request's walk slot past it. One hang path is still unbounded: the probe's object-store readability check is a plain filesystem sweep with nothing to reap, so a wedged filesystem can hold the slot past the deadline. Default 600. Accepted range is 1 to 3153600000 (100 years), since the node derives a deadline from this value and a larger one cannot be represented. |
+| `GITLAWB_IPFS_RESOLVE_BUDGET_SECS` | Shorter budget for the pre-walk CID resolve inside an admitted `/ipfs/{cid}` request: the lookup that maps the requested CID to its git oid(s), which runs while the scarce walk admission is already held. A well-formed CID with no pin row does no probe and no walk work, so without this it could hold a walk slot for the whole request budget while nothing walked, and enough such requests shed every real retrieval at admission. The effective deadline is the lesser of this and the remaining request budget, so a value above `GITLAWB_IPFS_REQUEST_BUDGET_SECS` degrades to the request budget. Only the resolve is on this clock; walk and probe work stay on the request budget, so a slow but progressing scan is never shed by it. Default 10. Accepted range is 1 to 3153600000 (100 years). |
 | `GITLAWB_IPFS_RATE_LIMIT` | Max `/ipfs/{cid}` requests per client IP per hour (route flood brake). 0 disables. Default 600. |
+| `GITLAWB_TASK_READ_RATE_LIMIT` | Max anonymous task-read requests (`GET /api/v1/tasks`, `GET /api/v1/tasks/{id}`) per client IP per hour, sharing the budget with the GraphQL/WS task-read brake. Uses `GITLAWB_TRUSTED_PROXY` to resolve the client IP. 0 disables. Default 1200 (above the `/ipfs` budget to allow list pages followed by per-task reads). |
 | `GITLAWB_TIGRIS_BUCKET` | Optional S3/Tigris shared repo storage bucket. |
 | `GITLAWB_PINATA_JWT` | Optional Pinata/IPFS warm-storage pinning. |
 | `GITLAWB_IRYS_URL` | Optional Irys/Arweave permanent anchoring. |

 Production note: change the default Postgres password before exposing a node publicly.

+Legacy-pin window: releases before the CID-resolver work stored the provider CID (Kubo dag-pb / Pinata) as a pinned object's resolver key. The `/ipfs/{cid}` resolver now recomputes the raw-content CID from the object bytes and refuses to serve a key that does not match, so `GET /api/v1/ipfs/pins` can still advertise an unrepaired legacy CID that 404s. Such a row is repaired opportunistically the next time a push carries the object again (its key is rewritten to the raw CID, the old value kept in `legacy_provider_cid`), but git negotiation omits objects the node already has, so most legacy rows never re-enter a push delta. A deferred one-shot startup sweep, not this opportunistic path, is what fully retires the advertise-then-404 window. Rows whose object bytes are gone stay withheld.
+
+### Upgrading: the PeerId rotates once
+
+This node's libp2p identity is now a keypair generated on first start and kept
+at `GITLAWB_P2P_KEY`, rather than one derived from the node DID. Every node
+therefore gets a new PeerId once, on the first start after upgrading, and keeps
+it from then on as long as that key file survives (put it on a persistent volume
+in a container).
+
+Two things to check before upgrading:
+
+- Any `GITLAWB_P2P_BOOTSTRAP` multiaddr that pins a peer's old PeerId with a
+  `/p2p/<PeerId>` suffix stops matching once that peer upgrades. Update the
+  suffix, or drop it and let identify supply the current one. Addresses without
+  the suffix keep working untouched.
+- `GITLAWB_P2P_KEY` must name a key file inside a directory, such as
+  `/data/keys/p2p.key` or `./keys/p2p.key`. A bare filename is refused at
+  startup, since the node will not keep its identity key in the working
+  directory.
+
+Peers found over `GITLAWB_BOOTSTRAP_PEERS` and the embedded seed list are
+unaffected, since those are HTTP URLs and carry no PeerId.
+
 ---

 ## Optional node staking

-Gitlawb Node includes optional Base L2 node-operator hooks. Operators can register a node DID, stake `$GITLAWB`, and post heartbeats.
+Twigpine Node includes optional Base L2 node-operator hooks. Operators can register a node DID, stake `$GITLAWB`, and post heartbeats.

 PoS is disabled unless these are configured:

@@ -441,8 +527,8 @@ Build:
 Output:

 ```txt
-dist/Gitlawb Node.app
-dist/Gitlawb Node.dmg
+dist/Twigpine Node.app
+dist/GitlawbNode-macOS.dmg
 ```

 Features:
@@ -456,7 +542,7 @@ Features:
 Unsigned local build:

 ```bash
-xattr -cr "dist/Gitlawb Node.app"
+xattr -cr "dist/Twigpine Node.app"
 ```

 ---
@@ -471,7 +557,7 @@ Short-term priorities:
 2. Add Docker and installer smoke tests.
 3. Improve operator docs and `gl doctor` checks.
 4. Harden peer writes and publish the signed-peer rollout plan.
-5. Close default-open write authorization and wire trusted UCAN delegation into repository permissions.
+5. Wire trusted UCAN delegation into repository write authorization.
 6. Add threaded, line-level pull-request discussions and enforce approval requirements on merges.
 7. Add metrics for pushes, fetches, pack sizes, peer sync, failed auth, and webhooks.

@@ -2,9 +2,9 @@

 ## Reporting a Vulnerability

-If you discover a security vulnerability in gitlawb, please **do not open a public issue**.
+If you discover a security vulnerability in Twigpine, please **do not open a public issue**.

-Report it privately by emailing **security@gitlawb.com** with:
+Report it privately by emailing **security@twigpine.com** with:
 - A description of the vulnerability
 - Steps to reproduce
 - Potential impact assessment
@@ -26,7 +26,7 @@ We will acknowledge receipt within 48 hours and aim to release a fix within 14 d

 **Content addressing**
 - Every git object is content-addressed via CIDv1 (SHA-256)
- Tamper-evident by construction — a modified object changes its CID
+- Tamper-evident by construction: a modified object changes its CID

 **UCAN token validation**
 - Bootstrap UCAN tokens are issued at registration.
@@ -35,8 +35,8 @@ We will acknowledge receipt within 48 hours and aim to release a fix within 14 d
 - Capability grants are not yet consulted by repository write authorization; see the limitations below.

 **Smart contracts (Base Sepolia testnet)**
- `GitlawbDIDRegistry` — on-chain DID → document registry
- `GitlawbNameRegistry` — human name → DID registry
+- `GitlawbDIDRegistry`: on-chain DID → document registry
+- `GitlawbNameRegistry`: human name → DID registry
 - Both auditable on-chain, no admin keys

 ---
@@ -54,9 +54,9 @@ We will acknowledge receipt within 48 hours and aim to release a fix within 14 d
 These are documented limitations of the current live release. They should be prioritized without breaking existing nodes during rolling upgrades.

 ### Repository write authorization defaults
- `git-receive-pack` verifies HTTP Signatures, but `GITLAWB_ENFORCE_OWNER_PUSH` defaults to `false` for compatibility during rollout.
- **Impact:** With the default setting, a valid signature authenticates the pusher but does not require that DID to be the repository owner.
- **Mitigation:** Set `GITLAWB_ENFORCE_OWNER_PUSH=true` on nodes where owner-only pushes are required. Confirm that every legitimate pusher uses the owner DID before enabling it.
+- `git-receive-pack` verifies HTTP Signatures and, with the default configuration, also requires the pusher DID to be the repository owner: `GITLAWB_ENFORCE_OWNER_PUSH` defaults to `true`.
+- **Impact:** With the default setting, a valid signature authenticates the pusher, and the pusher must be the repository owner. Authentication is not authorization on its own, since anyone can mint a `did:key` and sign, so owner enforcement is the only thing that prevents every signed caller from pushing to every repository, including private ones.
+- **Mitigation:** `GITLAWB_ENFORCE_OWNER_PUSH=false` is the explicit rolling-upgrade compatibility override. Disable it only for the duration of an upgrade whose pushers are not yet the repo owner, then turn it back on. Delegated and CI keys count as non-owners: a UCAN `git/push` capability is verified but not yet honored for authorization, so they cannot push while this is on.

 ### UCAN delegation and revocation
 - The middleware validates a supplied UCAN's complete proof chain, but a root token is accepted without an independently trusted issuer anchor. `Ucan::can` is not yet used by write handlers, so a UCAN does not grant scoped repository access.
@@ -73,7 +73,7 @@ These are documented limitations of the current live release. They should be pri
 ### Private repository reads
 - Repository and path-scoped visibility checks are enforced for repository API and Git content reads. A denied whole-repository or root read returns the same 404 shape as a missing repository, so the denial does not reveal private-repository existence.
 - Sparse-clone support exposes withheld path globs to callers who may read the repository root. Do not put sensitive information in withheld path names.
- `GET /api/v1/tasks`, `/api/v1/ipfs/pins`, and `/api/v1/arweave/anchors` are not repository-gated. Task records include a UCAN token; pin and anchor listings expose object and ref metadata.
+- Task list/get reads (`GET /api/v1/tasks`, `GET /api/v1/tasks/{id}`, and GraphQL `tasks` / `task` over HTTP and WebSocket) are repository/task-gated through shared visibility collectors and omit `ucan_token`. Slash-form `repo_id` rows (mirror-hosted tasks) can appear on list for the delegator or trimmed assignee, but `GET /api/v1/tasks/{id}` and GraphQL `task` fail closed for every caller because there is no local repository backing to serve. All of those transports share the same per-IP `GITLAWB_TASK_READ_RATE_LIMIT` budget. `/api/v1/ipfs/pins` and `/api/v1/arweave/anchors` are not repository-gated; pin and anchor listings expose object and ref metadata.
 - Changing a repository's visibility controls future serving, but cannot retract ref metadata or configured external pins and anchors already announced while the repository was public. Do not push secrets to an announceable repository.
 - **Impact:** Visibility policies protect the repository and Git content routes they gate, not every metadata endpoint or previously published content.
 - **Remaining boundary:** This read control does not address the independent write-authorization and UCAN-delegation limitations described above.
@@ -118,13 +118,13 @@ These are documented limitations of the current live release. They should be pri

 ## Threat Model

-gitlawb is designed to be secure against:
- **Unauthenticated writes** — HTTP Signature auth on protected write endpoints; see Known Limitations for owner and capability authorization gaps
- **Tampered git objects** — CIDv1 content addressing detects modification
- **Identity spoofing** — DIDs derived from public keys, unforgeable without the private key
- **Centralized takedown** — no single point of control; data on IPFS + Arweave
+Twigpine is designed to be secure against:
+- **Unauthenticated writes:** HTTP Signature auth on protected write endpoints; see Known Limitations for owner and capability authorization gaps
+- **Tampered git objects:** CIDv1 content addressing detects modification
+- **Identity spoofing:** DIDs derived from public keys, unforgeable without the private key
+- **Centralized takedown:** no single point of control; data on IPFS + Arweave

-gitlawb is **not yet** designed to defend against:
+Twigpine is **not yet** designed to defend against:
 - A compromised node operator (node operators are trusted for their own node)
 - Sybil attacks on the DHT (trust score system mitigates, not eliminates)
 - Timing attacks on signature verification (not constant-time compared in v0.1)
@@ -1,11 +1,11 @@
 {
-  "$comment": "Canonical seed list for the Gitlawb network. Merged with GITLAWB_BOOTSTRAP_PEERS at startup. PRs to add public nodes welcome.",
+  "$comment": "Canonical seed list for the Twigpine network. Merged with GITLAWB_BOOTSTRAP_PEERS at startup. PRs to add public nodes welcome.",
  "version": 1,
-  "updated": "2026-06-30",
+  "updated": "2026-10-02",
  "peers": [
    {
      "name": "gitlawb",
-      "operator": "Gitlawb",
+      "operator": "Twigpine",
      "did": "did:key:z6MkqDnb7Siv3Cwj7pGJq4T5EsUisECqR8KpnDLwcaZq5TPr",
      "http_url": "https://node.gitlawb.com",
      "p2p_multiaddr": null,
@@ -13,7 +13,7 @@
    },
    {
      "name": "gitlawb-node2",
-      "operator": "Gitlawb",
+      "operator": "Twigpine",
      "did": null,
      "http_url": "https://node2.gitlawb.com",
      "p2p_multiaddr": null,
@@ -21,7 +21,7 @@
    },
    {
      "name": "gitlawb-node3",
-      "operator": "Gitlawb",
+      "operator": "Twigpine",
      "did": null,
      "http_url": "https://node3.gitlawb.com",
      "p2p_multiaddr": null,
@@ -29,7 +29,7 @@
    },
    {
      "name": "gitlawb-manila",
-      "operator": "Gitlawb",
+      "operator": "Twigpine",
      "did": null,
      "http_url": "https://manila.gitlawb.com",
      "p2p_multiaddr": null,
@@ -42,6 +42,14 @@
      "http_url": "https://rapybus-gitlawb.com",
      "p2p_multiaddr": null,
      "added": "2026-05-16"
+    },
+    {
+      "name": "moonwake-ledger",
+      "operator": "Moonwake Ledger",
+      "did": "did:key:z6Mkh9AZ46Z3RkGZfFCom5vzX2535f54d6G2o2ZGkCqjwW8K",
+      "http_url": "https://node.moonwakeledger.com",
+      "p2p_multiaddr": null,
+      "added": "2026-10-02"
    }
  ]
 }
@@ -11,7 +11,7 @@ name = "git-remote-gitlawb"
 path = "src/main.rs"

 [dependencies]
-gitlawb-core = { path = "../gitlawb-core" }
+gitlawb-core = { path = "../gitlawb-core", features = ["redirect"] }
 anyhow = { workspace = true }
 reqwest = { workspace = true }
 tracing = { workspace = true }
@@ -199,9 +199,7 @@ fn handle_connect<R: Read>(
        other => bail!("unsupported git service: {other}"),
    }

-    let client = reqwest::blocking::Client::builder()
-        .timeout(std::time::Duration::from_secs(300))
-        .build()?;
+    let client = build_http_client()?;

    // ── Phase 1: ref advertisement (GET /info/refs?service=<service>) ─────────
    //
@@ -289,8 +287,10 @@ fn handle_connect<R: Read>(
    //    this into one POST deadlocks a multi-round fetch (#117).
    //
    //  git-receive-pack (push):
-    //    Git sends ref-update commands + the complete PACK blob, then closes its
-    //    write pipe. read_to_end is safe and correct here, a single POST.
+    //    Git sends the ref-update command list (pkt-lines ending in a flush),
+    //    then the complete PACK blob only when a command has a non-zero new-oid.
+    //    A delete-only push sends no pack, and git holds its write pipe open for
+    //    the report-status response, so EOF cannot delimit the request (#369).

    let post_url = format!("{}/{}", repo_base, service);

@@ -298,10 +298,7 @@ fn handle_connect<R: Read>(
        return negotiate_upload_pack(&client, &post_url, service, signing_key, stdin, &mut stdout);
    }

-    let mut request_body = Vec::new();
-    stdin
-        .read_to_end(&mut request_body)
-        .context("reading receive-pack request")?;
+    let request_body = read_receive_pack_request(stdin)?;
    tracing::debug!("pack request: {} bytes from git", request_body.len());

    if request_body.is_empty() {
@@ -320,6 +317,58 @@ fn handle_connect<R: Read>(
    )
 }

+// ── HTTP client ───────────────────────────────────────────────────────────────
+
+/// Total request timeout. A pack transfer can be large and slow, so this is far
+/// wider than the CLI's.
+const HTTP_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(300);
+
+/// The one client both phases use, with the redirect policy the signing surfaces
+/// need.
+///
+/// Every request this client sends can carry RFC 9421 `Signature` and
+/// `Signature-Input` headers: a push signs from the first request, and a fetch of a
+/// private repo signs on the retry. reqwest strips only `Authorization`, `Cookie`,
+/// `Proxy-Authorization` and `WWW-Authenticate` across hosts, so under the default
+/// `Policy::limited(10)` those signature headers rode a 302 to whatever origin the
+/// node named, and on a 307/308 the pack body went with them. Scope the follow to the
+/// origin that issued the redirect AND to an identical request-target, which is the
+/// same predicate `gl` uses.
+fn build_http_client() -> Result<reqwest::blocking::Client> {
+    Ok(reqwest::blocking::Client::builder()
+        .timeout(HTTP_TIMEOUT)
+        .redirect(reqwest::redirect::Policy::custom(same_origin_redirect))
+        .build()?)
+}
+
+/// Refuse any redirect that leaves the issuing origin or rewrites the request-target,
+/// and bound the chain.
+///
+/// Refusal is `stop`, not `error`: the 3xx comes back as an ordinary response and
+/// the caller reports it through the status path it already has.
+///
+/// `Policy::custom` replaces reqwest's built-in limit, so the chain bound is
+/// restated. It is not what makes the request finite (`HTTP_TIMEOUT` covers the whole
+/// chain); it is what keeps a node that redirects to itself from costing a request
+/// per round trip until that timeout. `>` and not `>=` because reqwest pushes the
+/// redirecting URL onto `previous` before consulting the policy, which is how
+/// `Policy::limited` reads the same counter.
+fn same_origin_redirect(attempt: reqwest::redirect::Attempt<'_>) -> reqwest::redirect::Action {
+    let Some(previous) = attempt.previous().last() else {
+        // Unreachable through reqwest, which pushes the redirecting URL first, but
+        // the safe reading of "cannot prove same-origin" is to refuse.
+        return attempt.stop();
+    };
+    if attempt.previous().len() > gitlawb_core::redirect::MAX_REDIRECTS {
+        return attempt.stop();
+    }
+    if gitlawb_core::redirect::may_follow(previous, attempt.url()) {
+        attempt.follow()
+    } else {
+        attempt.stop()
+    }
+}
+
 // ── Smart-protocol request builders ───────────────────────────────────────────

 const USER_AGENT: &str = "git/2.0 git-remote-gitlawb/0.1.0";
@@ -485,6 +534,108 @@ fn read_upload_pack_round<R: Read>(stdin: &mut R) -> Result<(Vec<u8>, RoundEnd)>
    }
 }

+/// Read one git-receive-pack request: optional `shallow` lines, then EITHER a
+/// command list (pkt-lines terminated by a flush) or a push-cert block
+/// (pkt-lines terminated by a "push-cert-end" line), then the negotiated
+/// push-options block, then the pack ONLY when a command carries a non-zero
+/// new-oid.
+///
+/// A delete-only push sends no pack, and git holds its write pipe open waiting
+/// for the report-status response, so a plain read_to_end deadlocks the two
+/// (#369). A signed push is the other trap: when the client sends a push-cert
+/// it sends NO command list at all (the commands live inside the cert), so the
+/// request starts with a "push-cert" line and no flush ever arrives. The
+/// negotiated capability list is on the push-cert line, or on the first
+/// command line after its NUL when there is no cert. EOF at any point also
+/// ends the read, covering git versions that close the pipe early.
+fn read_receive_pack_request<R: Read>(stdin: &mut R) -> Result<Vec<u8>> {
+    /// Read one pkt-line into `body` (length prefix + data), returning its data
+    /// or `None` on flush or EOF. Malformed lengths bail rather than collapse
+    /// into a synthesized flush (#192 F5).
+    fn read_pkt<R: Read>(stdin: &mut R, body: &mut Vec<u8>) -> Result<Option<Vec<u8>>> {
+        let mut len_bytes = [0u8; 4];
+        match stdin.read_exact(&mut len_bytes) {
+            Ok(()) => {}
+            Err(e) if e.kind() == io::ErrorKind::UnexpectedEof => return Ok(None),
+            Err(e) => return Err(e.into()),
+        }
+        let Ok(len_hex) = std::str::from_utf8(&len_bytes) else {
+            bail!("malformed pkt-line length prefix: non-UTF-8 bytes {len_bytes:02x?}");
+        };
+        let Ok(pkt_len) = usize::from_str_radix(len_hex, 16) else {
+            bail!("malformed pkt-line length prefix: non-hex {len_hex:?}");
+        };
+        body.extend_from_slice(&len_bytes);
+        if pkt_len == 0 {
+            return Ok(None);
+        }
+        if pkt_len < 4 {
+            bail!("invalid pkt-line length: {pkt_len}");
+        }
+        let mut data = vec![0u8; pkt_len - 4];
+        stdin
+            .read_exact(&mut data)
+            .context("reading pkt-line data")?;
+        body.extend_from_slice(&data);
+        Ok(Some(data))
+    }
+
+    /// "<40-hex old> <40-hex new> <refname>" — the command shape, whether it
+    /// appears in a command list or inside a push-cert.
+    fn is_command(data: &[u8]) -> bool {
+        data.len() >= 82
+            && data[..40].iter().all(|b| b.is_ascii_hexdigit())
+            && data[40] == b' '
+            && data[41..81].iter().all(|b| b.is_ascii_hexdigit())
+            && data[81] == b' '
+    }
+
+    let mut body = Vec::new();
+    let mut caps = String::new();
+    let mut pack_expected = false;
+    let mut in_cert = false;
+    loop {
+        let Some(data) = read_pkt(stdin, &mut body)? else {
+            // Flush ends a command list; EOF ends anything. A cert has no
+            // legal flush inside it, so the same break covers malformed
+            // input without hanging.
+            break;
+        };
+        if !in_cert && data.starts_with(b"push-cert") {
+            in_cert = true;
+            if let Some(nul) = data.iter().position(|&b| b == 0) {
+                caps = String::from_utf8_lossy(&data[nul + 1..]).into_owned();
+            }
+            continue;
+        }
+        if in_cert && data == b"push-cert-end\n" {
+            break;
+        }
+        if is_command(&data) {
+            if caps.is_empty() {
+                if let Some(nul) = data.iter().position(|&b| b == 0) {
+                    caps = String::from_utf8_lossy(&data[nul + 1..]).into_owned();
+                }
+            }
+            if data[41..81].iter().any(|&b| b != b'0') {
+                pack_expected = true;
+            }
+        }
+    }
+    // When both are negotiated the push-options block follows the cert; when
+    // only options are negotiated it follows the command flush. Either way it
+    // is pkt-lines to a flush.
+    if caps.split(' ').any(|c| c.starts_with("push-options")) {
+        while read_pkt(stdin, &mut body)?.is_some() {}
+    }
+    if pack_expected {
+        stdin
+            .read_to_end(&mut body)
+            .context("reading receive-pack pack")?;
+    }
+    Ok(body)
+}
+
 /// POST one self-contained stateless-RPC request body and stream the response to
 /// `stdout`. Signs the request when `signing_key` is present, so every round of a
 /// private-repo fetch carries the caller's signature (the node visibility-gates
@@ -858,6 +1009,438 @@ mod tests {
        String::from_utf8_lossy(&buf).into_owned()
    }

+    /// The signed headers must not survive a redirect off the node's origin, on
+    /// EITHER phase.
+    ///
+    /// This is the binary git runs for `clone`, `fetch` and `push`. It built its
+    /// client with a timeout and nothing else, so it ran reqwest's default
+    /// `Policy::limited(10)`, and reqwest's cross-host header stripping covers
+    /// `Authorization`, `Cookie`, `Proxy-Authorization` and `WWW-Authenticate` only.
+    /// `Signature` and `Signature-Input` came straight through. A push signs from the
+    /// first request, so a hostile node answering 302 was handed a working credential,
+    /// and a 307/308 would have taken the pack body along.
+    ///
+    /// Two mockito servers are two ports on one host, which is the boundary this
+    /// policy draws. The second server answers everything and expects nothing, with a
+    /// second mock matching on the `signature` header also at zero, so a followed
+    /// redirect fails whether or not the signature came with it. Phase 1 is driven
+    /// with a 302 and Phase 2 with a 308, the status that would carry the body.
+    ///
+    /// MUTATION (RED): drop the `.redirect(...)` line from `build_http_client`.
+    #[test]
+    fn signed_requests_do_not_follow_a_redirect_off_the_node_origin() {
+        let kp = Keypair::generate();
+        let client = build_http_client().unwrap();
+
+        let mut elsewhere = mockito::Server::new();
+        let never = elsewhere
+            .mock("GET", mockito::Matcher::Any)
+            .with_status(200)
+            .with_body("bytes from the redirect target")
+            .expect(0)
+            .create();
+        let never_post = elsewhere
+            .mock("POST", mockito::Matcher::Any)
+            .with_status(200)
+            .expect(0)
+            .create();
+        let signature_seen_get = elsewhere
+            .mock("GET", mockito::Matcher::Any)
+            .match_header("signature", mockito::Matcher::Any)
+            .with_status(200)
+            .expect(0)
+            .create();
+        let signature_seen_post = elsewhere
+            .mock("POST", mockito::Matcher::Any)
+            .match_header("signature", mockito::Matcher::Any)
+            .with_status(200)
+            .expect(0)
+            .create();
+
+        let mut node = mockito::Server::new();
+        let repo_base = format!("{}/zOwner/myrepo", node.url());
+        let bounce_get = node
+            .mock("GET", mockito::Matcher::Regex(r"/info/refs".to_string()))
+            .with_status(302)
+            .with_header("location", &format!("{}/info/refs", elsewhere.url()))
+            .expect(1)
+            .create();
+        let bounce_post = node
+            .mock(
+                "POST",
+                mockito::Matcher::Regex(r"/git-receive-pack$".to_string()),
+            )
+            .with_status(308)
+            .with_header("location", &format!("{}/git-receive-pack", elsewhere.url()))
+            .expect(1)
+            .create();
+
+        let refs_url = format!("{repo_base}/info/refs?service=git-receive-pack");
+        let advertisement = build_advertisement_request(&client, &refs_url, Some(&kp))
+            .send()
+            .unwrap();
+
+        let body = b"0009done\n".to_vec();
+        let post_url = format!("{repo_base}/git-receive-pack");
+        let pack_post =
+            build_pack_post_request(&client, &post_url, "git-receive-pack", &body, Some(&kp))
+                .body(body.clone())
+                .send()
+                .unwrap();
+
+        // The other origin first: it is the assertion that names the leak, and it must
+        // be the one that speaks when a followed redirect makes every one of these
+        // fail at once.
+        never.assert();
+        never_post.assert();
+        signature_seen_get.assert();
+        signature_seen_post.assert();
+        bounce_get.assert();
+        bounce_post.assert();
+
+        assert_eq!(
+            advertisement.status(),
+            302,
+            "a refused redirect stops rather than errors, so the caller sees the 3xx"
+        );
+        assert!(
+            !advertisement
+                .text()
+                .unwrap()
+                .contains("bytes from the redirect target"),
+            "the redirect target's bytes must never reach the caller"
+        );
+        assert_eq!(
+            pack_post.status(),
+            308,
+            "the pack POST stops at the redirect too"
+        );
+    }
+
+    /// A same-origin hop that REWRITES the request-target is refused, even though the
+    /// origin never changes.
+    ///
+    /// The signature binds `@path` as the literal path-and-query the helper signed, and
+    /// the node rebuilds it from the URI it received. An `info/refs` to `info/refs/`
+    /// bounce therefore presents a signature over a target the node never saw, so a
+    /// clone or push behind such a proxy 401s. Refusing the hop hands the caller the
+    /// 3xx that names what happened instead.
+    ///
+    /// The target mock expects zero hits and is asserted: mockito only checks
+    /// `.expect(N)` when `.assert()` runs, so an unbound or unasserted mock passes
+    /// vacuously.
+    #[test]
+    fn a_same_origin_path_changing_redirect_is_refused() {
+        let kp = Keypair::generate();
+        let client = build_http_client().unwrap();
+
+        let mut node = mockito::Server::new();
+        let bounce = node
+            .mock("GET", "/zOwner/myrepo/info/refs")
+            .with_status(301)
+            .with_header("location", "/zOwner/myrepo/info/refs/")
+            .expect(1)
+            .create();
+        let target = node
+            .mock("GET", "/zOwner/myrepo/info/refs/")
+            .with_status(200)
+            .with_body("bytes from the rewritten target")
+            .expect(0)
+            .create();
+
+        let refs_url = format!("{}/zOwner/myrepo/info/refs", node.url());
+        let resp = build_advertisement_request(&client, &refs_url, Some(&kp))
+            .send()
+            .unwrap();
+
+        assert_eq!(
+            resp.status(),
+            301,
+            "a refused redirect stops rather than errors, so the caller sees the 3xx"
+        );
+        assert!(
+            !resp
+                .text()
+                .unwrap()
+                .contains("bytes from the rewritten target"),
+            "the rewritten target's bytes must never reach the caller"
+        );
+        bounce.assert();
+        target.assert();
+    }
+
+    /// The other direction, and the only same-origin hop still followed: a redirect
+    /// that re-issues the IDENTICAL request-target.
+    ///
+    /// Without this the policy could be tightened to `Policy::none()` and nothing in
+    /// this crate would notice. It is also the first executed coverage of the chain
+    /// bound in `build_http_client`'s policy closure: the route answers 301 pointing at
+    /// itself, so bounded it is hit once for the original request plus `MAX_REDIRECTS`
+    /// follows and the call returns the 301 (a refused hop stops rather than errors).
+    /// Unbounded it would spin until `HTTP_TIMEOUT`, costing the node a request per
+    /// round trip. Its twin is `a_self_redirect_stops_at_the_chain_bound` in
+    /// `crates/gl/src/http.rs`.
+    #[test]
+    fn an_identical_target_redirect_is_still_followed_up_to_the_chain_bound() {
+        let kp = Keypair::generate();
+        let client = build_http_client().unwrap();
+
+        let mut node = mockito::Server::new();
+        let loop_route = node
+            .mock("GET", "/zOwner/myrepo/info/refs")
+            .with_status(301)
+            .with_header("location", "/zOwner/myrepo/info/refs")
+            .expect(gitlawb_core::redirect::MAX_REDIRECTS + 1)
+            .create();
+
+        let refs_url = format!("{}/zOwner/myrepo/info/refs", node.url());
+        let resp = build_advertisement_request(&client, &refs_url, Some(&kp))
+            .send()
+            .expect("the bound must end the chain, not the timeout");
+
+        assert_eq!(
+            resp.status(),
+            301,
+            "the chain ends by refusing the next hop, so the last 3xx is what comes back"
+        );
+        loop_route.assert();
+    }
+
+    // ── the node's own verification, run over what the helper actually sent ──
+
+    /// What the verifying mock made of a request that reached it.
+    ///
+    /// The empty slot (`None`) is its own state and means the mock was never hit at
+    /// all, which is what the refusal test asserts. It must stay distinguishable from
+    /// [`Verdict::WrongIdentity`], because "nobody verified anything" and "something
+    /// verified against the wrong key" are opposite findings.
+    ///
+    /// The payloads are read through `Debug` in the assertion messages and nowhere
+    /// else, which the dead-code pass does not count; they carry the detail that makes
+    /// a failure legible, so they stay.
+    #[derive(Debug)]
+    #[allow(dead_code)]
+    enum Verdict {
+        /// The chain accepted the signature AND the key it resolved is the test's DID.
+        Accepted,
+        /// The chain refused it. Carries the error so a failure reads as the actual
+        /// rejection rather than a bare hit count.
+        Rejected(String),
+        /// The chain accepted a signature made by somebody else. A key resolved from
+        /// the parsed `key_id` is read out of the artifact under verification, so an
+        /// accept on it alone proves consistency, never authenticity.
+        WrongIdentity { expected: String, got: String },
+    }
+
+    /// The node's `require_signature` verification, over a request this crate did not
+    /// necessarily build: parse the headers, recompute the content-digest from the
+    /// body, rebuild the signing string over `@method`/`@path`/`content-digest`,
+    /// Ed25519-verify. Returns the DID the signature resolved to, so a caller can pin
+    /// the identity.
+    ///
+    /// Its twin is the hand-copy in `crates/gl/src/http.rs`, which cannot import this
+    /// module (this is a binary crate's test module). Keep the two textually identical
+    /// apart from the mockito seam around them, so an edit to one is visibly an edit to
+    /// both.
+    ///
+    /// The production verifier both copies mirror is `crate::auth::require_signature` in
+    /// `crates/gitlawb-node/src/auth/mod.rs`. This is a re-implementation, not a call, so
+    /// an edit to that middleware has to land here too: otherwise the copies drift and
+    /// this test keeps passing against a rule the node has stopped applying.
+    ///
+    /// It asserts internally, which is deliberate but constrains its callers: inside
+    /// `with_body_from_request` those assertions fire on the server thread and reach
+    /// the client as a transport error, not as a recorded verdict. So the identity
+    /// check lives in the caller as a [`Verdict`] variant, never as an assert in here.
+    fn node_verifies(
+        method: &str,
+        path_and_query: &str,
+        body: &[u8],
+        sig_input: &str,
+        sig_header: &str,
+        content_digest: &str,
+    ) -> anyhow::Result<String> {
+        use gitlawb_core::http_sig::{build_signing_string, compute_content_digest, HttpSignature};
+        use gitlawb_core::identity::verify;
+        use std::collections::HashMap;
+
+        let sig = HttpSignature::parse(sig_input, sig_header)?;
+        sig.check_created()?;
+        assert!(
+            sig.missing_components().is_empty(),
+            "signature must cover all required components"
+        );
+        assert_eq!(sig.alg, "ed25519");
+        assert_eq!(
+            content_digest,
+            compute_content_digest(body),
+            "content-digest must match the body"
+        );
+        let vk = sig.key_id.to_verifying_key()?;
+        let mut values = HashMap::new();
+        values.insert("@method".to_string(), method.to_uppercase());
+        values.insert("@path".to_string(), path_and_query.to_string());
+        values.insert("content-digest".to_string(), content_digest.to_string());
+        let sig_params_value = sig_input.strip_prefix("sig1=").unwrap_or(sig_input);
+        let components: Vec<&str> = sig.components.iter().map(String::as_str).collect();
+        let signing_string = build_signing_string(&components, sig_params_value, &values)?;
+        let sig_array: [u8; 64] = sig.signature_bytes.as_slice().try_into()?;
+        verify(&vk, signing_string.as_bytes(), &sig_array)?;
+        Ok(sig.key_id.to_string())
+    }
+
+    /// Pull a header value off a received mockito request, or explain which one the
+    /// helper failed to send.
+    fn received_header(req: &mockito::Request, name: &str) -> String {
+        req.header(name)
+            .first()
+            .unwrap_or_else(|| panic!("the helper sent no {name} header"))
+            .to_str()
+            .unwrap()
+            .to_string()
+    }
+
+    /// Run [`node_verifies`] over a GET that arrived at the mock and record what the
+    /// node would have made of it, pinned to `expected_did`.
+    fn record_get_verdict(
+        req: &mockito::Request,
+        expected_did: &str,
+        slot: &std::sync::Arc<std::sync::Mutex<Option<Verdict>>>,
+    ) {
+        let verdict = match node_verifies(
+            "GET",
+            req.path_and_query(),
+            b"",
+            &received_header(req, "signature-input"),
+            &received_header(req, "signature"),
+            &received_header(req, "content-digest"),
+        ) {
+            Ok(did) if did == expected_did => Verdict::Accepted,
+            Ok(did) => Verdict::WrongIdentity {
+                expected: expected_did.to_string(),
+                got: did,
+            },
+            Err(e) => Verdict::Rejected(e.to_string()),
+        };
+        *slot.lock().unwrap() = Some(verdict);
+    }
+
+    /// The finding's repro, now a guard: a rewritten same-origin target must never
+    /// receive the advertisement's signature, and the proof is the node's own
+    /// verification, not a hit count.
+    ///
+    /// Post-fix the hop is refused, so the slot stays empty. Pre-fix the hop is
+    /// followed and the slot records the Ed25519 rejection of a signature made over
+    /// `.../info/refs?service=git-upload-pack` and presented at `.../info/refs/...`,
+    /// which is the 401 an operator behind such a proxy actually sees. The verdict is
+    /// asserted first, so a failure speaks about verification rather than about
+    /// reachability.
+    ///
+    /// Its paired positive control is
+    /// `a_direct_signed_advertisement_verifies_under_the_node_verifier`: without it, an
+    /// empty slot would be satisfied just as well by a harness that can never record
+    /// anything.
+    #[test]
+    fn a_rewritten_target_never_receives_the_signature() {
+        let kp = Keypair::generate();
+        let expected_did = kp.did().to_string();
+        let client = build_http_client().unwrap();
+        let slot = std::sync::Arc::new(std::sync::Mutex::new(None::<Verdict>));
+
+        let mut node = mockito::Server::new();
+        let bounce = node
+            .mock("GET", "/zOwner/myrepo/info/refs?service=git-upload-pack")
+            .with_status(301)
+            .with_header(
+                "location",
+                "/zOwner/myrepo/info/refs/?service=git-upload-pack",
+            )
+            .expect(1)
+            .create();
+        let recorder = slot.clone();
+        let did_for_target = expected_did.clone();
+        let target = node
+            .mock("GET", "/zOwner/myrepo/info/refs/?service=git-upload-pack")
+            .with_status(200)
+            .with_body_from_request(move |req| {
+                record_get_verdict(req, &did_for_target, &recorder);
+                Vec::new()
+            })
+            .expect(0)
+            .create();
+
+        let refs_url = format!(
+            "{}/zOwner/myrepo/info/refs?service=git-upload-pack",
+            node.url()
+        );
+        let resp = build_advertisement_request(&client, &refs_url, Some(&kp))
+            .send()
+            .unwrap();
+
+        let verdict = slot.lock().unwrap().take();
+        assert!(
+            verdict.is_none(),
+            "the node's own verifier must never see this request: the signature covers \
+             /zOwner/myrepo/info/refs?service=git-upload-pack and the rewritten target \
+             adds a trailing slash, so what arrives there is a stale request-target; \
+             recorded verdict: {verdict:?}"
+        );
+        assert_eq!(
+            resp.status(),
+            301,
+            "the caller sees the 3xx, not the rewritten target's answer"
+        );
+        bounce.assert();
+        target.assert();
+    }
+
+    /// The positive control for the test above, and the proof that the helper signs the
+    /// query it sends.
+    ///
+    /// A direct signed advertisement GET, no redirect anywhere, through the same
+    /// verifying mock. The verdict must be `Accepted`, which is what makes the refusal
+    /// test's empty slot attributable to the refusal rather than to a harness that
+    /// cannot record. The advertisement URL carries `?service=`, so a helper that
+    /// signed the bare path would land here as `Rejected`.
+    #[test]
+    fn a_direct_signed_advertisement_verifies_under_the_node_verifier() {
+        let kp = Keypair::generate();
+        let expected_did = kp.did().to_string();
+        let client = build_http_client().unwrap();
+        let slot = std::sync::Arc::new(std::sync::Mutex::new(None::<Verdict>));
+
+        let mut node = mockito::Server::new();
+        let recorder = slot.clone();
+        let did_for_target = expected_did.clone();
+        let route = node
+            .mock("GET", "/zOwner/myrepo/info/refs?service=git-upload-pack")
+            .with_status(200)
+            .with_body_from_request(move |req| {
+                record_get_verdict(req, &did_for_target, &recorder);
+                Vec::new()
+            })
+            .expect(1)
+            .create();
+
+        let refs_url = format!(
+            "{}/zOwner/myrepo/info/refs?service=git-upload-pack",
+            node.url()
+        );
+        let resp = build_advertisement_request(&client, &refs_url, Some(&kp))
+            .send()
+            .unwrap();
+        assert_eq!(resp.status(), 200);
+
+        let verdict = slot.lock().unwrap().take();
+        assert!(
+            matches!(verdict, Some(Verdict::Accepted)),
+            "a direct signed advertisement must verify under the node's own chain and \
+             resolve to {expected_did}, or the refusal test's empty slot proves \
+             nothing; recorded verdict: {verdict:?}"
+        );
+        route.assert();
+    }
+
    /// The regression that round-1 missed: the Phase-2 `git-upload-pack` POST was
    /// left unsigned, so an owner's fetch of a private repo cleared the (now signed)
    /// advertisement and then 404'd on the pack POST. Drive BOTH request builders
@@ -1137,49 +1720,13 @@ mod tests {
    /// to end (sign here, verify with the node's verifier), not reasoned.
    #[test]
    fn client_signature_verifies_under_node_verification_for_both_services() {
-        use gitlawb_core::http_sig::{build_signing_string, compute_content_digest, HttpSignature};
-        use gitlawb_core::identity::verify;
-        use std::collections::HashMap;
-
        let kp = Keypair::generate();
        let client = reqwest::blocking::Client::new();
        let body = b"0009done\n".to_vec();

-        // Re-implements the node's require_signature verification (auth/mod.rs):
-        // parse headers, recompute content-digest from the body, rebuild the signing
-        // string over @method/@path/content-digest, Ed25519-verify. Ok iff the node
-        // would accept it.
-        let node_verifies = |method: &str,
-                             path_and_query: &str,
-                             body: &[u8],
-                             sig_input: &str,
-                             sig_header: &str,
-                             content_digest: &str|
-         -> anyhow::Result<()> {
-            let sig = HttpSignature::parse(sig_input, sig_header)?;
-            sig.check_created()?;
-            assert!(
-                sig.missing_components().is_empty(),
-                "signature must cover all required components"
-            );
-            assert_eq!(sig.alg, "ed25519");
-            assert_eq!(
-                content_digest,
-                compute_content_digest(body),
-                "content-digest must match the body"
-            );
-            let vk = sig.key_id.to_verifying_key()?;
-            let mut values = HashMap::new();
-            values.insert("@method".to_string(), method.to_uppercase());
-            values.insert("@path".to_string(), path_and_query.to_string());
-            values.insert("content-digest".to_string(), content_digest.to_string());
-            let sig_params_value = sig_input.strip_prefix("sig1=").unwrap_or(sig_input);
-            let components: Vec<&str> = sig.components.iter().map(String::as_str).collect();
-            let signing_string = build_signing_string(&components, sig_params_value, &values)?;
-            let sig_array: [u8; 64] = sig.signature_bytes.as_slice().try_into()?;
-            verify(&vk, signing_string.as_bytes(), &sig_array)?;
-            Ok(())
-        };
+        // The verification chain is [`node_verifies`], the test-module helper the
+        // redirect verdict tests share. Same primitives the node's require_signature
+        // runs, over the request-target this crate transmits.
        // @path exactly as the node reconstructs it from the request it receives.
        let path_and_query = |req: &reqwest::blocking::Request| match req.url().query() {
            Some(q) => format!("{}?{}", req.url().path(), q),
@@ -1571,6 +2118,184 @@ mod tests {
        assert!(help.ends_with('\n'));
    }

+    // ── #369 delete-only push deadlock ───────────────────────────────────────
+
+    /// A delete-only receive-pack request ends at the command flush; git then
+    /// holds the write pipe open for the report and sends no pack. A reader that
+    /// blocks after the command list reproduces the hang that a Cursor (EOF)
+    /// hides. The read must return at the flush, inside the timeout.
+    #[test]
+    fn delete_only_receive_pack_request_returns_at_command_flush() {
+        let old = "a".repeat(40);
+        let zero = "0".repeat(40);
+        let mut seed = Vec::new();
+        seed.extend_from_slice(&pkt(&format!("{old} {zero} refs/heads/gone\n")));
+        seed.extend_from_slice(b"0000"); // command-list flush; git then blocks
+
+        let (tx, rx) = std::sync::mpsc::channel::<()>();
+        let mut reader = BlockAfterSeed {
+            seed: io::Cursor::new(seed),
+            gate: rx,
+        };
+
+        let (done_tx, done_rx) = std::sync::mpsc::channel::<Vec<u8>>();
+        let handle = std::thread::spawn(move || {
+            let _ = done_tx.send(read_receive_pack_request(&mut reader).unwrap());
+        });
+
+        let got = done_rx.recv_timeout(std::time::Duration::from_secs(2));
+        let _ = tx.send(()); // unblock the reader so the thread can exit
+        let _ = handle.join();
+
+        let body = got.expect(
+            "read_receive_pack_request blocked past the command flush waiting for \
+             a pack a delete-only push never sends (#369)",
+        );
+        assert!(
+            body.ends_with(b"0000"),
+            "the request body keeps the command list and its flush terminator"
+        );
+    }
+
+    /// A push carrying objects still reads the pack through to EOF, and the
+    /// command flush sits between the commands and the pack in the body.
+    #[test]
+    fn receive_pack_request_with_pack_reads_through_eof() {
+        let old = "a".repeat(40);
+        let new = "b".repeat(40);
+        let mut stream = Vec::new();
+        stream.extend_from_slice(&pkt(&format!(
+            "{old} {new} refs/heads/main\0 report-status-v2\n"
+        )));
+        stream.extend_from_slice(b"0000");
+        stream.extend_from_slice(b"PACK\x00\x00\x00\x02fakepack");
+
+        let mut stdin = io::Cursor::new(stream);
+        let body = read_receive_pack_request(&mut stdin).unwrap();
+        let pos = body.windows(4).position(|w| w == b"0000").unwrap();
+        assert_eq!(
+            &body[pos + 4..],
+            b"PACK\x00\x00\x00\x02fakepack",
+            "a non-zero new-oid means the pack follows the flush"
+        );
+    }
+
+    /// A signed delete-only push sends NO command list: the request starts
+    /// with the "push-cert" line, the commands live inside the cert, and the
+    /// cert ends with a "push-cert-end" pkt-line (not a flush). The reader
+    /// must return after push-cert-end without waiting on the pipe.
+    #[test]
+    fn signed_delete_only_push_returns_after_push_cert_end() {
+        let old = "a".repeat(40);
+        let zero = "0".repeat(40);
+        let mut seed = Vec::new();
+        seed.extend_from_slice(&pkt("push-cert\0 report-status-v2 push-cert\n"));
+        seed.extend_from_slice(&pkt("certificate version 0.1\n"));
+        seed.extend_from_slice(&pkt("pusher did:key:z6Mk\n"));
+        seed.extend_from_slice(&pkt("\n")); // empty line ends the headers
+        seed.extend_from_slice(&pkt(&format!("{old} {zero} refs/heads/gone\n")));
+        seed.extend_from_slice(&pkt("sig-line\n"));
+        seed.extend_from_slice(&pkt("push-cert-end\n"));
+        // git now blocks for the report; the pipe stays open.
+
+        let (tx, rx) = std::sync::mpsc::channel::<()>();
+        let mut reader = BlockAfterSeed {
+            seed: io::Cursor::new(seed),
+            gate: rx,
+        };
+
+        let (done_tx, done_rx) = std::sync::mpsc::channel::<Vec<u8>>();
+        let handle = std::thread::spawn(move || {
+            let _ = done_tx.send(read_receive_pack_request(&mut reader).unwrap());
+        });
+        let got = done_rx.recv_timeout(std::time::Duration::from_secs(2));
+        let _ = tx.send(());
+        let _ = handle.join();
+
+        let body = got.expect(
+            "read_receive_pack_request blocked past push-cert-end on a signed \
+             delete-only push (#369)",
+        );
+        assert!(
+            body.windows(6).any(|w| w == b"pusher"),
+            "the cert block is part of the request body"
+        );
+    }
+
+    /// A signed pack-bearing push: cert with the commands inside, then the
+    /// pack. Everything must land in the body.
+    #[test]
+    fn signed_pack_push_keeps_cert_and_pack() {
+        let old = "a".repeat(40);
+        let new = "b".repeat(40);
+        let mut stream = Vec::new();
+        stream.extend_from_slice(&pkt("push-cert\0 report-status-v2 push-cert\n"));
+        stream.extend_from_slice(&pkt("certificate version 0.1\n"));
+        stream.extend_from_slice(&pkt("\n"));
+        stream.extend_from_slice(&pkt(&format!("{old} {new} refs/heads/main\n")));
+        stream.extend_from_slice(&pkt("push-cert-end\n"));
+        stream.extend_from_slice(b"PACK\x00\x00\x00\x02fakepack");
+
+        let mut stdin = io::Cursor::new(stream);
+        let body = read_receive_pack_request(&mut stdin).unwrap();
+        assert!(
+            body.windows(5).any(|w| w == b"PACK\x00"),
+            "the pack follows the cert into the body"
+        );
+    }
+
+    /// A negotiated push-options block is pkt-lines ending in its own flush.
+    /// With no cert it follows the command flush; with a cert it follows
+    /// push-cert-end.
+    #[test]
+    fn receive_pack_request_consumes_a_push_options_block() {
+        let old = "a".repeat(40);
+        let new = "b".repeat(40);
+        let mut stream = Vec::new();
+        stream.extend_from_slice(&pkt(&format!(
+            "{old} {new} refs/heads/main\0 report-status-v2 push-options\n"
+        )));
+        stream.extend_from_slice(b"0000"); // command flush
+        stream.extend_from_slice(&pkt("ci.skip\n"));
+        stream.extend_from_slice(b"0000"); // options flush
+        stream.extend_from_slice(b"PACK\x00\x00\x00\x02fakepack");
+
+        let mut stdin = io::Cursor::new(stream);
+        let body = read_receive_pack_request(&mut stdin).unwrap();
+        assert!(body.windows(7).any(|w| w == b"ci.skip"), "options survived");
+        assert!(
+            body.ends_with(b"PACK\x00\x00\x00\x02fakepack"),
+            "the pack lands after the options block"
+        );
+    }
+
+    /// Cert and options together: the cert comes first and the options block
+    /// sits between push-cert-end and the pack.
+    #[test]
+    fn signed_push_with_options_reads_cert_then_options_then_pack() {
+        let old = "a".repeat(40);
+        let new = "b".repeat(40);
+        let mut stream = Vec::new();
+        stream.extend_from_slice(&pkt(
+            "push-cert\0 report-status-v2 push-cert push-options\n",
+        ));
+        stream.extend_from_slice(&pkt("certificate version 0.1\n"));
+        stream.extend_from_slice(&pkt("\n"));
+        stream.extend_from_slice(&pkt(&format!("{old} {new} refs/heads/main\n")));
+        stream.extend_from_slice(&pkt("push-cert-end\n"));
+        stream.extend_from_slice(&pkt("ci.skip\n"));
+        stream.extend_from_slice(b"0000"); // options flush
+        stream.extend_from_slice(b"PACK\x00\x00\x00\x02fakepack");
+
+        let mut stdin = io::Cursor::new(stream);
+        let body = read_receive_pack_request(&mut stdin).unwrap();
+        assert!(body.windows(7).any(|w| w == b"ci.skip"), "options survived");
+        assert!(
+            body.ends_with(b"PACK\x00\x00\x00\x02fakepack"),
+            "the pack lands after the options block"
+        );
+    }
+
    // ── #117 multi-round fetch negotiation ───────────────────────────────────

    /// Encode a git pkt-line: 4-byte hex length (incl. the 4 bytes) + data.
@@ -95,6 +95,36 @@ fn upload_pack(repo: &Path, advertise: bool, input: &[u8]) -> Vec<u8> {
    out.stdout
 }

+/// Run `git receive-pack --stateless-rpc [--advertise-refs] <repo>`, optionally
+/// feeding `input` on stdin. Returns raw stdout (the wire bytes). The receive
+/// half of `upload_pack`, used so the shim can serve a real push.
+fn receive_pack(repo: &Path, advertise: bool, input: &[u8]) -> Vec<u8> {
+    let mut cmd = Command::new("git");
+    cmd.arg("receive-pack").arg("--stateless-rpc");
+    if advertise {
+        cmd.arg("--advertise-refs");
+    }
+    cmd.arg(repo)
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped());
+    let mut child = cmd.spawn().expect("spawn git receive-pack");
+    let mut stdin = child.stdin.take().unwrap();
+    let input = input.to_vec();
+    let writer = std::thread::spawn(move || {
+        let _ = stdin.write_all(&input);
+        // drop closes stdin
+    });
+    let out = child.wait_with_output().expect("wait receive-pack");
+    writer.join().ok();
+    assert!(
+        out.status.success(),
+        "git receive-pack failed: {}",
+        String::from_utf8_lossy(&out.stderr)
+    );
+    out.stdout
+}
+
 #[derive(Clone, Copy)]
 enum ShimMode {
    /// Faithful v0 serving: pipe each POST to `git upload-pack --stateless-rpc`.
@@ -107,6 +137,11 @@ enum ShimMode {
 struct Shim {
    base_url: String,
    posts: Arc<AtomicUsize>,
+    /// Advertisement (GET /info/refs) count. One `git fetch` performs exactly
+    /// one advertisement, so tests assert this is EXACTLY one: any fetch-level
+    /// retry that comes back re-advertises and goes red on the spot, which the
+    /// lower-bound POST assertions alone cannot see (#275 round 6).
+    gets: Arc<AtomicUsize>,
    stop: Arc<AtomicBool>,
    handle: Option<std::thread::JoinHandle<()>>,
 }
@@ -136,15 +171,17 @@ fn start_shim(repo: PathBuf, mode: ShimMode) -> Shim {
    let addr = listener.local_addr().unwrap();
    let base_url = format!("http://{addr}");
    let posts = Arc::new(AtomicUsize::new(0));
+    let gets = Arc::new(AtomicUsize::new(0));
    let stop = Arc::new(AtomicBool::new(false));

    let posts_t = posts.clone();
+    let gets_t = gets.clone();
    let stop_t = stop.clone();
    let handle = std::thread::spawn(move || {
        while !stop_t.load(Ordering::SeqCst) {
            match listener.accept() {
                Ok((stream, _)) => {
-                    handle_conn(stream, &repo, mode, &posts_t);
+                    handle_conn(stream, &repo, mode, &posts_t, &gets_t);
                }
                Err(ref e) if e.kind() == std::io::ErrorKind::WouldBlock => {
                    std::thread::sleep(Duration::from_millis(5));
@@ -157,6 +194,7 @@ fn start_shim(repo: PathBuf, mode: ShimMode) -> Shim {
    Shim {
        base_url,
        posts,
+        gets,
        stop,
        handle: Some(handle),
    }
@@ -182,7 +220,13 @@ fn normalize_accepted_stream(stream: &TcpStream) {
    stream.set_read_timeout(Some(Duration::from_secs(30))).ok();
 }

-fn handle_conn(stream: TcpStream, repo: &Path, mode: ShimMode, posts: &AtomicUsize) {
+fn handle_conn(
+    stream: TcpStream,
+    repo: &Path,
+    mode: ShimMode,
+    posts: &AtomicUsize,
+    gets: &AtomicUsize,
+) {
    normalize_accepted_stream(&stream);
    let mut reader = BufReader::new(stream);

@@ -218,7 +262,22 @@ fn handle_conn(stream: TcpStream, repo: &Path, mode: ShimMode, posts: &AtomicUsi
        reader.read_exact(&mut body).ok();
    }

-    let (content_type, payload) = if method == "GET" && target.contains("/info/refs") {
+    let (content_type, payload) = if method == "GET" && target.contains("service=git-receive-pack")
+    {
+        // Receive-pack advertisement, same wrapper shape as upload-pack.
+        let adv = receive_pack(repo, true, b"");
+        let mut wrapped = pkt(b"# service=git-receive-pack\n");
+        wrapped.extend_from_slice(b"0000");
+        wrapped.extend_from_slice(&adv);
+        ("application/x-git-receive-pack-advertisement", wrapped)
+    } else if method == "POST" && target.ends_with("/git-receive-pack") {
+        posts.fetch_add(1, Ordering::SeqCst);
+        (
+            "application/x-git-receive-pack-result",
+            receive_pack(repo, false, &body),
+        )
+    } else if method == "GET" && target.contains("/info/refs") {
+        gets.fetch_add(1, Ordering::SeqCst);
        // v0 advertisement, wrapped exactly as the node's info_refs does.
        let adv = upload_pack(repo, true, b"");
        let mut wrapped = pkt(b"# service=git-upload-pack\n");
@@ -293,6 +352,34 @@ fn fetch_with_helper(clone: &Path, node_url: &str) -> (bool, std::process::Outpu
    run_bounded(cmd, Duration::from_secs(30))
 }

+/// Run `git push` in `local` through the helper, same wiring and hard timeout as
+/// [`fetch_with_helper`]. `args` carries the push spec, e.g.
+/// `["--delete", "doomed"]`.
+fn push_with_helper(local: &Path, node_url: &str, args: &[&str]) -> (bool, std::process::Output) {
+    let helper_bin = PathBuf::from(env!("CARGO_BIN_EXE_git-remote-gitlawb"));
+    let helper_dir = helper_bin.parent().unwrap().to_path_buf();
+    let path_env = match std::env::var_os("PATH") {
+        Some(p) => {
+            let mut dirs = vec![helper_dir.clone()];
+            dirs.extend(std::env::split_paths(&p));
+            std::env::join_paths(dirs).unwrap()
+        }
+        None => helper_dir.clone().into_os_string(),
+    };
+
+    let mut cmd = Command::new("git");
+    cmd.arg("-C")
+        .arg(local)
+        .arg("push")
+        .arg("origin")
+        .args(args)
+        .env("PATH", path_env)
+        .env("LC_ALL", "C")
+        .env("GITLAWB_NODE", node_url)
+        .env("GITLAWB_KEY", "/nonexistent-key-for-anon-push");
+    run_bounded(cmd, Duration::from_secs(30))
+}
+
 /// A synthetic non-success `ExitStatus` used only on the unix timeout path, where
 /// `reap_group` has already consumed the child and the later `wait` returns ECHILD.
 /// The timeout branch already reports `completed == false`, so the exact status is
@@ -935,6 +1022,7 @@ fn shim_answers_a_connection_that_arrives_non_blocking() {

    let server = std::thread::spawn(move || {
        let posts = AtomicUsize::new(0);
+        let gets = AtomicUsize::new(0);
        let stream = loop {
            match listener.accept() {
                Ok((s, _)) => break s,
@@ -949,7 +1037,7 @@ fn shim_answers_a_connection_that_arrives_non_blocking() {
        // the assertions below run on every platform.
        stream.set_nonblocking(true).unwrap();
        accepted_tx.send(()).unwrap();
-        handle_conn(stream, &repo, ShimMode::Normal, &posts);
+        handle_conn(stream, &repo, ShimMode::Normal, &posts, &gets);
    });

    let mut client = TcpStream::connect(addr).unwrap();
@@ -1013,6 +1101,17 @@ fn real_git_multi_round_fetch_completes() {
        posts >= 2,
        "fixture did not force multi-round negotiation (observed {posts} POST(s)); the bridging path was not exercised"
    );
+    // Exactly one advertisement: the POST bound above is a floor, so a
+    // reintroduced fetch-level retry could satisfy it by running a second full
+    // fetch instead of multi-round negotiation. Re-advertising is what a
+    // second fetch cannot avoid, so this is the committed guard on the
+    // retry's absence.
+    let gets = shim.gets.load(Ordering::SeqCst);
+    assert_eq!(
+        gets, 1,
+        "expected exactly one advertisement GET per fetch; {gets} means a \
+         fetch-level retry is back"
+    );

    // The fetched tip is present and the clone's object graph is intact.
    let server_head = String::from_utf8(git(&server, &["rev-parse", "HEAD"])).unwrap();
@@ -1105,6 +1204,17 @@ fn real_git_withheld_shaped_first_post() {
        "helper hung on a withheld-shaped response (should forward-and-terminate, not deadlock). stderr:\n{stderr}"
    );

+    // Exactly one advertisement, pass or fail: the POST assertions below are
+    // floors, so a reintroduced fetch-level retry could mask the withheld
+    // shape behind a second full fetch. A retry cannot avoid re-advertising,
+    // so this is the committed guard on its absence.
+    let gets = shim.gets.load(Ordering::SeqCst);
+    assert_eq!(
+        gets, 1,
+        "expected exactly one advertisement GET per fetch; {gets} means a \
+         fetch-level retry is back"
+    );
+
    if out.status.success() {
        // Real git accepted the mid-negotiation pack: the withheld multi-round
        // path works end to end with no extra handling.
@@ -1242,3 +1352,77 @@ fn run_bounded_bounds_join_when_leader_exits_leaving_a_pipe_holder() {
         held pipes do not stall the joins)"
    );
 }
+
+/// #369 end-to-end: a real `git push --delete` through the helper completes.
+/// The receive-pack request for a delete-only push is commands + flush with NO
+/// pack, and git holds its write pipe open for the report — so a helper that
+/// reads stdin to EOF deadlocks against it. The shim pipes the POST to a real
+/// `git receive-pack --stateless-rpc`, so the branch actually has to be deleted
+/// server-side for the assertions to pass.
+#[test]
+fn real_git_push_delete_completes() {
+    let dir = tempfile::tempdir().unwrap();
+    let server = dir.path().join("server.git");
+    let local = dir.path().join("local");
+
+    // A bare repo holding the doomed branch, plus a live one.
+    git(
+        dir.path(),
+        &[
+            "init",
+            "--bare",
+            server.to_str().unwrap(),
+            "--initial-branch=main",
+        ],
+    );
+    let work = dir.path().join("work");
+    git(
+        dir.path(),
+        &["init", work.to_str().unwrap(), "--initial-branch=main"],
+    );
+    git(&work, &["config", "user.email", "t@t"]);
+    git(&work, &["config", "user.name", "t"]);
+    std::fs::write(work.join("f"), "x").unwrap();
+    git(&work, &["add", "f"]);
+    git(&work, &["commit", "-m", "c"]);
+    git(&work, &["branch", "doomed"]);
+    git(&work, &["push", server.to_str().unwrap(), "main", "doomed"]);
+
+    // The pusher: a local repo whose origin is the gitlawb remote pointing at
+    // the shim. Nothing needs to be fetched first; git takes the old-oid from
+    // the receive-pack advertisement.
+    git(
+        dir.path(),
+        &["init", local.to_str().unwrap(), "--initial-branch=main"],
+    );
+    git(
+        &local,
+        &["remote", "add", "origin", "gitlawb://did:key:z6MkTest/repo"],
+    );
+
+    let shim = start_shim(server.clone(), ShimMode::Normal);
+    let (completed, out) = push_with_helper(&local, &shim.base_url, &["--delete", "doomed"]);
+    let posts = shim.posts.load(Ordering::SeqCst);
+
+    assert!(
+        completed,
+        "git push --delete did not complete within the timeout (deadlock signature). stderr:\n{}",
+        String::from_utf8_lossy(&out.stderr)
+    );
+    assert!(
+        out.status.success(),
+        "git push --delete failed. stderr:\n{}",
+        String::from_utf8_lossy(&out.stderr)
+    );
+    assert_eq!(posts, 1, "a delete-only push is exactly one POST");
+    let gone = Command::new("git")
+        .arg("-C")
+        .arg(&server)
+        .args(["rev-parse", "--verify", "--quiet", "refs/heads/doomed"])
+        .output()
+        .unwrap();
+    assert!(
+        !gone.status.success(),
+        "refs/heads/doomed must be deleted in the server repo"
+    );
+}
@@ -1,6 +1,6 @@
 [package]
 name = "gitlawb-attest"
-description = "External Attestation v1: pluggable provenance attachments for gitlawb ref-update certs"
+description = "External Attestation v1: pluggable provenance attachments for Twigpine ref-update certs"
 version = "0.7.1" # x-release-please-version
 edition.workspace = true
 rust-version.workspace = true
@@ -105,16 +105,30 @@ impl Attestation {
        let bytes = canonical_signing_bytes(&self.type_, &self.payload, &self.cert_hash)?;

        let vk = verifying_key_from_did_key(&self.signer)?;
+        self.verify_sig_with_key(&vk, &bytes)?;
+
+        Ok(vk)
+    }
+
+    /// The signature check itself, against a key the caller supplies.
+    ///
+    /// Split out so the STRICTNESS of the check stays provable on its own.
+    /// `verifying_key_from_did_key` now refuses a small-order key, so a weak
+    /// signer can no longer reach this line through `verify_signature`. Without
+    /// this seam there would be nothing left that fails when `verify_strict` is
+    /// downgraded to the non-strict `verify`, and that regression would land
+    /// silently. Check order in `verify_signature` is unchanged.
+    fn verify_sig_with_key(&self, vk: &VerifyingKey, bytes: &[u8]) -> Result<()> {
        let sig_bytes: [u8; 64] = B64U
            .decode(&self.sig)
            .map_err(|e| Error::Signature(format!("base64url: {e}")))?
            .try_into()
            .map_err(|_| Error::Signature("signature must be 64 bytes".to_string()))?;
        let sig = Signature::from_bytes(&sig_bytes);
-        vk.verify_strict(&bytes, &sig)
+        vk.verify_strict(bytes, &sig)
            .map_err(|e| Error::Signature(format!("ed25519: {e}")))?;

-        Ok(vk)
+        Ok(())
    }

    /// Reparse `payload` as `P`. Errors if the type discriminator does not
@@ -181,7 +195,7 @@ fn canonical_signing_bytes(
    Ok(out)
 }

-fn did_key_from_verifying_key(key: &VerifyingKey) -> String {
+pub(crate) fn did_key_from_verifying_key(key: &VerifyingKey) -> String {
    let mut buf = Vec::with_capacity(ED25519_MULTICODEC.len() + 32);
    buf.extend_from_slice(&ED25519_MULTICODEC);
    buf.extend_from_slice(&key.to_bytes());
@@ -202,6 +216,19 @@ fn verifying_key_from_did_key(did: &str) -> Result<VerifyingKey> {
            "did:key must use base58btc (z-prefix): {did}"
        )));
    }
+    // Refuse an oversized id before decoding. base58 decoding is quadratic in
+    // its input and `signer` comes off an attacker-supplied attestation, so a
+    // short request could otherwise buy a large decode. An ed25519 did:key
+    // method-id is a fixed 48 characters, so this bound is slack rather than a
+    // behavior change, and it only helps ahead of the decode. Mirrors the same
+    // cap in gitlawb-core's Did::to_verifying_key, which this crate cannot call
+    // (gitlawb-core is a dev-dependency here).
+    const MAX_METHOD_ID_LEN: usize = 64;
+    if method_id.len() > MAX_METHOD_ID_LEN {
+        return Err(Error::Did(
+            "did:key method-specific id too long".to_string(),
+        ));
+    }
    let (base, bytes) =
        multibase::decode(method_id).map_err(|e| Error::Did(format!("multibase: {e}")))?;
    if base != multibase::Base::Base58Btc {
@@ -217,8 +244,20 @@ fn verifying_key_from_did_key(did: &str) -> Result<VerifyingKey> {
    let key_bytes: [u8; 32] = bytes[ED25519_MULTICODEC.len()..]
        .try_into()
        .expect("length checked above");
-    VerifyingKey::from_bytes(&key_bytes)
-        .map_err(|e| Error::Did(format!("invalid ed25519 key: {e}")))
+    let key = VerifyingKey::from_bytes(&key_bytes)
+        .map_err(|e| Error::Did(format!("invalid ed25519 key: {e}")))?;
+
+    // `from_bytes` only decompresses, so it accepts a small-order point. This
+    // crate parses did:key itself rather than going through gitlawb-core's
+    // `Did`, and it cannot go through it (gitlawb-core is a dev-dependency
+    // here), so the rejection is mirrored rather than inherited. Signature
+    // verification is already strict, so this is defense in depth against a
+    // future consumer of this parser that does not verify strictly.
+    if key.is_weak() {
+        return Err(Error::Did("small-order ed25519 key".to_string()));
+    }
+
+    Ok(key)
 }

 #[cfg(test)]
@@ -465,6 +504,23 @@ mod tests {
        assert!(matches!(err, Error::Did(_)));
    }

+    #[test]
+    fn verify_rejects_oversized_method_id_before_decoding() {
+        let sk = fresh();
+        let mut att = dummy_attestation(&sk, sample_cert_hash());
+        // A well-formed prefix on an absurdly long id must fail the length
+        // bound, not the decoder: base58 decode cost is quadratic in input
+        // length, so reaching it at all is the defect.
+        att.signer = format!("did:key:z{}", "1".repeat(256));
+        let err = att.verify_signature(sample_cert_hash()).unwrap_err();
+        // Match the bound's own message: any other Did error (a decode or
+        // length mismatch downstream) means the oversized input was decoded.
+        match err {
+            Error::Did(m) => assert!(m.contains("too long"), "unexpected: {m}"),
+            e => panic!("expected Error::Did, got {e:?}"),
+        }
+    }
+
    #[test]
    fn verify_rejects_bad_signature_base64() {
        let sk = fresh();
@@ -483,6 +539,43 @@ mod tests {
        assert!(matches!(err, Error::Signature(_)));
    }

+    /// The identity-point forgery must be rejected: the shared attestation
+    /// verifier is a cert-bound provenance gate, so accepting the weak-key
+    /// signature would let anyone mint a forged attestation that verifies.
+    /// Strict verification rejects small-order public keys and R (the identity
+    /// point here), which ordinary verification does not.
+    #[test]
+    fn verify_rejects_identity_point_forgery() {
+        let cert_hash = sample_cert_hash();
+        let mut att = dummy_attestation(&fresh(), cert_hash);
+
+        // Public key A = identity point (0,1); signature R = identity, S = 0.
+        // The equation `[S]B = R + [k]A` then holds for any k and any message.
+        let identity = [
+            1u8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+            0, 0, 0, 0,
+        ];
+        let mut buf = Vec::with_capacity(34);
+        buf.extend_from_slice(&ED25519_MULTICODEC);
+        buf.extend_from_slice(&identity);
+        att.signer = format!(
+            "did:key:{}",
+            multibase::encode(multibase::Base::Base58Btc, &buf)
+        );
+        let mut sig = [0u8; 64];
+        sig[..32].copy_from_slice(&identity);
+        att.sig = B64U.encode(sig);
+
+        let err = att.verify_signature(cert_hash).unwrap_err();
+        match err {
+            Error::Did(msg) => assert!(
+                msg.contains("small-order"),
+                "identity-point signer must be refused as a small-order key, got: {msg}"
+            ),
+            other => panic!("expected Error::Did for an identity-point signer, got {other:?}"),
+        }
+    }
+
    /// A payload that happens to contain a `cert_hash` field of its own does
    /// not interfere with the outer binding: the attestation envelope's
    /// `cert_hash` is the only field consulted by `verify_signature`, and the
@@ -559,6 +652,111 @@ mod tests {
    /// weak (small-order) public key satisfies the verification equation
    /// but must be rejected. The identity point is such a key: with R the
    /// identity and S = 0, [S]B - [k]A is the identity for any message.
+    /// gitlawb-attest parses did:key itself and never routes through
+    /// gitlawb-core's `Did`, which it cannot: gitlawb-core is a dev-dependency
+    /// here. So the choke-point rejection has to be mirrored in this parser or
+    /// this crate keeps handing out keys the rest of the system rejects.
+    #[test]
+    fn verifying_key_from_did_key_rejects_a_small_order_key() {
+        let mut weak = [0u8; 32];
+        weak[0] = 1; // compressed identity point
+        let mut buf = Vec::with_capacity(ED25519_MULTICODEC.len() + 32);
+        buf.extend_from_slice(&ED25519_MULTICODEC);
+        buf.extend_from_slice(&weak);
+        let did = format!(
+            "did:key:{}",
+            multibase::encode(multibase::Base::Base58Btc, &buf)
+        );
+
+        let err =
+            verifying_key_from_did_key(&did).expect_err("a small-order did:key must not resolve");
+        match err {
+            Error::Did(msg) => assert!(
+                msg.contains("small-order"),
+                "rejection must name the small-order key, got: {msg}"
+            ),
+            other => panic!("expected Error::Did, got {other:?}"),
+        }
+    }
+
+    /// base58 decoding is quadratic in its input and `signer` is attacker
+    /// controlled, so an oversized method-id must be refused BEFORE the decode
+    /// or a hostile attestation buys a large decode for a short request. The
+    /// assertion is on the length error specifically: a multibase error here
+    /// would mean the cap ran too late to matter.
+    #[test]
+    fn verifying_key_from_did_key_refuses_an_oversized_method_id_before_decoding() {
+        let oversized = format!("did:key:z{}", "1".repeat(65_536));
+        let err = verifying_key_from_did_key(&oversized)
+            .expect_err("an oversized method-id must be refused");
+        match err {
+            Error::Did(msg) => assert!(
+                msg.contains("too long"),
+                "must fail on the length cap, not after decoding: {msg}"
+            ),
+            other => panic!("expected Error::Did, got {other:?}"),
+        }
+    }
+
+    /// Strictness regression guard. The parser now refuses a small-order key,
+    /// so `verify_signature` can no longer carry a weak signer down to the
+    /// signature check; this drives that check directly to prove it is still
+    /// STRICT. Downgrading `verify_strict` to `verify` makes this go red,
+    /// which nothing else in the suite would catch.
+    #[test]
+    fn verify_sig_with_key_is_strict_about_a_weak_key() {
+        let mut weak_key_bytes = [0u8; 32];
+        weak_key_bytes[0] = 1; // compressed identity point
+        let weak_vk = VerifyingKey::from_bytes(&weak_key_bytes).unwrap();
+        assert!(weak_vk.is_weak());
+
+        let cert_hash = sample_cert_hash();
+        let mut att = dummy_attestation(&SigningKey::generate(&mut OsRng), cert_hash);
+        // R = identity, S = 0: satisfies the non-strict verification equation
+        // for any message under a small-order key.
+        let mut forged = [0u8; 64];
+        forged[0] = 1;
+        att.sig = B64U.encode(forged);
+
+        let bytes = canonical_signing_bytes(&att.type_, &att.payload, &att.cert_hash).expect("jcs");
+        let err = att
+            .verify_sig_with_key(&weak_vk, &bytes)
+            .expect_err("strict verification must reject a small-order key");
+        assert!(
+            matches!(err, Error::Signature(_)),
+            "expected a signature error, got {err:?}"
+        );
+    }
+
+    /// Accept control for the seam: a genuine signature still verifies through
+    /// the same helper, so the guard above is not simply rejecting everything.
+    #[test]
+    fn verify_sig_with_key_accepts_a_genuine_signature() {
+        let cert_hash = sample_cert_hash();
+        let sk = SigningKey::generate(&mut OsRng);
+        let att = dummy_attestation(&sk, cert_hash);
+        let bytes = canonical_signing_bytes(&att.type_, &att.payload, &att.cert_hash).expect("jcs");
+        att.verify_sig_with_key(&sk.verifying_key(), &bytes)
+            .expect("a real signature must verify");
+    }
+
+    /// Control for the guard above: a real did:key must still resolve.
+    #[test]
+    fn verifying_key_from_did_key_still_accepts_a_real_key() {
+        let sk = SigningKey::from_bytes(&[3u8; 32]);
+        let vk = sk.verifying_key();
+        let mut buf = Vec::with_capacity(ED25519_MULTICODEC.len() + 32);
+        buf.extend_from_slice(&ED25519_MULTICODEC);
+        buf.extend_from_slice(&vk.to_bytes());
+        let did = format!(
+            "did:key:{}",
+            multibase::encode(multibase::Base::Base58Btc, &buf)
+        );
+
+        let got = verifying_key_from_did_key(&did).expect("a real did:key must resolve");
+        assert_eq!(got.to_bytes(), vk.to_bytes());
+    }
+
    #[test]
    fn verify_rejects_weak_key_signature() {
        let mut weak_key_bytes = [0u8; 32];
@@ -572,7 +770,13 @@ mod tests {
        let forged_sig = B64U.encode(forged);

        // Build an attestation with the weak key as signer and the forged
-        // signature. The weak-key check happens at verify time.
+        // signature. Rejection now happens at DID RESOLUTION rather than at
+        // verify time: `verifying_key_from_did_key` refuses a small-order key
+        // before `verify_strict` is reached, so the observed error is
+        // `Error::Did`, not `Error::Signature`. `verify_strict` remains the
+        // second layer for malleability cases that do not involve a weak
+        // public key (a small-order R under an honest key), which this fixture
+        // cannot construct and therefore no longer covers.
        let mut att = dummy_attestation(&SigningKey::generate(&mut OsRng), cert_hash);
        let mut buf = Vec::with_capacity(ED25519_MULTICODEC.len() + 32);
        buf.extend_from_slice(&ED25519_MULTICODEC);
@@ -584,9 +788,12 @@ mod tests {
        att.sig = forged_sig;

        let err = att.verify_signature(cert_hash).unwrap_err();
-        assert!(
-            matches!(err, Error::Signature(_)),
-            "signature under a weak (small-order) public key must be rejected"
-        );
+        match err {
+            Error::Did(msg) => assert!(
+                msg.contains("small-order"),
+                "weak signer must be refused as a small-order key, got: {msg}"
+            ),
+            other => panic!("expected Error::Did for a small-order signer, got {other:?}"),
+        }
    }
 }
@@ -1,4 +1,4 @@
-//! Pluggable provenance attachments for gitlawb ref-update certs.
+//! Pluggable provenance attachments for Twigpine ref-update certs.
 //!
 //! A `gitlawb/ref-update/v1` cert proves who pushed and where a ref moved to.
 //! It does not say how the commit was produced. This crate adds an optional
@@ -121,7 +121,12 @@ impl Registry {
        attestation: &Attestation,
        expected_cert_hash: [u8; 32],
    ) -> Result<VerifiedAttestation> {
-        attestation.verify_signature(expected_cert_hash)?;
+        // The key that actually verified is the trust anchor for the signer a
+        // consumer sees. `verify_signature` recovers it from the artifact's
+        // `signer` field and rejects a field that does not match the signing
+        // key, so this is the canonical DID of the key that signed — never a
+        // raw artifact string that could disagree with it.
+        let verified_key = attestation.verify_signature(expected_cert_hash)?;

        let fully = match self.by_type.get(attestation.type_.as_str()) {
            Some(v) => {
@@ -138,30 +143,55 @@ impl Registry {

        Ok(VerifiedAttestation {
            type_: attestation.type_.clone(),
-            signer: attestation.signer.clone(),
+            signer: crate::attestation::did_key_from_verifying_key(&verified_key),
            cert_hash: attestation.cert_hash.clone(),
            fully_verified: fully,
        })
    }

    /// Verify a batch, then enforce `RequireAll`.
+    ///
+    /// Under [`Policy::RequireAll`] this never short-circuits: an attestation
+    /// that fails to verify — bad signature, cert-hash mismatch, or malformed
+    /// payload, the actual `Err` paths of [`Registry::verify`] — is dropped
+    /// from the result rather than aborting the batch. A valid attestation of
+    /// a type with NO registered verifier is not one of those: under
+    /// `RequireAll` it verifies to `Ok` with `fully_verified: false` and stays
+    /// in the result, preserving the distinction between *untrusted* and
+    /// *invalid* (see `require_all_is_lenient_on_unknown_types_in_the_batch`).
+    /// Any third party can attach an attestation, so a hard error here would
+    /// let an unrelated `attacker/spam/v1` deny-of-service the whole cert (the
+    /// DoS vector the module docstring warns about). The `required_types`
+    /// check below is the real gate, and it only counts `fully_verified`
+    /// entries — so a required type represented solely by a malformed
+    /// attestation still fails with [`Error::RequiredMissing`].
+    ///
+    /// Under `AcceptKnown` and `RejectUnknown` the batch stays strict: the
+    /// first verification error is surfaced to the caller.
    pub fn verify_all(
        &self,
        attestations: &[Attestation],
        expected_cert_hash: [u8; 32],
    ) -> Result<Vec<VerifiedAttestation>> {
-        let mut verified = Vec::with_capacity(attestations.len());
-        for a in attestations {
-            verified.push(self.verify(a, expected_cert_hash)?);
+        if self.policy != Policy::RequireAll {
+            let mut verified = Vec::with_capacity(attestations.len());
+            for a in attestations {
+                verified.push(self.verify(a, expected_cert_hash)?);
+            }
+            return Ok(verified);
        }
-        if self.policy == Policy::RequireAll {
-            for required in &self.required_types {
-                let present = verified
-                    .iter()
-                    .any(|v| &v.type_ == required && v.fully_verified);
-                if !present {
-                    return Err(Error::RequiredMissing(required.clone()));
-                }
+
+        // RequireAll: drop entries that fail to verify instead of aborting.
+        let verified: Vec<VerifiedAttestation> = attestations
+            .iter()
+            .filter_map(|a| self.verify(a, expected_cert_hash).ok())
+            .collect();
+        for required in &self.required_types {
+            let present = verified
+                .iter()
+                .any(|v| &v.type_ == required && v.fully_verified);
+            if !present {
+                return Err(Error::RequiredMissing(required.clone()));
            }
        }
        Ok(verified)
@@ -171,7 +201,7 @@ impl Registry {
 #[cfg(test)]
 mod tests {
    use super::*;
-    use crate::attestation::{Attestation, AttestationPayload};
+    use crate::attestation::{did_key_from_verifying_key, Attestation, AttestationPayload};
    use ed25519_dalek::SigningKey;
    use rand::rngs::OsRng;
    use serde::{Deserialize, Serialize};
@@ -319,6 +349,38 @@ mod tests {
            .any(|v| v.type_ == "other/v1" && !v.fully_verified));
    }

+    /// The strict policies must surface a batch member's verification error
+    /// through `verify_all`, not launder it into a partial result. This binds
+    /// the POLICY BOUNDARY itself: only the `RequireAll` arm may drop failing
+    /// entries, and an apparently harmless cleanup that applies its
+    /// `filter_map(Result::ok)` shape to the strict arms would silently turn
+    /// the fail-closed contracts into fail-open — with every per-attestation
+    /// `verify` test still green, because none of them go through the batch
+    /// path. The failing entry is a wrong-cert-hash attestation (signed over a
+    /// different hash than the batch expects), a genuine `Err` under every
+    /// policy.
+    #[test]
+    fn strict_policies_propagate_a_batch_members_error() {
+        let sk = fresh();
+        let cert_hash = sample_hash();
+        let mut wrong_hash = sample_hash();
+        wrong_hash[0] ^= 0xff;
+
+        for policy in [Policy::AcceptKnown, Policy::RejectUnknown] {
+            let mut reg = Registry::new().with_policy(policy);
+            reg.register(DemoVerifier);
+            let good = signed_demo(&sk, cert_hash, "ok");
+            let bad = signed_demo(&sk, wrong_hash, "ok");
+            let err = reg
+                .verify_all(&[good, bad], cert_hash)
+                .expect_err("a strict policy must fail the whole batch");
+            assert!(
+                matches!(err, Error::Signature(_) | Error::CertHashMismatch),
+                "expected the underlying verification error, got {err:?}"
+            );
+        }
+    }
+
    /// A required type that is present but unverified (no verifier registered
    /// even though the type appears in `required_types`) must fail
    /// `verify_all` with `RequiredMissing`. Catches a misconfiguration where
@@ -336,6 +398,65 @@ mod tests {
        assert!(matches!(err, Error::RequiredMissing(t) if t == "demo/v1"));
    }

+    /// A cert-hash that does not match `sample_hash()`, so an attestation
+    /// signed against it fails `verify_signature` when the batch is checked
+    /// against `sample_hash()`.
+    fn other_hash() -> [u8; 32] {
+        let mut h = sample_hash();
+        h[0] ^= 0xff;
+        h
+    }
+
+    /// Under `RequireAll`, a malformed attestation (here: signed against the
+    /// wrong cert hash, so it fails signature verification) attached alongside a
+    /// valid required one must not abort the batch. The junk is dropped and the
+    /// cert still verifies. Before the fix, `verify_all` propagated the error
+    /// via `?` and the whole batch failed — the DoS vector the docstring warns
+    /// about.
+    #[test]
+    fn require_all_drops_a_malformed_extra_and_still_accepts() {
+        let sk = fresh();
+        let cert_hash = sample_hash();
+        let mut reg = Registry::new()
+            .with_policy(Policy::RequireAll)
+            .require_types(["demo/v1"]);
+        reg.register(DemoVerifier);
+
+        // Signed against `other_hash()`, so it will not verify under `cert_hash`.
+        let malformed = signed_demo(&sk, other_hash(), "boom");
+        let real = signed_demo(&sk, cert_hash, "ok");
+
+        let verified = reg
+            .verify_all(&[malformed, real], cert_hash)
+            .expect("a malformed extra must not abort the batch under RequireAll");
+        assert_eq!(verified.len(), 1, "the malformed entry should be dropped");
+        assert!(verified
+            .iter()
+            .any(|v| v.type_ == "demo/v1" && v.fully_verified));
+    }
+
+    /// Leniency must not open a hole: if the *only* attestation of a required
+    /// type is malformed, dropping it leaves the required type absent and
+    /// `verify_all` must still fail with `RequiredMissing`. Guards against a fix
+    /// that skips errors so eagerly it lets an unverified required type pass.
+    #[test]
+    fn require_all_rejects_when_only_a_malformed_required_type_is_present() {
+        let sk = fresh();
+        let cert_hash = sample_hash();
+        let mut reg = Registry::new()
+            .with_policy(Policy::RequireAll)
+            .require_types(["demo/v1"]);
+        reg.register(DemoVerifier);
+
+        // The sole demo/v1 is bound to the wrong cert hash → fails verification.
+        let malformed_required = signed_demo(&sk, other_hash(), "ok");
+
+        let err = reg
+            .verify_all(&[malformed_required], cert_hash)
+            .unwrap_err();
+        assert!(matches!(err, Error::RequiredMissing(t) if t == "demo/v1"));
+    }
+
    #[test]
    fn payload_check_failure_rejects() {
        let sk = fresh();
@@ -374,4 +495,52 @@ mod tests {
        let err = reg.verify(&att, cert_hash).unwrap_err();
        assert!(matches!(err, Error::Payload(_)));
    }
+
+    /// The signer a consumer sees must be the canonical DID of the key that
+    /// actually verified, never a raw artifact string that could disagree.
+    /// `verify_signature` derives the key from the artifact's `signer` field
+    /// and the parser is canonical (base58btc, exact length, ed25519
+    /// multicodec), so a forged field fails verification outright: the
+    /// must-not here is that no `VerifiedAttestation` ever carries a signer
+    /// that did not verify.
+    #[test]
+    fn verified_signer_is_anchored_to_the_key_that_verified() {
+        let signer = fresh();
+        let other = fresh();
+        assert_ne!(
+            signer.verifying_key().to_bytes(),
+            other.verifying_key().to_bytes(),
+            "fixture precondition: distinct signing keys"
+        );
+
+        let cert_hash = sample_hash();
+        let reg = Registry::new();
+
+        // Control: an honest artifact reports the signer's canonical DID, and
+        // it agrees with the artifact field.
+        let att = signed_demo(&signer, cert_hash, "ok");
+        let v = reg.verify(&att, cert_hash).unwrap();
+        let expected = did_key_from_verifying_key(&signer.verifying_key());
+        assert_eq!(v.signer, expected, "signer must be the verified key's DID");
+        assert_eq!(v.signer, att.signer, "honest artifact: field and key agree");
+
+        // Must-not, both directions: a signer field naming a DIFFERENT key
+        // than the one that signed must fail verification, so the artifact
+        // string can never redirect the reported signer.
+        let mut forged = signed_demo(&signer, cert_hash, "ok");
+        forged.signer = did_key_from_verifying_key(&other.verifying_key());
+        let err = reg.verify(&forged, cert_hash).unwrap_err();
+        assert!(
+            matches!(err, Error::Signature(_)),
+            "a signer field that does not match the signing key must fail, got {err:?}"
+        );
+
+        let mut forged_rev = signed_demo(&other, cert_hash, "ok");
+        forged_rev.signer = did_key_from_verifying_key(&signer.verifying_key());
+        let err_rev = reg.verify(&forged_rev, cert_hash).unwrap_err();
+        assert!(
+            matches!(err_rev, Error::Signature(_)),
+            "mismatched signer field must fail verification, got {err_rev:?}"
+        );
+    }
 }
@@ -1,6 +1,6 @@
 [package]
 name = "gitlawb-core"
-description = "Core cryptographic primitives for the gitlawb network: DIDs, CIDs, UCAN, HTTP Signatures, ref-update certificates"
+description = "Core cryptographic primitives for the Twigpine network: DIDs, CIDs, UCAN, HTTP Signatures, ref-update certificates"
 version = "0.7.1" # x-release-please-version
 edition.workspace = true
 rust-version.workspace = true
@@ -22,6 +22,7 @@ multihash-codetable = { workspace = true }
 cid = { workspace = true }
 chrono = { workspace = true }
 uuid = { workspace = true }
+url = { workspace = true, optional = true }
 zeroize = { version = "1", features = ["derive"] }
 pkcs8 = { version = "0.10", features = ["pem", "std"] }
 curve25519-dalek = "4"
@@ -30,3 +31,9 @@ chacha20poly1305 = "0.10"

 [dev-dependencies]
 tokio = { workspace = true }
+# Not optional here: the redirect matrix must run under a bare
+# `cargo test -p gitlawb-core`, with no feature selected.
+url = { workspace = true }
+
+[features]
+redirect = ["dep:url"]
@@ -5,6 +5,28 @@
 //!
 //! The schema is frozen at v1. All fields are mandatory for forward compatibility.
 //! Nodes that receive a certificate with an unknown version MUST reject it.
+//!
+//! ## Replay
+//!
+//! A certificate proves *who* asserted a ref update and *what* they asserted.
+//! It does not prove the assertion is fresh. [`RefUpdateBody::seq`] and
+//! [`RefUpdateBody::nonce`] are inside the signed bytes (see
+//! [`RefUpdateBody::to_signing_bytes`]), so a verifier can trust the values,
+//! but deciding whether a nonce is being reused or a `seq` is going backwards
+//! needs state.
+//!
+//! This crate is deliberately stateless — it is embedded in both nodes and
+//! clients — so replay is **not** enforced here and cannot be. Ownership of
+//! the check sits with the verifier that accepts and persists certificates,
+//! which must:
+//!
+//! - reject a `nonce` it has already accepted for the same `(repo, ref_name)`;
+//! - reject a `seq` that does not advance past the last accepted `seq` for
+//!   that `(repo, ref_name)`.
+//!
+//! Until such a verifier exists, replay of ref-update certificates is not
+//! enforced anywhere in this workspace. This contract exists so the first one
+//! inherits the requirement instead of re-deriving it.

 use std::collections::HashSet;

@@ -19,6 +41,30 @@ use crate::{Error, Result};
 /// The certificate type discriminant. Always `"gitlawb/ref-update/v1"`.
 pub const CERT_TYPE: &str = "gitlawb/ref-update/v1";

+/// Exact url-safe unpadded base64 length of a 64-byte Ed25519 signature.
+/// The tolerant threshold scan requires this canonical unpadded length before
+/// attempting decode: the length gate is a DoS bound that keeps an
+/// attacker-sized sig string from costing allocation before the signature is
+/// rejected. A longer string is not meaningless (the same 64 bytes can appear
+/// in padded or non-url-safe encodings) but no in-tree signer produces one,
+/// so the scan skips non-canonical entries rather than decoding them.
+const ENCODED_SIG_LEN: usize = 86;
+
+/// Hard ceiling on signature entries a certificate may carry into the
+/// tolerant threshold scan. A legitimate certificate holds one entry per
+/// maintainer (a re-sign after key rotation may briefly leave a second), so
+/// this is far above any real signer set, while turning the scan's worst case
+/// from "one curve verification per attacker-appended entry" into a small
+/// constant. Rejection is loud (`Err`), not a silent prefix scan: scanning
+/// only the first N entries would let junk placed ahead of the real
+/// signatures starve them out of the window, which is the #349 denial again
+/// in positional form. An attacker positioned to inflate the list past this
+/// ceiling can already suppress the certificate outright. The loud
+/// rejection does deny a legitimately over-long certificate; that is an
+/// explicit, detectable denial above the cap rather than a silent window
+/// junk placement could exploit.
+const MAX_SIGNATURE_ENTRIES: usize = 64;
+
 /// A signature on a ref-update certificate.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct CertSignature {
@@ -41,11 +87,19 @@ pub struct RefUpdateBody {
    pub from: String,
    /// The new commit hash (SHA-256 hex).
    pub to: String,
-    /// Monotonically increasing sequence number. Prevents replay.
+    /// Monotonically increasing sequence number, scoped to `(repo, ref_name)`.
+    ///
+    /// Signed, but not enforced by this crate: core holds no state. The
+    /// verifier that persists certificates must reject a `seq` that does not
+    /// advance for the same repo and ref. See the module docs on replay.
    pub seq: u64,
    /// RFC 3339 timestamp of this update.
    pub timestamp: DateTime<Utc>,
-    /// Random nonce for deduplication.
+    /// Random nonce, unique per certificate.
+    ///
+    /// Signed, but not enforced by this crate: core holds no state. The
+    /// verifier that persists certificates must reject a nonce it has already
+    /// accepted for the same repo and ref. See the module docs on replay.
    pub nonce: String,
 }

@@ -108,9 +162,15 @@ impl RefUpdateCert {
        Ok(())
    }

-    /// Verify all signatures on this certificate.
+    /// Verify all signatures on this certificate, failing closed.
    ///
-    /// Returns the list of DIDs whose signatures are valid.
+    /// On success returns the list of DIDs whose signatures are valid. This is
+    /// strict: it returns an error if *any* signature entry is malformed or
+    /// invalid. Because signature entries live outside the signed body (anyone
+    /// can append one without key material), do not use this to make a
+    /// threshold decision on an untrusted certificate — a single appended junk
+    /// entry would reject the whole cert. Use [`Self::satisfies_threshold`],
+    /// which ignores invalid entries and counts only the valid signers.
    pub fn verify_all(&self) -> Result<Vec<Did>> {
        use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine};
        let signing_bytes = self.body.to_signing_bytes()?;
@@ -139,12 +199,65 @@ impl RefUpdateCert {
    /// from the provided set of authorized maintainer DIDs.
    ///
    /// Counts distinct signer DIDs, not signature entries: a repeated
-    /// signature from the same maintainer counts once.
+    /// signature from the same maintainer counts once. Malformed and invalid
+    /// signature entries are skipped, not fatal — entries live outside the
+    /// signed body, so any third party can append junk without key material,
+    /// and short-circuiting on the first bad one (as [`Self::verify_all`]
+    /// does) would let that junk deny an otherwise valid certificate (#349).
+    ///
+    /// Tolerance must not become an unbounded verification path (#349 review
+    /// round 2): entries are attacker-extensible, so the work done for them is
+    /// bounded BEFORE the loop, not discovered inside it:
+    ///
+    /// - a list over [`MAX_SIGNATURE_ENTRIES`] is rejected outright (see the
+    ///   constant for why loud rejection beats a silent prefix scan);
+    /// - an encoded signature is length-gated before base64 work — a 64-byte
+    ///   Ed25519 signature is exactly [`ENCODED_SIG_LEN`] characters in
+    ///   url-safe unpadded base64, so an oversized blob never allocates;
+    /// - entries whose signer is not in `maintainers`, or is already counted,
+    ///   are skipped on a string compare, before any decode or curve work —
+    ///   so even within the cap, junk naming strangers costs no crypto;
+    /// - the scan returns as soon as the threshold is met.
    pub fn satisfies_threshold(&self, maintainers: &[Did], threshold: usize) -> Result<bool> {
-        let valid = self.verify_all()?;
-        let distinct_signers: HashSet<&Did> =
-            valid.iter().filter(|d| maintainers.contains(d)).collect();
-        Ok(distinct_signers.len() >= threshold)
+        use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine};
+        if threshold == 0 {
+            return Err(Error::RefCert("threshold must be at least 1".to_string()));
+        }
+        if self.signatures.len() > MAX_SIGNATURE_ENTRIES {
+            return Err(Error::RefCert(format!(
+                "certificate carries {} signature entries (max {})",
+                self.signatures.len(),
+                MAX_SIGNATURE_ENTRIES
+            )));
+        }
+        let signing_bytes = self.body.to_signing_bytes()?;
+        let mut counted: HashSet<&Did> = HashSet::new();
+
+        for cert_sig in &self.signatures {
+            if counted.contains(&cert_sig.signer) || !maintainers.contains(&cert_sig.signer) {
+                continue;
+            }
+            if cert_sig.sig.len() != ENCODED_SIG_LEN {
+                continue;
+            }
+            let Ok(vk) = cert_sig.signer.to_verifying_key() else {
+                continue;
+            };
+            let Ok(sig_bytes_vec) = URL_SAFE_NO_PAD.decode(&cert_sig.sig) else {
+                continue;
+            };
+            let Ok(sig_bytes) = <[u8; 64]>::try_from(sig_bytes_vec) else {
+                continue;
+            };
+            if verify(&vk, &signing_bytes, &sig_bytes).is_ok() {
+                counted.insert(&cert_sig.signer);
+                if counted.len() >= threshold {
+                    return Ok(true);
+                }
+            }
+        }
+
+        Ok(counted.len() >= threshold)
    }

    /// Validate the certificate structure (not signatures).
@@ -377,8 +490,151 @@ mod tests {
        assert!(!cert.satisfies_threshold(&maintainers, 2).unwrap());
    }

+    /// An attacker can append junk signature entries to a cert (they live
+    /// outside the signed body, so no key material is needed). `satisfies_threshold`
+    /// must ignore them and still count the real signers — before the fix,
+    /// `verify_all`'s `?` on the junk made the whole check error out, denying a
+    /// valid 2-of-2 cert.
    #[test]
-    fn threshold_zero_is_always_satisfied() {
+    fn satisfies_threshold_ignores_appended_junk_signatures() {
+        let kp1 = Keypair::generate();
+        let kp2 = Keypair::generate();
+        let mut cert = RefUpdateCert::new(
+            kp1.did(),
+            "refs/heads/main".to_string(),
+            dummy_hash('0'),
+            dummy_hash('a'),
+            1,
+            &kp1,
+        )
+        .unwrap();
+        cert.countersign(&kp2).unwrap();
+
+        // Junk sits AHEAD of the real entries so the scan must actually skip
+        // it: appended last, the threshold-2 early return at the second real
+        // signature would fire before any junk is examined and the test would
+        // stay green even if the skip logic never ran.
+        cert.signatures.insert(
+            0,
+            CertSignature {
+                signer: kp1.did(),
+                sig: "!!!not-base64!!!".to_string(),
+            },
+        );
+        // Valid base64 but only 3 bytes — fails the 64-byte length check.
+        cert.signatures.insert(
+            0,
+            CertSignature {
+                signer: kp2.did(),
+                sig: "AAAA".to_string(),
+            },
+        );
+        // Exactly ENCODED_SIG_LEN of valid base64 decoding to garbage —
+        // exercises the decode+verify path before the skip, not just the
+        // length gate.
+        cert.signatures.insert(
+            0,
+            CertSignature {
+                signer: kp1.did(),
+                sig: "A".repeat(ENCODED_SIG_LEN),
+            },
+        );
+
+        let maintainers = vec![kp1.did(), kp2.did()];
+        assert!(cert.satisfies_threshold(&maintainers, 2).unwrap());
+    }
+
+    /// A junk entry that names a real maintainer DID but carries a signature
+    /// that does not verify must not be counted — otherwise appending garbage
+    /// under a maintainer's DID could forge a threshold.
+    #[test]
+    fn satisfies_threshold_does_not_count_a_forged_maintainer_signature() {
+        let kp1 = Keypair::generate();
+        let kp2 = Keypair::generate();
+        let mut cert = RefUpdateCert::new(
+            kp1.did(),
+            "refs/heads/main".to_string(),
+            dummy_hash('0'),
+            dummy_hash('a'),
+            1,
+            &kp1,
+        )
+        .unwrap();
+        // Well-formed 64-byte signature from kp2, but over unrelated bytes, so
+        // it fails verification against the cert body.
+        cert.signatures.push(CertSignature {
+            signer: kp2.did(),
+            sig: kp2.sign_b64(b"unrelated bytes"),
+        });
+
+        let maintainers = vec![kp1.did(), kp2.did()];
+        // Only kp1 actually signed the body: 2-of-2 must fail, 1-of-2 holds.
+        assert!(!cert.satisfies_threshold(&maintainers, 2).unwrap());
+        assert!(cert.satisfies_threshold(&maintainers, 1).unwrap());
+    }
+
+    /// The tolerant scan must not be an unbounded verification path: a list
+    /// past MAX_SIGNATURE_ENTRIES is rejected loudly before any curve work,
+    /// while a list at the cap still verifies normally.
+    #[test]
+    fn satisfies_threshold_rejects_an_oversized_signature_list() {
+        let kp = Keypair::generate();
+        let mut cert = RefUpdateCert::new(
+            kp.did(),
+            "refs/heads/main".to_string(),
+            dummy_hash('0'),
+            dummy_hash('a'),
+            1,
+            &kp,
+        )
+        .unwrap();
+        let junk = || CertSignature {
+            signer: Keypair::generate().did(),
+            sig: "AAAA".to_string(),
+        };
+        while cert.signatures.len() < MAX_SIGNATURE_ENTRIES {
+            cert.signatures.push(junk());
+        }
+        let maintainers = vec![kp.did()];
+        // At the cap: still fine, the real signature counts.
+        assert!(cert.satisfies_threshold(&maintainers, 1).unwrap());
+        // One past the cap: loud rejection, no tolerant scan.
+        cert.signatures.push(junk());
+        let err = cert.satisfies_threshold(&maintainers, 1).unwrap_err();
+        assert!(err.to_string().contains("signature entries"));
+    }
+
+    /// An entry whose encoded signature is not the exact base64 length of a
+    /// 64-byte Ed25519 signature is skipped before any base64 decode — an
+    /// attacker-sized sig string must cost neither allocation nor curve work,
+    /// and must not affect the verdict.
+    #[test]
+    fn satisfies_threshold_skips_an_oversized_encoded_signature() {
+        let kp1 = Keypair::generate();
+        let kp2 = Keypair::generate();
+        let mut cert = RefUpdateCert::new(
+            kp1.did(),
+            "refs/heads/main".to_string(),
+            dummy_hash('0'),
+            dummy_hash('a'),
+            1,
+            &kp1,
+        )
+        .unwrap();
+        // Names a real maintainer, valid base64 alphabet, but megabytes long:
+        // the length gate must refuse it without decoding.
+        cert.signatures.push(CertSignature {
+            signer: kp2.did(),
+            sig: "A".repeat(2 * 1024 * 1024),
+        });
+
+        let maintainers = vec![kp1.did(), kp2.did()];
+        assert!(cert.satisfies_threshold(&maintainers, 1).unwrap());
+        assert!(!cert.satisfies_threshold(&maintainers, 2).unwrap());
+    }
+
+    #[test]
+    fn threshold_zero_is_rejected() {
        let kp = Keypair::generate();
        let cert = RefUpdateCert::new(
            kp.did(),
@@ -389,6 +645,120 @@ mod tests {
            &kp,
        )
        .unwrap();
-        assert!(cert.satisfies_threshold(&[], 0).unwrap());
+
+        // Zero must never read as "no signatures required", even with an empty
+        // maintainer set.
+        let err = cert.satisfies_threshold(&[], 0).unwrap_err();
+        assert!(
+            err.to_string().contains("threshold must be at least 1"),
+            "zero must be refused as an invalid threshold"
+        );
+    }
+
+    /// The threshold is checked before any signature work, so a zero is
+    /// reported as a threshold error even when the certificate's signatures
+    /// are also invalid. Otherwise a broken certificate would let a caller
+    /// mistake the failure for a legitimate verification result.
+    #[test]
+    fn threshold_zero_is_rejected_before_signature_verification() {
+        let kp = Keypair::generate();
+        let other = Keypair::generate();
+        let mut cert = RefUpdateCert::new(
+            kp.did(),
+            "refs/heads/main".to_string(),
+            dummy_hash('0'),
+            dummy_hash('a'),
+            1,
+            &kp,
+        )
+        .unwrap();
+        cert.signatures[0].sig = other.sign_b64(b"unrelated bytes");
+
+        let err = cert.satisfies_threshold(&[kp.did()], 0).unwrap_err();
+        assert!(
+            err.to_string().contains("threshold must be at least 1"),
+            "the invalid threshold must win over the invalid signature"
+        );
+    }
+
+    /// `seq` and `nonce` are the verifier's replay inputs, so they have to be
+    /// inside the signed body; tapping either one must invalidate the
+    /// certificate rather than leave it verifying with different values.
+    #[test]
+    fn replay_fields_are_covered_by_the_signature() {
+        let kp = Keypair::generate();
+        let cert = RefUpdateCert::new(
+            kp.did(),
+            "refs/heads/main".to_string(),
+            dummy_hash('0'),
+            dummy_hash('a'),
+            7,
+            &kp,
+        )
+        .unwrap();
+        cert.verify_all().expect("baseline must verify");
+
+        let mut bumped = cert.clone();
+        bumped.body.seq = 8;
+        assert!(
+            matches!(bumped.verify_all(), Err(Error::SignatureInvalid)),
+            "changing seq must invalidate the signature"
+        );
+
+        let mut renamed = cert.clone();
+        renamed.body.nonce = "different-nonce".to_string();
+        assert!(
+            matches!(renamed.verify_all(), Err(Error::SignatureInvalid)),
+            "changing nonce must invalidate the signature"
+        );
+    }
+
+    /// A signature minted by a real key must not verify when the signer is
+    /// spelled in another multibase alphabet. Otherwise the same identity
+    /// would satisfy a maintainer allowlist under several DID strings.
+    ///
+    /// Base64url is deliberate: its 47-character method-id fits under the
+    /// pre-existing length bound, so this test reaches the canonical-prefix
+    /// rule and fails on the pre-fix code instead of being refused for length.
+    #[test]
+    fn verify_all_rejects_a_non_canonical_did_spelling() {
+        let kp = Keypair::generate();
+        let mut cert = RefUpdateCert::new(
+            kp.did(),
+            "refs/heads/main".to_string(),
+            dummy_hash('0'),
+            dummy_hash('a'),
+            1,
+            &kp,
+        )
+        .unwrap();
+
+        let mut prefixed = vec![0xed, 0x01];
+        prefixed.extend_from_slice(&kp.verifying_key().to_bytes());
+        cert.signatures[0].signer = format!(
+            "did:key:{}",
+            multibase::encode(multibase::Base::Base64Url, &prefixed)
+        )
+        .parse()
+        .expect("method is still `key`");
+
+        let err = cert
+            .verify_all()
+            .expect_err("a non-canonical did:key spelling must not resolve");
+        assert!(
+            err.to_string().contains("base58btc"),
+            "rejection must come from the canonical-prefix rule, not the length guard"
+        );
+        assert!(
+            !matches!(
+                cert.satisfies_threshold(&[cert.signatures[0].signer.clone()], 1),
+                Ok(true)
+            ),
+            "the alternate spelling must not satisfy a threshold"
+        );
+
+        // The canonical spelling of the same key is unaffected.
+        cert.signatures[0].signer = kp.did();
+        assert!(cert.satisfies_threshold(&[kp.did()], 1).unwrap());
    }
 }
@@ -64,6 +64,19 @@ impl Cid {
    }
 }

+/// True when `s` parses as a CIDv1 with the raw codec — the exact shape
+/// [`Cid::from_git_object_bytes`] produces and the `/ipfs` resolver looks up.
+/// A legacy provider CID (Kubo dag-pb, Pinata CIDv0) parses to a different
+/// version or codec and returns `false`, marking it an opportunistic-repair
+/// candidate. Decidable from the string alone (no object bytes), so the pin path
+/// can gate the byte-read/recompute cost on it and leave non-legacy rows at the
+/// existing DB-only skip cost. An unparseable string is non-canonical (`false`).
+pub fn is_raw_cidv1(s: &str) -> bool {
+    s.parse::<CidGeneric<64>>()
+        .map(|c| c.version() == cid::Version::V1 && c.codec() == RAW)
+        .unwrap_or(false)
+}
+
 impl fmt::Display for Cid {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(f, "{}", self.0)
@@ -177,6 +190,38 @@ mod tests {
        assert!(result.is_err());
    }

+    #[test]
+    fn is_raw_cidv1_classifies_codec_from_string() {
+        // The canonical resolver key: CIDv1 + raw codec → not a repair candidate.
+        let raw = Cid::from_git_object_bytes(b"blob 5\0hello");
+        assert!(
+            is_raw_cidv1(raw.as_str()),
+            "from_git_object_bytes output is CIDv1/raw"
+        );
+
+        // A CIDv0 (Pinata dag-pb legacy shape) → repair candidate.
+        assert!(
+            !is_raw_cidv1("QmYwAPJzv5CZsnA625s3Xf2nemtYgPpHdWEz79ojWnPbdG"),
+            "a CIDv0 dag-pb value is a legacy-repair candidate"
+        );
+
+        // A CIDv1 with the dag-pb codec (the Kubo above-block-size root) over the
+        // same multihash → still a repair candidate (codec, not just version).
+        let parsed = raw.as_str().parse::<CidGeneric<64>>().unwrap();
+        const DAG_PB: u64 = 0x70;
+        let dagpb = CidGeneric::<64>::new_v1(DAG_PB, *parsed.hash()).to_string();
+        assert!(
+            !is_raw_cidv1(&dagpb),
+            "a CIDv1 dag-pb value is a legacy-repair candidate"
+        );
+
+        // Garbage is non-canonical.
+        assert!(
+            !is_raw_cidv1("not-a-cid"),
+            "an unparseable string is non-canonical"
+        );
+    }
+
    #[test]
    fn sha256_hex_of_empty_input_is_well_known() {
        // SHA-256("") is a fixed constant; verifies the hasher is wired correctly.
@@ -1,11 +1,11 @@
-//! Decentralized Identifier (DID) types for the gitlawb network.
+//! Decentralized Identifier (DID) types for the Twigpine network.
 //!
 //! Supported methods:
 //!   - `did:key`      — ephemeral, derived directly from an Ed25519 public key
 //!   - `did:web`      — anchored to a domain
 //!   - `did:gitlawb`  — native, anchored to the libp2p DHT
 //!
-//! The canonical DID for a gitlawb actor is `did:key` during bootstrap
+//! The canonical DID for a Twigpine actor is `did:key` during bootstrap
 //! and migrates to `did:gitlawb` once DHT anchoring is live.

 use ed25519_dalek::VerifyingKey;
@@ -20,6 +20,19 @@ use crate::{Error, Result};
 /// Value: 0xed01 encoded as a varint.
 const ED25519_MULTICODEC: &[u8] = &[0xed, 0x01];

+/// The only multibase prefix a `did:key` method-specific id may carry.
+///
+/// The DID:key spec encodes Ed25519 keys as base58btc, which is multibase
+/// `z`. `multibase::decode` happily accepts every other alphabet, and each
+/// spelling of the same key bytes would produce a distinct DID string while
+/// gates, allowlists, and database rows compare DIDs as strings — one
+/// identity would read as several. Rejecting the alternatives here keeps the
+/// mapping from `did:key` to key single-valued for every caller of
+/// [`Did::to_verifying_key`] (HTTP signatures, UCAN, ref certs, peer
+/// announces). `gitlawb-attest` enforces the same rule for attestation
+/// signers.
+const DID_KEY_MULTIBASE_PREFIX: char = 'z';
+
 /// A Decentralized Identifier.
 #[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
 #[serde(transparent)]
@@ -71,7 +84,14 @@ impl Did {

    /// Resolve the Ed25519 verifying key from a `did:key`.
    ///
-    /// Returns `Err` if the DID is not a `did:key` or the key bytes are invalid.
+    /// Only the canonical base58btc (`z`) spelling is accepted. Other
+    /// multibase encodings decode to the same key but spell a different DID
+    /// string, so accepting them would let one identity appear under several
+    /// identifiers to string-comparing gates and allowlists.
+    ///
+    /// Returns `Err` if the DID is not a `did:key`, uses a non-canonical
+    /// multibase prefix, the key bytes are invalid, or the key is a
+    /// small-order (weak) point.
    pub fn to_verifying_key(&self) -> Result<VerifyingKey> {
        if !self.is_did_key() {
            return Err(Error::InvalidDid(format!(
@@ -93,9 +113,27 @@ impl Did {
            ));
        }

-        let (_, bytes) =
+        // Cheap prefix gate before the decode, then the base check on what
+        // multibase actually parsed. The first keeps a clearly non-canonical id
+        // off the decoder. The second is belt-and-braces — multibase maps a
+        // leading `z` only to `Base58Btc` — but it states the requirement
+        // directly instead of relying on that mapping, and `gitlawb-attest`
+        // makes the same pair of checks.
+        if !self.method_id().starts_with(DID_KEY_MULTIBASE_PREFIX) {
+            return Err(Error::InvalidDid(
+                "did:key must use base58btc (z-prefix)".to_string(),
+            ));
+        }
+
+        let (base, bytes) =
            multibase::decode(self.method_id()).map_err(|e| Error::InvalidDid(e.to_string()))?;

+        if base != Base::Base58Btc {
+            return Err(Error::InvalidDid(
+                "did:key must use base58btc (z-prefix)".to_string(),
+            ));
+        }
+
        if !bytes.starts_with(ED25519_MULTICODEC) {
            return Err(Error::InvalidDid(
                "not an ed25519 multicodec key".to_string(),
@@ -106,7 +144,25 @@ impl Did {
            .try_into()
            .map_err(|_| Error::InvalidDid("ed25519 key must be 32 bytes".to_string()))?;

-        VerifyingKey::from_bytes(&key_bytes).map_err(|e| Error::InvalidDid(e.to_string()))
+        let key =
+            VerifyingKey::from_bytes(&key_bytes).map_err(|e| Error::InvalidDid(e.to_string()))?;
+
+        // `from_bytes` only decompresses, so it accepts a small-order point.
+        // Such a key satisfies the verification equation for any message, and
+        // every X25519 shared secret derived from it is the all-zero key: the
+        // order-1 and order-2 points convert to the all-zero Montgomery u
+        // directly, and the order-4 and order-8 points are annihilated by the
+        // scalar clamping instead. `is_weak` covers the whole torsion set.
+        // Mixed-order points are correctly NOT rejected, since clamping clears
+        // the cofactor component. Rejecting
+        // at resolution is what makes this the choke point: every consumer that
+        // resolves a DID through here inherits the rejection, and a recipient
+        // that cannot resolve already fails closed downstream.
+        if key.is_weak() {
+            return Err(Error::InvalidDid("small-order ed25519 key".to_string()));
+        }
+
+        Ok(key)
    }

    /// Return the full DID string as a `&str`.
@@ -152,7 +208,7 @@ impl From<Did> for String {
    }
 }

-/// A DID Document for a gitlawb actor.
+/// A DID Document for a Twigpine actor.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct DidDocument {
    #[serde(rename = "@context")]
@@ -369,4 +425,154 @@ mod tests {
        );
        did.to_verifying_key().expect("a real did:key must resolve");
    }
+
+    /// `multibase::decode` accepts every alphabet, so the same Ed25519 key can
+    /// be spelled as several `did:key` strings. Only base58btc is canonical;
+    /// the rest must be refused so one identity cannot appear as several to
+    /// gates and allowlists that compare DID strings.
+    #[test]
+    fn to_verifying_key_rejects_non_base58btc_multibase() {
+        use ed25519_dalek::SigningKey;
+        let signing = SigningKey::from_bytes(&[9u8; 32]);
+        let mut prefixed = ED25519_MULTICODEC.to_vec();
+        prefixed.extend_from_slice(&signing.verifying_key().to_bytes());
+
+        let canonical = Did::from_verifying_key(&signing.verifying_key());
+        assert_eq!(
+            canonical.to_verifying_key().unwrap().to_bytes(),
+            signing.verifying_key().to_bytes(),
+            "base58btc is the accepted spelling"
+        );
+
+        // These spellings fit inside the method-id bound, so they reach the
+        // canonical-prefix check and must be refused there.
+        for base in [Base::Base64Url, Base::Base32Lower] {
+            let alternative = format!("did:key:{}", multibase::encode(base, &prefixed));
+            let did: Did = alternative.parse().expect("method is still `key`");
+            assert_ne!(
+                did.as_str(),
+                canonical.as_str(),
+                "the same key must not compare equal under another spelling"
+            );
+
+            let err = did
+                .to_verifying_key()
+                .expect_err("non-canonical multibase must be refused");
+            assert!(
+                matches!(err, Error::InvalidDid(_)),
+                "must be an InvalidDid, got {err:?}"
+            );
+            assert!(
+                err.to_string().contains("base58btc"),
+                "the error must name the canonical encoding: {err}"
+            );
+        }
+
+        // A base16 spelling of the same key is longer than the method-id
+        // bound, so it is refused by the length guard before the prefix check.
+        // Either way it must not resolve.
+        for base in [Base::Base16Lower, Base::Base16Upper] {
+            let alternative = format!("did:key:{}", multibase::encode(base, &prefixed));
+            let did: Did = alternative.parse().expect("method is still `key`");
+            assert!(
+                matches!(did.to_verifying_key(), Err(Error::InvalidDid(_))),
+                "a base16 spelling of a real key must not resolve"
+            );
+        }
+    }
+
+    /// base58flickr (`Z`) is the non-bitcoin sibling alphabet. Its spelling of
+    /// a real key fits under the length bound, so it must be refused for the
+    /// prefix, not merely for its size.
+    #[test]
+    fn to_verifying_key_rejects_a_base58flickr_spelling() {
+        use ed25519_dalek::SigningKey;
+        let signing = SigningKey::from_bytes(&[11u8; 32]);
+        let mut prefixed = ED25519_MULTICODEC.to_vec();
+        prefixed.extend_from_slice(&signing.verifying_key().to_bytes());
+
+        let did: Did = format!(
+            "did:key:{}",
+            multibase::encode(Base::Base58Flickr, &prefixed)
+        )
+        .parse()
+        .expect("method is still `key`");
+
+        let err = did
+            .to_verifying_key()
+            .expect_err("base58flickr must be refused");
+        assert!(
+            err.to_string().contains("base58btc"),
+            "the error must name the canonical encoding: {err}"
+        );
+    }
+
+    /// An empty method-specific id must be refused by the prefix rule without
+    /// panicking on a missing first character.
+    #[test]
+    fn to_verifying_key_rejects_an_empty_method_id() {
+        let did: Did = "did:key:".parse().expect("method is still `key`");
+        let err = did
+            .to_verifying_key()
+            .expect_err("an empty method id must be refused");
+        assert!(
+            err.to_string().contains("base58btc"),
+            "the error must name the canonical encoding: {err}"
+        );
+    }
+
+    /// A canonical base58btc spelling is not enough on its own: the bytes must
+    /// also carry the ed25519-pub multicodec. This pins the check that stops a
+    /// key of another type from resolving as an Ed25519 key.
+    #[test]
+    fn to_verifying_key_rejects_base58btc_with_a_wrong_multicodec() {
+        use ed25519_dalek::SigningKey;
+        let signing = SigningKey::from_bytes(&[12u8; 32]);
+        let mut buf = vec![0x12, 0x00]; // sha2-256, not ed25519-pub
+        buf.extend_from_slice(&signing.verifying_key().to_bytes());
+
+        let did: Did = format!("did:key:{}", multibase::encode(Base::Base58Btc, &buf))
+            .parse()
+            .expect("method is still `key`");
+
+        let err = did
+            .to_verifying_key()
+            .expect_err("a non-ed25519 multicodec must be refused");
+        assert!(
+            err.to_string().contains("ed25519"),
+            "the error must name the expected multicodec: {err}"
+        );
+    }
+
+    /// A `did:key` can encode a small-order (weak) Ed25519 point. Such a key
+    /// satisfies the verification equation for any message and, converted to
+    /// Montgomery form for X25519, yields the all-zero shared secret, so an
+    /// envelope sealed to it is decryptable by anyone. Resolution is the choke
+    /// point every `Did`-based consumer shares, so it fails closed here.
+    #[test]
+    fn to_verifying_key_rejects_a_small_order_key() {
+        // The compressed identity point, the canonical small-order key.
+        let mut weak = [0u8; 32];
+        weak[0] = 1;
+        let mut prefixed = Vec::with_capacity(ED25519_MULTICODEC.len() + 32);
+        prefixed.extend_from_slice(ED25519_MULTICODEC);
+        prefixed.extend_from_slice(&weak);
+        let did: Did = format!(
+            "did:key:{}",
+            multibase::encode(multibase::Base::Base58Btc, &prefixed)
+        )
+        .parse()
+        .expect("a small-order did:key is still well-formed");
+
+        let err = did
+            .to_verifying_key()
+            .expect_err("a small-order did:key must not resolve");
+        match err {
+            Error::InvalidDid(msg) => assert!(
+                msg.contains("small-order"),
+                "rejection must name the small-order key, got: {msg}"
+            ),
+            other => panic!("expected Error::InvalidDid, got {other:?}"),
+        }
+    }
 }
@@ -11,6 +11,22 @@ use zeroize::Zeroizing;
 /// X25519 public key (Montgomery u) for an Ed25519 verifying key.
 fn x25519_public(vk: &VerifyingKey) -> Result<[u8; 32]> {
    use curve25519_dalek::edwards::CompressedEdwardsY;
+
+    // Every X25519 shared secret derived from a small-order point is the
+    // all-zero key: order-1 and order-2 convert to the all-zero Montgomery u
+    // directly, order-4 and order-8 are annihilated by the scalar clamping in
+    // x25519_secret_from_seed. Either way the per-recipient wrap could be
+    // rebuilt by anyone. Resolution already refuses such a key (see
+    // Did::to_verifying_key); this guard covers the SEAL side on its own terms
+    // for a caller that obtained the key some other way. The open side
+    // (open_blob's attacker-supplied `eph`) skips any entry whose exchange with
+    // this reader yields the all-zero shared secret
+    // (`yields_all_zero_shared_secret`), regardless of whether the bytes decode
+    // to an Edwards point.
+    if vk.is_weak() {
+        return Err(anyhow::anyhow!("verifying key is a small-order point"));
+    }
+
    let edwards = CompressedEdwardsY::from_slice(vk.as_bytes())
        .ok()
        .and_then(|c| c.decompress())
@@ -18,6 +34,26 @@ fn x25519_public(vk: &VerifyingKey) -> Result<[u8; 32]> {
    Ok(edwards.to_montgomery().to_bytes())
 }

+/// True when the X25519 exchange between an attacker-supplied `u` and this
+/// reader's scalar yields the all-zero shared secret.
+///
+/// Asked of the RESULT, not of the encoding, and that distinction is the whole
+/// guard. The previous version decompressed `u` to an Edwards point and asked
+/// `is_small_order`, which reports "safe" for any low-order encoding that does
+/// not decompress at all. Measured across the seven standard low-order
+/// encodings, six were caught and `u = p - 1` was not: it fails to decompress,
+/// so the check returned false, and the exchange is still the all-zero secret.
+/// An entry built on it unwraps for every reader.
+///
+/// Enumerating encodings cannot be exhaustive, because the attack depends on the
+/// shared secret rather than on the bytes that produced it, and twist and
+/// non-canonical inputs keep arriving. Asking the result needs no list.
+fn yields_all_zero_shared_secret(u: &[u8; 32], scalar: &[u8; 32]) -> bool {
+    use curve25519_dalek::montgomery::MontgomeryPoint;
+
+    MontgomeryPoint(*u).mul_clamped(*scalar).to_bytes() == [0u8; 32]
+}
+
 /// X25519 secret scalar for an Ed25519 seed (SHA-512 of seed, lower 32, clamped).
 /// Returns the scalar wrapped in `Zeroizing`, and scrubs the intermediate
 /// SHA-512 digest, so no copy of this secret material lingers in freed memory.
@@ -49,6 +85,13 @@ use serde::{Deserialize, Serialize};
 const MAGIC: &[u8] = b"GLENC";
 const VERSION: u8 = 2;

+/// Upper bound on the blinded recipient list. The envelope author controls
+/// this count and `open_blob` pays one X25519 exchange per entry before it can
+/// answer "not a recipient", so the count has to be capped rather than
+/// optimized (#419). Legitimate envelopes carry a handful of authorized
+/// readers, so 256 is far above real use.
+const MAX_RECIPIENTS: usize = 256;
+
 #[derive(Serialize, Deserialize)]
 struct Recipient {
    eph: String,   // base64 ephemeral x25519 pubkey (32B)
@@ -68,6 +111,14 @@ pub fn seal_blob(plaintext: &[u8], recipients: &[VerifyingKey]) -> Result<Vec<u8
    if recipients.is_empty() {
        return Err(anyhow::anyhow!("seal_blob: no recipients"));
    }
+    // Same bound open_blob enforces: an envelope over the cap could never be
+    // opened, so refuse to seal it (#419).
+    if recipients.len() > MAX_RECIPIENTS {
+        return Err(anyhow::anyhow!(
+            "seal_blob: {} recipients, over the {MAX_RECIPIENTS} cap",
+            recipients.len()
+        ));
+    }
    let mut content_key = [0u8; 32];
    OsRng.fill_bytes(&mut content_key);
    let body_cipher = XChaCha20Poly1305::new_from_slice(&content_key)
@@ -126,9 +177,20 @@ pub fn open_blob(envelope: &[u8], keypair: &Keypair) -> Result<Vec<u8>> {
    let header: Header =
        serde_json::from_slice(envelope.get(p..p + hlen).context("truncated header")?)
            .context("decode header")?;
+    // Reject before the recipient scan: every entry below is an X25519
+    // exchange, and a non-recipient would run the whole list (#419).
+    if header.recipients.len() > MAX_RECIPIENTS {
+        return Err(anyhow::anyhow!(
+            "envelope carries {} recipients, over the {MAX_RECIPIENTS} cap",
+            header.recipients.len()
+        ));
+    }
    let body = &envelope[p + hlen..];

-    let my_x = XSecret::from(*x25519_secret_from_seed(&keypair.to_seed()));
+    // The raw scalar is kept alongside the box secret so the exchange can be
+    // tested BEFORE a box is built from an attacker-supplied ephemeral.
+    let my_x_scalar = x25519_secret_from_seed(&keypair.to_seed());
+    let my_x = XSecret::from(*my_x_scalar);

    // Identities are blinded: no entry says which recipient it belongs to, so
    // try each one. The ChaChaBox AEAD tag authenticates, so exactly the
@@ -140,7 +202,12 @@ pub fn open_blob(envelope: &[u8], keypair: &Keypair) -> Result<Vec<u8>> {
            .ok()
            .and_then(|b| <[u8; 32]>::try_from(b.as_slice()).ok())
        {
-            Some(b) => XPublic::from(b),
+            // An ephemeral whose exchange with this reader is the all-zero
+            // shared secret would unwrap for anyone, so the entry is skipped.
+            // Tested on the exchange itself: an encoding-shaped check misses
+            // every low-order input that does not decompress to Edwards.
+            Some(b) if !yields_all_zero_shared_secret(&b, &my_x_scalar) => XPublic::from(b),
+            Some(_) => continue,
            None => continue,
        };
        // from_slice panics on a wrong length, and the envelope is attacker
@@ -293,4 +360,436 @@ mod tests {
        header["nonce"] = bad_nonce;
        assert!(open_blob(&reframe(&header), &reader).is_err());
    }
+
+    #[test]
+    fn over_cap_recipient_list_is_rejected_before_the_scan() {
+        // #419: the envelope author controls the recipient count and each entry
+        // costs an X25519 exchange, so the list is capped. Reframe a valid
+        // envelope with MAX_RECIPIENTS + 1 junk entries.
+        let reader = Keypair::generate();
+        let env = seal_blob(b"private blob contents", &[reader.verifying_key()]).unwrap();
+
+        let mut p = MAGIC.len() + 1;
+        let hlen = u32::from_le_bytes(env[p..p + 4].try_into().unwrap()) as usize;
+        p += 4;
+        let header_bytes = &env[p..p + hlen];
+        let body = &env[p + hlen..];
+
+        let mut header: serde_json::Value = serde_json::from_slice(header_bytes).unwrap();
+        let junk = serde_json::json!({
+            "eph": B64.encode([0u8; 32]),
+            "nonce": B64.encode([0u8; 24]),
+            "wrap": B64.encode([0u8; 48]),
+        });
+        header["recipients"] = serde_json::Value::Array(vec![junk.clone(); MAX_RECIPIENTS + 1]);
+
+        let hj = serde_json::to_vec(&header).unwrap();
+        let mut fat = Vec::new();
+        fat.extend_from_slice(MAGIC);
+        fat.push(VERSION);
+        fat.extend_from_slice(&(hj.len() as u32).to_le_bytes());
+        fat.extend_from_slice(&hj);
+        fat.extend_from_slice(body);
+
+        let err = open_blob(&fat, &reader).unwrap_err();
+        assert!(
+            err.to_string().contains("recipients"),
+            "expected the recipient-cap error, got: {err}"
+        );
+
+        // Exactly at the cap the envelope is still processed end to end.
+        header["recipients"] = serde_json::Value::Array(vec![junk; MAX_RECIPIENTS]);
+        let hj = serde_json::to_vec(&header).unwrap();
+        let mut at_cap = Vec::new();
+        at_cap.extend_from_slice(MAGIC);
+        at_cap.push(VERSION);
+        at_cap.extend_from_slice(&(hj.len() as u32).to_le_bytes());
+        at_cap.extend_from_slice(&hj);
+        at_cap.extend_from_slice(body);
+        let err = open_blob(&at_cap, &reader).unwrap_err();
+        assert!(
+            !err.to_string().contains("over the"),
+            "at-cap envelope must reach the scan, got: {err}"
+        );
+    }
+
+    #[test]
+    fn seal_refuses_a_recipient_list_over_the_cap() {
+        // An envelope over the cap can never be opened, so seal must reject it
+        // instead of minting one.
+        let keys: Vec<VerifyingKey> = (0..MAX_RECIPIENTS + 1)
+            .map(|_| Keypair::generate().verifying_key())
+            .collect();
+        let err = seal_blob(b"blob", &keys).unwrap_err();
+        assert!(
+            err.to_string().contains("recipients"),
+            "expected the recipient-cap error, got: {err}"
+        );
+    }
+
+    /// Every standard low-order X25519 encoding must be detected by
+    /// `yields_all_zero_shared_secret`, including encodings that do not
+    /// decompress to an Edwards point (notably `u = p - 1`). The open side
+    /// rejects on the exchange result, not on decode shape, because an
+    /// encoding-shaped check misses non-decompressing low-order inputs while
+    /// still yielding the all-zero shared secret.
+    #[test]
+    fn all_zero_shared_secret_is_detected_for_every_standard_low_order_encoding() {
+        // The seven standard X25519 low-order encodings, driven as a set rather
+        // than as the two that happen to decompress. The encoding-shaped check
+        // this replaced reported `p - 1` as safe, and its exchange is all-zero.
+        let mut one = [0u8; 32];
+        one[0] = 1;
+        let hex = |h: &str| -> [u8; 32] {
+            let mut out = [0u8; 32];
+            for i in 0..32 {
+                out[i] = u8::from_str_radix(&h[i * 2..i * 2 + 2], 16).unwrap();
+            }
+            out
+        };
+        let vectors = [
+            ("u=0", [0u8; 32]),
+            ("u=1", one),
+            (
+                "order8-a",
+                hex("e0eb7a7c3b41b8ae1656e3faf19fc46ada098deb9c32b1fd866205165f49b800"),
+            ),
+            (
+                "order8-b",
+                hex("5f9c95bca3508c24b1d0b1559c83ef5b04445cc4581c8e86d8224eddd09f1157"),
+            ),
+            (
+                "p-1",
+                hex("ecffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff7f"),
+            ),
+            (
+                "p",
+                hex("edffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff7f"),
+            ),
+            (
+                "p+1",
+                hex("eeffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff7f"),
+            ),
+        ];
+        let kp = Keypair::generate();
+        let scalar = x25519_secret_from_seed(&kp.to_seed());
+        for (name, u) in vectors {
+            assert!(
+                yields_all_zero_shared_secret(&u, &scalar),
+                "{name} is a low-order encoding and must be detected"
+            );
+        }
+
+        // And the other direction, or a guard that returned true always would
+        // pass everything above while making the envelope unopenable.
+        let real_u = x25519_public(&kp.verifying_key()).unwrap();
+        assert!(
+            !yields_all_zero_shared_secret(&real_u, &scalar),
+            "a real key must not be treated as low-order"
+        );
+    }
+
+    /// The open side must skip an entry whose header-supplied `eph` is a
+    /// low-order point: such an entry forces the all-zero shared secret and
+    /// would unwrap for any reader. A poisoned entry among honest ones must
+    /// not DoS the envelope — the honest entry still opens.
+    #[test]
+    fn open_blob_skips_a_low_order_ephemeral_and_still_opens_honest_entry() {
+        use curve25519_dalek::montgomery::MontgomeryPoint;
+        let reader = Keypair::generate();
+        let env = seal_blob(b"private blob contents", &[reader.verifying_key()]).unwrap();
+
+        // Split the envelope framing into header JSON and body.
+        let mut p = MAGIC.len() + 1;
+        let hlen = u32::from_le_bytes(env[p..p + 4].try_into().unwrap()) as usize;
+        p += 4;
+        let header_bytes = &env[p..p + hlen];
+        let body = &env[p + hlen..];
+
+        // A low-order Montgomery u (u = 0) encoded as the entry's eph.
+        let low_order_u = MontgomeryPoint([0u8; 32]).to_bytes();
+        let poisoned = serde_json::json!({
+            "eph": B64.encode(low_order_u),
+            "nonce": B64.encode([0u8; 24]),
+            "wrap": B64.encode([0u8; 32]),
+        });
+
+        let reframe = |header: &serde_json::Value| -> Vec<u8> {
+            let hj = serde_json::to_vec(header).unwrap();
+            let mut out = Vec::new();
+            out.extend_from_slice(MAGIC);
+            out.push(VERSION);
+            out.extend_from_slice(&(hj.len() as u32).to_le_bytes());
+            out.extend_from_slice(&hj);
+            out.extend_from_slice(body);
+            out
+        };
+
+        let mut header: serde_json::Value = serde_json::from_slice(header_bytes).unwrap();
+        // Prepend the poisoned entry. The reader's honest entry must still
+        // unwrap, so the envelope opens despite the low-order entry.
+        let recipients = header["recipients"].as_array_mut().unwrap();
+        recipients.insert(0, poisoned);
+        assert_eq!(
+            open_blob(&reframe(&header), &reader).unwrap(),
+            b"private blob contents",
+            "the honest entry must still open despite a low-order eph entry"
+        );
+    }
+
+    /// Must-not for the open-side guard: an envelope whose ONLY entry is built
+    /// on a low-order ephemeral must not open for any reader. Without the
+    /// guard the all-zero shared secret unwraps for everyone, so the attacker
+    /// controls the plaintext and the envelope's confidentiality is void.
+    #[test]
+    fn open_blob_rejects_a_poisoned_only_envelope() {
+        use chacha20poly1305::{aead::Aead, KeyInit, XChaCha20Poly1305, XNonce};
+        use crypto_box::aead::AeadCore;
+        use curve25519_dalek::montgomery::MontgomeryPoint;
+
+        // Build a poisoned envelope from scratch: one recipient entry whose
+        // eph is the low-order u = 0. The wrap is the encryption of a content
+        // key under the all-zero shared secret, so without the guard any
+        // reader's ChaChaBox against u = 0 decrypts it.
+        let mut content_key: [u8; 32] = Default::default();
+        OsRng.fill_bytes(&mut content_key);
+        let attacker_x = XSecret::generate(&mut OsRng);
+        let zero_box = ChaChaBox::new(&XPublic::from(<[u8; 32]>::default()), &attacker_x);
+        let nonce = ChaChaBox::generate_nonce(&mut OsRng);
+        let wrap = zero_box.encrypt(&nonce, &content_key[..]).unwrap();
+
+        let mut body_nonce: [u8; 24] = Default::default();
+        OsRng.fill_bytes(&mut body_nonce);
+        let body_cipher = XChaCha20Poly1305::new_from_slice(&content_key).unwrap();
+        let body = body_cipher
+            .encrypt(
+                XNonce::from_slice(&body_nonce),
+                b"attacker-chosen plaintext".as_slice(),
+            )
+            .unwrap();
+
+        let header = serde_json::json!({
+            "alg": "xchacha20poly1305",
+            "nonce": B64.encode(body_nonce),
+            "recipients": [{
+                "eph": B64.encode(MontgomeryPoint([0u8; 32]).to_bytes()),
+                "nonce": B64.encode(nonce),
+                "wrap": B64.encode(wrap),
+            }],
+        });
+        let header_json = serde_json::to_vec(&header).unwrap();
+        let mut env = Vec::new();
+        env.extend_from_slice(MAGIC);
+        env.push(VERSION);
+        env.extend_from_slice(&(header_json.len() as u32).to_le_bytes());
+        env.extend_from_slice(&header_json);
+        env.extend_from_slice(&body);
+
+        let reader = Keypair::generate();
+        let err = open_blob(&env, &reader).unwrap_err();
+        assert!(
+            err.to_string().contains("not a recipient"),
+            "a poisoned-only envelope must not open for anyone, got: {err}"
+        );
+    }
+
+    /// The same must-not, driven with the low-order encoding that does NOT map
+    /// to an Edwards point.
+    ///
+    /// This is the case an encoding-shaped guard cannot see. `u = p - 1` is a
+    /// legal X25519 input whose exchange is the all-zero secret for any scalar,
+    /// and `MontgomeryPoint::to_edwards` returns None for it, so a check written
+    /// as "decompress, then ask is_small_order" reports it as safe. Measured
+    /// across the seven standard low-order encodings: six are caught that way
+    /// and this one is not.
+    ///
+    /// Found by a cross-family review of the head that first added the guard,
+    /// and the comment beside that guard asserted this case was already handled
+    /// by a decode path that does not exist. That is why the fix rejects on the
+    /// RESULT of the exchange rather than on a list of encodings: the result is
+    /// what the attack actually depends on, and it needs no enumeration to be
+    /// exhaustive.
+    #[test]
+    fn open_blob_rejects_a_non_edwards_low_order_ephemeral() {
+        // No blanket `use ...::Aead` here. Two aead versions are reachable, and a
+        // blanket import silently resolved the wrap through the one open_blob does
+        // NOT decrypt with, producing a ciphertext that could never open. The test
+        // then passed with the guard, without it, and with the original pre-fix
+        // guard alike, proving nothing. Every call below names its trait.
+        use chacha20poly1305::{KeyInit, XChaCha20Poly1305, XNonce};
+        use crypto_box::aead::AeadCore;
+
+        // u = p - 1 = 2^255 - 20, little endian.
+        let mut eph = [0xffu8; 32];
+        eph[0] = 0xec;
+        eph[31] = 0x7f;
+
+        let mut content_key: [u8; 32] = Default::default();
+        OsRng.fill_bytes(&mut content_key);
+        let attacker_x = XSecret::generate(&mut OsRng);
+        let zero_box = ChaChaBox::new(&XPublic::from(<[u8; 32]>::default()), &attacker_x);
+        let nonce = ChaChaBox::generate_nonce(&mut OsRng);
+        let wrap = crypto_box::aead::Aead::encrypt(&zero_box, &nonce, &content_key[..]).unwrap();
+
+        let mut body_nonce: [u8; 24] = Default::default();
+        OsRng.fill_bytes(&mut body_nonce);
+        let body_cipher = XChaCha20Poly1305::new_from_slice(&content_key).unwrap();
+        let body = chacha20poly1305::aead::Aead::encrypt(
+            &body_cipher,
+            XNonce::from_slice(&body_nonce),
+            b"attacker-chosen plaintext".as_slice(),
+        )
+        .unwrap();
+
+        let header = serde_json::json!({
+            "alg": "xchacha20poly1305",
+            "nonce": B64.encode(body_nonce),
+            "recipients": [{
+                "eph": B64.encode(eph),
+                "nonce": B64.encode(nonce),
+                "wrap": B64.encode(wrap),
+            }],
+        });
+        let header_json = serde_json::to_vec(&header).unwrap();
+        let mut env = Vec::new();
+        env.extend_from_slice(MAGIC);
+        env.push(VERSION);
+        env.extend_from_slice(&(header_json.len() as u32).to_le_bytes());
+        env.extend_from_slice(&header_json);
+        env.extend_from_slice(&body);
+
+        let reader = Keypair::generate();
+        let err = open_blob(&env, &reader).unwrap_err();
+        assert!(
+            err.to_string().contains("not a recipient"),
+            "an envelope whose only entry uses a non-Edwards low-order eph must not \
+             open for anyone, got: {err}"
+        );
+    }
+
+    /// The compressed identity point: a well-formed Ed25519 encoding that is
+    /// small-order, so every shared secret derived from it is all-zero.
+    fn weak_verifying_key() -> VerifyingKey {
+        let mut weak = [0u8; 32];
+        weak[0] = 1;
+        let vk = VerifyingKey::from_bytes(&weak).expect("identity point decompresses");
+        assert!(vk.is_weak(), "fixture precondition: key is small-order");
+        vk
+    }
+
+    /// A small-order recipient key converts to Montgomery u = 0, and X25519
+    /// against u = 0 is the all-zero shared secret for ANY scalar, so the
+    /// wrapping box is reconstructable with no secret at all. Reject at the
+    /// primitive so the seal side is safe regardless of how the caller
+    /// obtained the key.
+    #[test]
+    fn x25519_public_rejects_a_small_order_key() {
+        let weak_vk = weak_verifying_key();
+
+        assert!(
+            x25519_public(&weak_vk).is_err(),
+            "a small-order key must not yield an x25519 public key"
+        );
+    }
+
+    /// Control: a legitimate key still converts, and to a non-zero u.
+    #[test]
+    fn x25519_public_still_accepts_a_real_key() {
+        let kp = Keypair::generate();
+        let u = x25519_public(&kp.verifying_key()).expect("a real key must convert");
+        assert_ne!(u, [0u8; 32], "a real key must not map to the zero point");
+    }
+
+    /// The guard has to propagate: no envelope may be produced for a weak
+    /// recipient even when the DID choke point is bypassed entirely.
+    #[test]
+    fn seal_blob_refuses_a_small_order_recipient() {
+        let weak_vk = weak_verifying_key();
+
+        assert!(
+            seal_blob(b"withheld", &[weak_vk]).is_err(),
+            "sealing to a weak recipient must fail"
+        );
+
+        // And a mixed set must fail too: one weak recipient exposes the shared
+        // content key, so a partial envelope is not an acceptable outcome.
+        let honest = Keypair::generate();
+        assert!(
+            seal_blob(b"withheld", &[honest.verifying_key(), weak_vk]).is_err(),
+            "a mixed honest+weak recipient set must fail closed, not seal partially"
+        );
+    }
+
+    /// Control for the two above: the ordinary seal/open round trip is intact.
+    #[test]
+    fn legit_seal_open_round_trip_still_works() {
+        let reader = Keypair::generate();
+        let env = seal_blob(b"withheld blob", &[reader.verifying_key()]).expect("seal");
+        assert_eq!(open_blob(&env, &reader).expect("open"), b"withheld blob");
+    }
+
+    /// The defect this fix exists to close, kept as an executable regression
+    /// rather than an assertion about absence. It runs the real attack: craft a
+    /// small-order did:key, get it into a recipient set alongside an honest
+    /// reader, then rebuild the wrapping box from the all-zero shared secret
+    /// that a small-order recipient forces. If any guard regresses, this does
+    /// not merely fail, it fails printing the plaintext it recovered.
+    #[test]
+    fn attacker_cannot_recover_plaintext_via_a_weak_recipient() {
+        use crate::did::Did;
+        use std::str::FromStr;
+
+        let weak_vk = weak_verifying_key();
+
+        // Layer 1: the attacker's did:key string must not resolve at all.
+        let weak_did = Did::from_verifying_key(&weak_vk).to_string();
+        assert!(
+            Did::from_str(&weak_did)
+                .expect("still a well-formed did:key")
+                .to_verifying_key()
+                .is_err(),
+            "a small-order did:key must not resolve"
+        );
+
+        // Layer 2: even handed the key directly, sealing must refuse. One weak
+        // recipient would expose the single shared content key, so the honest
+        // reader's blob would be readable by anyone.
+        let honest = Keypair::generate();
+        let secret_plaintext = b"WITHHELD BLOB CONTENTS";
+        let envelope = match seal_blob(secret_plaintext, &[honest.verifying_key(), weak_vk]) {
+            Err(_) => return, // no envelope exists, nothing to attack
+            Ok(env) => env,
+        };
+
+        // Only reachable if a guard regressed. Run the attack and report what
+        // it got, so the failure names the actual exposure.
+        let mut p = MAGIC.len() + 1;
+        let hlen = u32::from_le_bytes(envelope[p..p + 4].try_into().unwrap()) as usize;
+        p += 4;
+        let header: serde_json::Value = serde_json::from_slice(&envelope[p..p + hlen]).unwrap();
+        let body = &envelope[p + hlen..];
+        let body_nonce = B64.decode(header["nonce"].as_str().unwrap()).unwrap();
+
+        // X25519 against u = 0 is the all-zero shared secret for any scalar, so
+        // the sealer's box is reconstructable with a key the attacker picks.
+        let zero_box = ChaChaBox::new(
+            &XPublic::from(<[u8; 32]>::default()),
+            &XSecret::generate(&mut OsRng),
+        );
+        for r in header["recipients"].as_array().unwrap() {
+            let n = B64.decode(r["nonce"].as_str().unwrap()).unwrap();
+            let w = B64.decode(r["wrap"].as_str().unwrap()).unwrap();
+            if let Ok(content_key) = zero_box.decrypt(n.as_slice().into(), w.as_slice()) {
+                let cipher = XChaCha20Poly1305::new_from_slice(&content_key).unwrap();
+                let recovered = cipher
+                    .decrypt(XNonce::from_slice(&body_nonce), body)
+                    .expect("body decrypts once the content key is out");
+                panic!(
+                    "REGRESSION: attacker recovered plaintext with no private key: {:?}",
+                    String::from_utf8_lossy(&recovered)
+                );
+            }
+        }
+        panic!("an envelope was sealed to a small-order recipient; the seal guard regressed");
+    }
 }
@@ -1,6 +1,6 @@
 //! HTTP Signatures (RFC 9421) for gitlawb.
 //!
-//! Every write request to a gitlawb node must be signed by the actor's
+//! Every write request to a Twigpine node must be signed by the actor's
 //! Ed25519 private key. The signature covers:
 //!   - `@method`         — HTTP method (uppercase)
 //!   - `@path`           — request path and query
@@ -14,13 +14,13 @@
 use base64::{engine::general_purpose::STANDARD, Engine};
 use chrono::Utc;
 use sha2::{Digest, Sha256};
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};

 use crate::did::Did;
 use crate::identity::Keypair;
 use crate::{Error, Result};

-/// The component identifiers covered by every gitlawb signature.
+/// The component identifiers covered by every Twigpine signature.
 pub const COVERED_COMPONENTS: &[&str] = &["@method", "@path", "content-digest"];

 /// The three headers produced by RFC 9421 signing.
@@ -67,11 +67,20 @@ impl HttpSignature {
        let components_str = &rest[open + 1..close];
        let params_str = &rest[close + 1..]; // starts with ';'

-        // "\"@method\" \"@path\" \"content-digest\"" → ["@method", "@path", "content-digest"]
-        let components: Vec<String> = components_str
-            .split_whitespace()
-            .map(|s| s.trim_matches('"').to_string())
-            .collect();
+        // RFC 9421 §2.1: a component identifier appears at most once. Gitlawb
+        // only supports bare names; parameters are rejected so reordered
+        // equivalent forms cannot bypass duplicate detection as distinct strings.
+        let mut components = Vec::new();
+        let mut seen = HashSet::new();
+        for token in components_str.split_whitespace() {
+            let name = parse_covered_component(token)?;
+            if !seen.insert(name.clone()) {
+                return Err(Error::HttpSignature(format!(
+                    "duplicate covered component '{name}' in Signature-Input"
+                )));
+            }
+            components.push(name);
+        }

        let params = parse_params(params_str)?;

@@ -113,15 +122,48 @@ impl HttpSignature {
        })
    }

-    /// Reject if the `created` timestamp is more than 5 minutes from now.
+    /// Reject a `created` timestamp outside the acceptance window, which is
+    /// ASYMMETRIC: up to 5 minutes into the past, but only 60 seconds into the
+    /// future.
+    ///
+    /// The two directions are two comparisons rather than one distance, because
+    /// they are not the same event. A signature that arrives late is ordinary:
+    /// clocks run slow, requests queue, and networks are not instant, so the
+    /// past side keeps the full 300-second allowance. A signature stamped ahead
+    /// of this node's clock is not ordinary, and folding both sides into an
+    /// `abs()` let a signer extend its own validity: signed at T with
+    /// `created = T + 300`, it stayed acceptable until roughly T + 600, so the
+    /// effective window was double the one this function advertises. The
+    /// forward tolerance exists only to absorb honest drift between two hosts,
+    /// which is what 60 seconds buys; it is the same allowance the gossip
+    /// path's freshness check uses for the same reason.
    pub fn check_created(&self) -> Result<()> {
+        const MAX_AGE_SECS: i64 = 300;
+        const MAX_FUTURE_SECS: i64 = 60;
+
        let now = Utc::now().timestamp();
-        let skew = (now - self.created).abs();
-        if skew > 300 {
+
+        // Saturating, not plain subtraction. `created` is parsed as an
+        // unrestricted i64 straight from the Signature-Input header, so a
+        // sender picks it: at i64::MIN the past-side subtraction overflows,
+        // which panics a debug build and wraps a release one into an absurd
+        // value that can read as acceptable. Saturating gives the answer the
+        // check wants at both extremes, since a timestamp that far away is
+        // refused by whichever side it saturates toward.
+        let age = now.saturating_sub(self.created);
+        if age > MAX_AGE_SECS {
            return Err(Error::HttpSignature(format!(
-                "clock skew too large: {skew}s (max 300s)"
+                "clock skew too large: created {age}s in the past (max {MAX_AGE_SECS}s)"
            )));
        }
+
+        let ahead = self.created.saturating_sub(now);
+        if ahead > MAX_FUTURE_SECS {
+            return Err(Error::HttpSignature(format!(
+                "clock skew too large: created {ahead}s in the future (max {MAX_FUTURE_SECS}s)"
+            )));
+        }
+
        Ok(())
    }

@@ -203,6 +245,33 @@ pub fn compute_content_digest(body: &[u8]) -> String {
    format!("sha-256=:{}:", STANDARD.encode(hasher.finalize()))
 }

+/// Parse one covered-component token from Signature-Input's parenthesized list.
+///
+/// Gitlawb only supports bare component names (`"@method"`, `"@path"`,
+/// `"content-digest"`). RFC 9421 permits parameters on identifiers; they are
+/// rejected rather than accepted in a form `missing_components()` and
+/// `build_signing_string()` cannot handle consistently.
+fn parse_covered_component(token: &str) -> Result<String> {
+    let token = token.trim();
+    let open = token
+        .strip_prefix('"')
+        .ok_or_else(|| Error::HttpSignature("component must be a quoted string".into()))?;
+    let close = open
+        .find('"')
+        .ok_or_else(|| Error::HttpSignature("unterminated component string".into()))?;
+    let name = &open[..close];
+    if name.is_empty() {
+        return Err(Error::HttpSignature("empty component name".into()));
+    }
+    let trailing = open[close + 1..].trim();
+    if !trailing.is_empty() {
+        return Err(Error::HttpSignature(format!(
+            "component parameters are not supported on '{name}'"
+        )));
+    }
+    Ok(name.to_string())
+}
+
 /// Parse `;key="value";key2=value` parameter string into a map.
 fn parse_params(s: &str) -> Result<HashMap<String, String>> {
    let mut map = HashMap::new();
@@ -272,6 +341,91 @@ mod tests {
        assert!(missing.contains(&"content-digest"));
    }

+    /// RFC 9421 §2.1: a component identifier must not appear twice in the
+    /// covered list. Repeating one is not a way to say anything, but it does
+    /// repeat a line in the signing string, so the size of what a verifier
+    /// builds (and what a node persists alongside a claim) is set by how many
+    /// times the caller chose to write the same name.
+    #[test]
+    fn parse_rejects_component_parameters() {
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let sig_input = format!(
+            r#"sig1=("@method" "@path" "content-digest";sf);keyid="{did}";alg="ed25519";created=1000"#
+        );
+        let err = HttpSignature::parse(
+            &sig_input,
+            "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:",
+        )
+        .expect_err("parameterized components must be refused");
+        assert!(
+            err.to_string().contains("component parameters"),
+            "got: {err}"
+        );
+    }
+
+    #[test]
+    fn parse_rejects_reordered_equivalent_component_parameters() {
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let sig_input = format!(
+            r#"sig1=("content-digest";sf;tr "content-digest";tr;sf "@method" "@path");keyid="{did}";alg="ed25519";created=1000"#
+        );
+        let err = HttpSignature::parse(
+            &sig_input,
+            "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:",
+        )
+        .expect_err("reordered equivalent parameters must not bypass as distinct components");
+        assert!(
+            err.to_string().contains("component parameters"),
+            "got: {err}"
+        );
+    }
+
+    #[test]
+    fn parse_many_distinct_components_completes_quickly() {
+        use std::time::{Duration, Instant};
+
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let names: Vec<String> = (0..20_000).map(|i| format!("\"c{i}\"")).collect();
+        let inner = names.join(" ");
+        let sig_input = format!(r#"sig1=({inner});keyid="{did}";alg="ed25519";created=1000"#);
+        let start = Instant::now();
+        let result = HttpSignature::parse(
+            &sig_input,
+            "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:",
+        );
+        let elapsed = start.elapsed();
+        assert!(
+            elapsed < Duration::from_secs(2),
+            "duplicate tracking must stay linear, took {elapsed:?}"
+        );
+        assert!(result.is_ok(), "distinct names should parse: {:?}", result);
+    }
+
+    #[test]
+    fn parse_rejects_duplicate_components() {
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let sig_input = format!(
+            r#"sig1=("@method" "@path" "@path" "content-digest");keyid="{did}";alg="ed25519";created=1000"#
+        );
+        let err = HttpSignature::parse(&sig_input, "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:")
+            .expect_err("a repeated component must be refused");
+        assert!(
+            err.to_string().contains("duplicate"),
+            "the error must name the duplication, got: {err}"
+        );
+
+        // The control: the same list without the repeat still parses.
+        let ok = format!(
+            r#"sig1=("@method" "@path" "content-digest");keyid="{did}";alg="ed25519";created=1000"#
+        );
+        HttpSignature::parse(&ok, "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:")
+            .expect("a distinct component list must still parse");
+    }
+
    #[test]
    fn verify_signature_end_to_end() {
        use crate::identity::verify;
@@ -335,6 +489,75 @@ mod tests {
        assert!(sig.check_created().is_err());
    }

+    /// Build a parseable signature whose `created` is `offset` seconds from now
+    /// (negative for the past), so a test can pin one side of the window.
+    fn sig_created_at_offset(offset: i64) -> HttpSignature {
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let created = Utc::now().timestamp() + offset;
+        let sig_input = format!(
+            r#"sig1=("@method" "@path" "content-digest");keyid="{did}";alg="ed25519";created={created}"#
+        );
+        HttpSignature::parse(&sig_input, "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:").unwrap()
+    }
+
+    /// Build a signature at an ABSOLUTE `created`, so a test can drive the
+    /// extremes an offset from now cannot reach.
+    fn sig_created_absolute(created: i64) -> HttpSignature {
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let sig_input = format!(
+            r#"sig1=("@method" "@path" "content-digest");keyid="{did}";alg="ed25519";created={created}"#
+        );
+        HttpSignature::parse(&sig_input, "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:").unwrap()
+    }
+
+    /// `created` is parsed as an unrestricted i64 from a header the sender
+    /// writes, so the window check has to survive both ends of the type. Before
+    /// the saturating form this panicked with "attempt to subtract with
+    /// overflow" on the past side, which is a debug-build crash any client
+    /// could trigger by sending one header value.
+    #[test]
+    fn extreme_created_values_are_refused_rather_than_overflowing() {
+        for created in [i64::MIN, i64::MIN + 1, i64::MAX - 1, i64::MAX] {
+            let err = sig_created_absolute(created)
+                .check_created()
+                .expect_err("an extreme created must be refused, not accepted");
+            let msg = err.to_string();
+            assert!(
+                msg.contains("past") || msg.contains("future"),
+                "created={created} must be refused by a named direction, got: {msg}"
+            );
+        }
+    }
+
+    #[test]
+    fn created_301s_in_past_rejected() {
+        let err = sig_created_at_offset(-301).check_created().unwrap_err();
+        assert!(err.to_string().contains("past"), "{err}");
+    }
+
+    #[test]
+    fn created_299s_in_past_accepted() {
+        assert!(sig_created_at_offset(-299).check_created().is_ok());
+    }
+
+    #[test]
+    fn created_61s_in_future_rejected() {
+        let err = sig_created_at_offset(61).check_created().unwrap_err();
+        assert!(err.to_string().contains("future"), "{err}");
+    }
+
+    #[test]
+    fn created_30s_in_future_accepted() {
+        assert!(sig_created_at_offset(30).check_created().is_ok());
+    }
+
+    #[test]
+    fn created_exactly_now_accepted() {
+        assert!(sig_created_at_offset(0).check_created().is_ok());
+    }
+
    #[test]
    fn fresh_signature_passes_clock_skew() {
        let kp = Keypair::generate();
@@ -343,6 +566,26 @@ mod tests {
        assert!(sig.check_created().is_ok());
    }

+    #[test]
+    fn extreme_created_timestamps_reject_without_wrapping() {
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let parse = |created: i64| {
+            let sig_input =
+                format!(r#"sig1=("@method");keyid="{did}";alg="ed25519";created={created}"#);
+            HttpSignature::parse(&sig_input, "sig1=:AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA:").unwrap()
+        };
+        // i64::MIN used to panic in debug (subtraction overflow) and wrap to a
+        // passing skew in release; now - 2^63 wrapped to i64::MIN, whose abs()
+        // wrapped back to itself and read as fresh. Both must reject.
+        for created in [i64::MIN, i64::MAX, Utc::now().timestamp() - i64::MAX] {
+            assert!(
+                parse(created).check_created().is_err(),
+                "created={created} must not pass the freshness gate"
+            );
+        }
+    }
+
    #[test]
    fn parse_error_missing_sig1_prefix() {
        let err = HttpSignature::parse(
@@ -1,6 +1,6 @@
 //! Keypair generation, storage, and signing.
 //!
-//! A gitlawb identity is an Ed25519 keypair. The public key encodes into a
+//! A Twigpine identity is an Ed25519 keypair. The public key encodes into a
 //! `did:key` DID. The private key is stored as PKCS#8 PEM on disk.

 use ed25519_dalek::{Signature, Signer, SigningKey, VerifyingKey};
@@ -11,7 +11,7 @@ use zeroize::Zeroizing;
 use crate::did::Did;
 use crate::{Error, Result};

-/// An Ed25519 keypair that is the root identity for a gitlawb actor.
+/// An Ed25519 keypair that is the root identity for a Twigpine actor.
 #[derive(Clone)]
 pub struct Keypair {
    signing_key: SigningKey,
@@ -77,6 +77,14 @@ impl Keypair {
 }

 /// Verify an Ed25519 signature.
+///
+/// Strict verification: rejects small-order `R` and small-order public keys
+/// (the identity point, and any point of low order). Ordinary `verify` accepts
+/// a signature forged with the identity point as the public key plus
+/// `R = identity, S = 0`, which verifies for *any* message. `identity::verify`
+/// is the shared primitive behind HTTP request authentication, UCANs, and
+/// certificates, so weak-key acceptance is an authentication bypass, not a
+/// malleability nuance.
 pub fn verify(verifying_key: &VerifyingKey, msg: &[u8], sig_bytes: &[u8; 64]) -> Result<()> {
    let sig = Signature::from_bytes(sig_bytes);
    verifying_key
@@ -208,6 +216,38 @@ mod tests {
        );
    }

+    /// The identity-point forgery: with public key A = identity, R = identity,
+    /// and S = 0, the equation `[S]B = R + [k]A` holds for every message,
+    /// because `[k]·identity = identity`. Ordinary (non-strict) Ed25519
+    /// verification accepts it, so the shared `verify` primitive must use
+    /// strict verification, which rejects small-order R and public keys.
+    #[test]
+    fn verify_rejects_identity_point_forgery() {
+        use ed25519_dalek::Verifier;
+        // The identity point (0,1) compresses to y = 1 with sign bit 0.
+        let identity = [
+            1u8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+            0, 0, 0, 0,
+        ];
+        let vk = VerifyingKey::from_bytes(&identity).expect("identity point is on the curve");
+        let mut sig_bytes = [0u8; 64];
+        sig_bytes[..32].copy_from_slice(&identity);
+        let msg = b"arbitrary message the key owner never signed";
+
+        // Prove the forged signature satisfies the ordinary verification
+        // equation, so the strict check below is what actually defends the
+        // boundary (not a signature that was already invalid everywhere).
+        assert!(
+            vk.verify(msg, &Signature::from_bytes(&sig_bytes)).is_ok(),
+            "identity-point forgery must satisfy ordinary verification (this is why strict is needed)"
+        );
+
+        assert!(
+            verify(&vk, msg, &sig_bytes).is_err(),
+            "strict verification must reject the identity-point forgery"
+        );
+    }
+
    #[test]
    fn verify_rejects_weak_key_signature() {
        // Regression guard for strict verification: a signature forged under a
@@ -5,7 +5,16 @@ pub mod encrypt;
 pub mod error;
 pub mod http_sig;
 pub mod identity;
+// `url` is the one dependency here that drags a tail (idna, then the icu
+// crates), and gitlawb-core is allowlisted to stay embeddable. Every client that
+// needs this predicate already parses URLs, so they opt in and nothing else
+// pays. `test` is in the cfg so `cargo test -p gitlawb-core` still compiles and
+// runs the matrix below with no feature selected; without it the tests would
+// silently not run, which is the failure this module exists to prevent.
+#[cfg(any(feature = "redirect", test))]
+pub mod redirect;
 pub mod sanitize;
+pub mod scan_token;
 pub mod ucan;

 pub use error::Error;
@@ -0,0 +1,241 @@
+//! Redirect policy shared by every Twigpine HTTP client that signs its requests.
+//!
+//! Both `gl` (async) and `git-remote-gitlawb` (blocking) attach RFC 9421
+//! `Signature` and `Signature-Input` headers, and reqwest strips only
+//! `Authorization`, `Cookie`, `Proxy-Authorization` and `WWW-Authenticate` when a
+//! redirect crosses hosts. A signature would survive that hop, and it binds
+//! `@method`, `@path` and `content-digest` with no authority component, so a node
+//! answering 302 could hand a working credential to a host of its choosing and read
+//! as the caller anywhere until the clock-skew window closes. On a 307/308 the
+//! request body goes along with it, which for the remote helper is the pack.
+//!
+//! That same missing authority component is why the predicate also pins the
+//! request-target: `@path` is signed as the client sent it and verified as the node
+//! received it, so a same-origin hop that rewrites the path or the query (a
+//! trailing-slash or query normalization) makes the signature cover a target the
+//! node never saw and the read 401s. Only a hop that re-issues the identical target,
+//! an http-to-https upgrade being the one that matters in practice, is followed.
+//!
+//! The decision lives here rather than in either client because the two used to
+//! disagree: `gl` was scoped to the origin while the remote helper, the binary that
+//! actually runs `git clone gitlawb://`, still ran reqwest's default and followed
+//! anywhere. One predicate is what keeps a future third client from repeating that.
+//!
+//! The type is `url::Url`, which is what `reqwest::Url` re-exports, so both clients
+//! pass their attempt URLs straight in.
+
+/// Longest redirect chain followed. `reqwest::redirect::Policy::custom` replaces
+/// reqwest's built-in limit, so the bound has to be restated by every client that
+/// installs a custom policy; the value is reqwest's own default.
+///
+/// This counts FOLLOWS, matching `Policy::limited`: reqwest pushes the redirecting
+/// URL onto `previous` before consulting the policy, and `Limit(max)` refuses once
+/// `previous.len() > max`, so a caller comparing against this constant must use `>`
+/// too or it permits one hop fewer than it says.
+pub const MAX_REDIRECTS: usize = 10;
+
+/// Follow a redirect only when it stays on the origin that issued it AND re-issues
+/// the identical request-target.
+///
+/// `Policy::none()` would have been the simpler answer, but one same-origin redirect
+/// shape is legitimate here: a node fronted by a proxy that upgrades http to https,
+/// or otherwise re-issues the same path and query, so the policy is scoped rather
+/// than switched off.
+///
+/// Path and query must match exactly, and that is the request-target clause rather
+/// than an origin one. `@path` is signed as the client sent it and verified as the
+/// node received it, so a hop that rewrites either half leaves a signature covering
+/// a target nobody asked for and the node answers 401. Refusing the hop turns a
+/// confusing 401 into the 3xx that names what actually happened. One policy covers
+/// signed and unsigned callers alike, for the same reason the predicate is shared:
+/// two rules would drift.
+///
+/// The clause pins `@path`, and only `@path`. A Twigpine signature also covers
+/// `@method` and `content-digest`, and both of those are still open on a followed
+/// hop: on a 301, 302 or 303, reqwest 0.12.28 delegates to tower-http's
+/// `FollowRedirect`, which rewrites a POST to a GET and empties the body
+/// (tower-http-0.6.8 `src/follow_redirect/mod.rs:273-285`), while its
+/// `drop_payload_headers` removes only `Content-Type`, `Content-Length`,
+/// `Content-Encoding` and `Transfer-Encoding`. So `Signature`, `Signature-Input` and
+/// `Content-Digest` ride along on a request that no longer has the method or the body
+/// they were computed over. Only 307 and 308 preserve both. This predicate returning
+/// true therefore makes a GET-shaped hop safe to replay against the node's verifier
+/// and says nothing about a bodied one: a signed write must not rely on it alone.
+///
+/// `Url::query` is `None` for `/a` and `Some("")` for `/a?`, and those are two
+/// different request-targets on the node side too, so the comparison is strict and
+/// needs no special case.
+///
+/// Host and port must match exactly. Port is compared as `Url::port`, which is
+/// `None` for a scheme's default port, so http -> https on the same host compares
+/// equal while http -> http on a different port does not. A downgrade from https to
+/// http is refused as well: the target is the same host, but the signature would go
+/// out in cleartext, which is the same credential leak by a slower route.
+///
+/// Host comparison rides on `url`'s parse-time normalization (lowercasing and IDN
+/// -> punycode), so the spellings an attacker reaches for do not open a gap. That
+/// is a property of the parsed `Url`, not of this function, which is why the test
+/// matrix pins it: a move to raw string comparison would silently lose it.
+pub fn may_follow(previous: &url::Url, next: &url::Url) -> bool {
+    let same_origin = next.host_str() == previous.host_str() && next.port() == previous.port();
+    let same_target = next.path() == previous.path() && next.query() == previous.query();
+    let downgraded = previous.scheme() == "https" && next.scheme() != "https";
+    same_origin && same_target && !downgraded
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Every branch of the decision, both ways, plus the spellings that would slip
+    /// past a comparison less careful than `Url`'s own normalization.
+    #[test]
+    fn may_follow_covers_each_origin_branch() {
+        let url = |s: &str| url::Url::parse(s).unwrap();
+        let cases: &[(&str, &str, bool, &str)] = &[
+            (
+                "http://node.example/a",
+                "http://node.example/b",
+                false,
+                "same origin but the path changed: @path is signed as sent and verified as received",
+            ),
+            (
+                "http://node.example/a",
+                "https://node.example/a",
+                true,
+                "http to https on one host with an identical request-target: both ports are \
+                 the scheme default, so the proxy upgrade is still followed",
+            ),
+            (
+                "https://node.example/a",
+                "https://node.example/a/",
+                false,
+                "trailing-slash normalization changes the request-target, so it is refused",
+            ),
+            (
+                "https://node.example:8443/a",
+                "https://node.example:8443/a",
+                true,
+                "same explicit port",
+            ),
+            (
+                "http://node.example/a",
+                "http://attacker.example/a",
+                false,
+                "different host",
+            ),
+            (
+                "http://node.example/a",
+                "http://node.example:8080/a",
+                false,
+                "same host, different port",
+            ),
+            (
+                "https://node.example/a",
+                "http://node.example/a",
+                false,
+                "https downgraded to cleartext on the same host",
+            ),
+            (
+                "https://node.example/a",
+                "http://node.example:443/a",
+                false,
+                "a downgrade dressed up as the https port",
+            ),
+            // The rows below pass today because `Url::parse` normalizes the host, not
+            // because anything here compares case-insensitively or decodes IDN. They
+            // are the variants an attacker reaches for, so they are pinned: swapping
+            // this predicate for a raw string comparison must break the suite.
+            // Each of these pairs an IDENTICAL path on both sides, deliberately. The
+            // request-target clause below would make every one of them false on the
+            // path alone, and a row that is false for two reasons has stopped pinning
+            // either. With the paths equal, the host or port comparison is the only
+            // thing left that can decide them.
+            (
+                "https://node.example/a",
+                "https://NODE.EXAMPLE/a",
+                true,
+                "same host in a different case: parse lowercases it",
+            ),
+            (
+                "https://node.example/a",
+                "https://node.example./a",
+                false,
+                "a trailing dot is a different host to url, so the redirect is refused",
+            ),
+            (
+                "https://exämple.test/a",
+                "https://xn--exmple-cua.test/a",
+                true,
+                "unicode host and its punycode spelling are one host after parse",
+            ),
+            (
+                "https://node.example/a",
+                "https://user:[email protected]/a",
+                true,
+                "userinfo is not part of the origin: same host, still followed",
+            ),
+            (
+                "https://node.example/a",
+                "https://[email protected]/a",
+                false,
+                "the node's name smuggled into userinfo: the host is the attacker's",
+            ),
+            (
+                "https://node.example/a",
+                "https://attacker.example/a#node.example",
+                false,
+                "the node's name pushed into the fragment: the host is the attacker's",
+            ),
+            // The request-target clause, both directions. `@path` is the only thing a
+            // Twigpine signature binds the request to, so a hop that rewrites it hands
+            // the node a signature over a target it never received.
+            (
+                "https://node.example/a?x=1",
+                "https://node.example/a?x=1",
+                true,
+                "identical request-target: the same-origin hop that is still followed",
+            ),
+            (
+                "https://node.example/a?x=1",
+                "https://node.example/a?x=2",
+                false,
+                "same path but a different query: the request-target covers the query too",
+            ),
+            (
+                "https://node.example/a",
+                "https://node.example/a?x=1",
+                false,
+                "a query added where there was none",
+            ),
+            (
+                "http://node.example/a",
+                "http://node.example/a?",
+                false,
+                "an empty query added where there was none: a missing query and an empty \
+                 one are different request-targets",
+            ),
+            (
+                "https://node.example/a?x=1",
+                "https://node.example/a",
+                false,
+                "the query dropped where there was one: the comparison is symmetric, and \
+                 nothing else in the matrix pins that direction",
+            ),
+            (
+                "https://node.example/a#x",
+                "https://node.example/a#y",
+                true,
+                "a fragment-only difference is still followed: a fragment never reaches \
+                 the wire, so it is no part of the request-target the node verifies",
+            ),
+        ];
+        for (previous, next, expected, why) in cases {
+            assert_eq!(
+                may_follow(&url(previous), &url(next)),
+                *expected,
+                "{previous} -> {next} ({why})"
+            );
+        }
+    }
+}
@@ -0,0 +1,499 @@
+//! Sealed continuation token for the node's bounded legacy CID scan (INV-13).
+//!
+//! The `/ipfs/{cid}` resolver's legacy scan stops at a row ceiling and sheds a
+//! retryable 503. To let a holder buried past that ceiling still be reached, the
+//! shed carries the scan position so the caller can echo it back and resume. The
+//! whole point of the design is that the node keeps NO server-side scan state: the
+//! position rides in the caller's token.
+//!
+//! That makes the token an EMITTED continuation derived from a FETCHED row, and on
+//! a scan that served nothing every fetched row is by construction a private or
+//! quarantined repo the caller may not read. The row's `created_at` leaks its
+//! creation time and its `id` carries the owner's DID, so both halves are withheld
+//! fields and the token must be CONFIDENTIAL, not merely tamper-evident:
+//!
+//!   * AEAD-sealed (XChaCha20-Poly1305), never base64-of-plaintext and never
+//!     signed plaintext. Integrity is not confidentiality.
+//!   * A fresh `OsRng` nonce on EVERY seal. Under a stream cipher a repeated nonce
+//!     means repeated keystream, and an attacker who can force the node to seal a
+//!     position whose plaintext they know XORs two tokens and recovers a withheld
+//!     row's fields in full, strictly worse than emitting plaintext.
+//!   * FIXED-WIDTH plaintext. AEAD ciphertext is plaintext-length plus the tag, and
+//!     every field of a scan position varies in length, so a variable encoding would
+//!     make token LENGTH a side channel for the sealed row (a short name under a
+//!     short owner vs a long one) and for the candidate oid's width (40 hex on a
+//!     sha1 repo, 64 on a sha256 one). Every token this module mints is byte-identical
+//!     in length. Each field is padded to its own separate width, which is a per-field
+//!     constant and so still leaks nothing about a given row or candidate.
+//!   * The canonical CID as associated data, so a token minted while scanning for
+//!     one CID does not authenticate when replayed against another.
+//!
+//! Every failure to open (wrong key, tampered bytes, wrong CID, expired, malformed)
+//! returns the same `None`. The caller treats that as "no token" and starts at the
+//! front, so no failure class is distinguishable and the token is no oracle.
+
+use base64::{engine::general_purpose::URL_SAFE_NO_PAD as B64URL, Engine};
+use chacha20poly1305::{
+    aead::{Aead, KeyInit, OsRng, Payload},
+    XChaCha20Poly1305, XNonce,
+};
+use rand::RngCore;
+
+/// Keyset position of the last row a truncated scan fetched.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ScanPosition {
+    /// The row's raw stored `created_at` text, the first half of the keyset cursor.
+    pub created_at_key: String,
+    /// The row's `id`, the tiebreaking half of the keyset cursor.
+    pub id: String,
+    /// The oid hex of the CANDIDATE this position resumes. One CID can map to several
+    /// git oids, so a bare row cursor names a row without naming whose walk it belongs
+    /// to. The candidate is named by IDENTITY rather than by position in the candidate
+    /// list: that list is ordered by hex and mutates between rungs, so an index would
+    /// name a different candidate the moment anything is pinned or unpinned.
+    pub sha256_hex: String,
+}
+
+/// Plaintext version byte, so a future layout change is a clean open-failure
+/// (treated as absent) rather than a misparse.
+///
+/// Bumped to 2 when the two halves stopped sharing one width (see [`ID_WIDTH`]), and
+/// to 3 when the position gained the candidate oid it resumes (see [`OID_WIDTH`]).
+/// A token minted under an earlier layout is a different length and a different
+/// framing, so it opens to `None` and the caller restarts at the front, which is the
+/// safe direction: a misparse would resume at a fabricated row and skip coverage.
+const VERSION: u8 = 3;
+
+/// Byte width the `created_at` half is padded to. Every value stored here is a
+/// serialized timestamp, about 30 bytes, so 64 is roomy for the field's whole domain.
+/// It is deliberately NOT widened to match [`ID_WIDTH`]: padding both halves to the id
+/// width would nearly double every token for a field that can never use the space.
+const CREATED_WIDTH: usize = 64;
+
+/// Byte width the `id` half is padded to.
+///
+/// The bound is set by the WRITERS, not by what a typical id happens to look like.
+/// `upsert_mirror_repo` builds `repos.id` as `{owner}/{name}`, and the slug validators
+/// in the node's `repo_store` admit an owner of up to 255 bytes and a name of up to
+/// 100, so 356 bytes is reachable through the ordinary write path and repo names are
+/// peer-controllable. 384 clears that with margin.
+///
+/// Under-sizing this is not a cosmetic bug. A row at a truncation boundary whose id
+/// exceeds the width fails the seal, the handler sheds a 503 with no continuation, and
+/// a tokenless shed is byte-identical to the wrapped-scan response whose contract is
+/// "the absence of a token means the ladder is over". The boundary row is deterministic
+/// for a stable inventory, so every retry reproduces it and every row past it becomes
+/// permanently unreachable. Anything past the width still fails loudly rather than
+/// silently truncating a cursor into one that resumes at the wrong row.
+const ID_WIDTH: usize = 384;
+
+/// Byte width the candidate oid half is padded to.
+///
+/// Git mints exactly two oid widths and this field carries BOTH. A production repo is
+/// created by `store::init_bare` with `git init --bare --object-format=sha1`, so its
+/// oids are 40 hex; only the sha256 test fixtures mint 64. The field is therefore
+/// length-prefixed like the two row halves rather than a bare fixed 64: a 64-only
+/// field would fail every seal on a real deployment, and a failed seal sheds a
+/// tokenless 503 that is byte-identical to "your ladder is over".
+///
+/// The padding to 64 is what keeps the WIDTH off the wire. Without it a 40-hex token
+/// is 24 bytes shorter than a 64-hex one, and token length would say which object
+/// format the holder's repo uses.
+const OID_WIDTH: usize = 64;
+
+/// `version | created_len:u16 | created[CREATED_WIDTH] | id_len:u16 | id[ID_WIDTH] | oid_len:u16 | oid[OID_WIDTH] | expires:i64`
+const PLAINTEXT_LEN: usize = 1 + 2 + CREATED_WIDTH + 2 + ID_WIDTH + 2 + OID_WIDTH + 8;
+
+/// Nonce width for XChaCha20-Poly1305.
+const NONCE_LEN: usize = 24;
+
+/// A fresh random 32-byte sealing key from the OS CSPRNG.
+pub fn new_key() -> [u8; 32] {
+    let mut key = [0u8; 32];
+    OsRng.fill_bytes(&mut key);
+    key
+}
+
+// The two halves below are deliberately separate and adjacent. The framing pair owns
+// "every token is the same length"; the AEAD pair owns "the contents are confidential
+// and CID-bound". Keeping them apart is what lets each property be exercised (and
+// broken) without disturbing the other.
+
+/// Encode a position into the FIXED-WIDTH plaintext:
+/// `version | created_len:u16 | created[CREATED_WIDTH] | id_len:u16 | id[ID_WIDTH] | oid_len:u16 | oid[OID_WIDTH] | expires:i64`
+///
+/// The padding is the point. AEAD ciphertext is plaintext-length plus the tag, and every
+/// field of a scan position varies in length, so a length-prefixed encoding with no
+/// padding would make token LENGTH a side channel for the sealed row and for the
+/// candidate's object format. Each field is padded to its OWN fixed width, which keeps
+/// every minted token the same length while letting the id half carry the range the
+/// write path actually admits.
+fn encode_position(pos: &ScanPosition, expires_at_unix: i64) -> anyhow::Result<Vec<u8>> {
+    let mut out = vec![0u8; PLAINTEXT_LEN];
+    out[0] = VERSION;
+    let mut at = 1;
+    for (field, width) in [
+        (pos.created_at_key.as_bytes(), CREATED_WIDTH),
+        (pos.id.as_bytes(), ID_WIDTH),
+        (pos.sha256_hex.as_bytes(), OID_WIDTH),
+    ] {
+        if field.len() > width {
+            // Loud rather than truncating: a clipped cursor resumes at the wrong row and
+            // silently skips coverage, which is the availability half of the bug this
+            // token exists to fix.
+            anyhow::bail!(
+                "scan token field is {} bytes, over the {width}-byte fixed width",
+                field.len()
+            );
+        }
+        out[at..at + 2].copy_from_slice(&(field.len() as u16).to_le_bytes());
+        at += 2;
+        out[at..at + field.len()].copy_from_slice(field);
+        at += width;
+    }
+    out[at..at + 8].copy_from_slice(&expires_at_unix.to_le_bytes());
+    Ok(out)
+}
+
+/// Decode what [`encode_position`] wrote. `None` on any structural mismatch.
+fn decode_position(bytes: &[u8]) -> Option<(ScanPosition, i64)> {
+    if bytes.len() != PLAINTEXT_LEN || bytes[0] != VERSION {
+        return None;
+    }
+    let mut at = 1;
+    let mut fields = [const { String::new() }; 3];
+    for (slot, width) in fields.iter_mut().zip([CREATED_WIDTH, ID_WIDTH, OID_WIDTH]) {
+        let len = u16::from_le_bytes([bytes[at], bytes[at + 1]]) as usize;
+        at += 2;
+        if len > width {
+            return None;
+        }
+        *slot = String::from_utf8(bytes[at..at + len].to_vec()).ok()?;
+        at += width;
+    }
+    let expires_at = i64::from_le_bytes(bytes[at..at + 8].try_into().ok()?);
+    let [created_at_key, id, sha256_hex] = fields;
+    // A zero-length candidate is a third state the encoder never mints. The front-of-table
+    // sentinel is empty ROW halves with a real oid, so an empty oid would hand the resume
+    // path a candidate that names nothing; refuse it like every other malformed frame.
+    if sha256_hex.is_empty() {
+        return None;
+    }
+    Some((
+        ScanPosition {
+            created_at_key,
+            id,
+            sha256_hex,
+        },
+        expires_at,
+    ))
+}
+
+/// AEAD-seal `plaintext` under `key`, bound to `cid`, framed as `nonce || ciphertext`.
+fn seal_bytes(key: &[u8; 32], cid: &str, plaintext: &[u8]) -> anyhow::Result<Vec<u8>> {
+    let cipher = XChaCha20Poly1305::new_from_slice(key)
+        .map_err(|e| anyhow::anyhow!("scan token key: {e}"))?;
+    // A FRESH nonce per seal, from the OS CSPRNG. Under a stream cipher a repeated nonce
+    // repeats the keystream, and two tokens sealed under one nonce XOR to the difference
+    // of their plaintexts, which recovers a withheld row in full when the attacker can
+    // force one of the two positions. This draw is the property the whole confidentiality
+    // claim rests on.
+    let mut nonce = [0u8; NONCE_LEN];
+    OsRng.fill_bytes(&mut nonce);
+    let sealed = cipher
+        .encrypt(
+            XNonce::from_slice(&nonce),
+            Payload {
+                msg: plaintext,
+                // The canonical CID as associated data: a token minted while scanning for
+                // one CID does not authenticate against another, so it cannot be replayed
+                // to seed a different scan.
+                aad: cid.as_bytes(),
+            },
+        )
+        .map_err(|e| anyhow::anyhow!("scan token seal: {e}"))?;
+    let mut out = Vec::with_capacity(NONCE_LEN + sealed.len());
+    out.extend_from_slice(&nonce);
+    out.extend_from_slice(&sealed);
+    Ok(out)
+}
+
+/// Open what [`seal_bytes`] framed. `None` on any failure, including a wrong `cid`.
+fn open_bytes(key: &[u8; 32], cid: &str, raw: &[u8]) -> Option<Vec<u8>> {
+    if raw.len() <= NONCE_LEN + 16 {
+        return None;
+    }
+    let (nonce, sealed) = raw.split_at(NONCE_LEN);
+    let cipher = XChaCha20Poly1305::new_from_slice(key).ok()?;
+    cipher
+        .decrypt(
+            XNonce::from_slice(nonce),
+            Payload {
+                msg: sealed,
+                aad: cid.as_bytes(),
+            },
+        )
+        .ok()
+}
+
+/// Seal `pos` under `key`, bound to `cid`, expiring at `expires_at_unix`.
+///
+/// Returns the base64url (no pad) token. Errors only when a field exceeds
+/// its own fixed width ([`CREATED_WIDTH`], [`ID_WIDTH`], [`OID_WIDTH`]) or the AEAD
+/// itself fails, never silently truncates.
+pub fn seal_scan_token(
+    key: &[u8; 32],
+    cid: &str,
+    pos: &ScanPosition,
+    expires_at_unix: i64,
+) -> anyhow::Result<String> {
+    let plaintext = encode_position(pos, expires_at_unix)?;
+    Ok(B64URL.encode(seal_bytes(key, cid, &plaintext)?))
+}
+
+/// Open a token minted by [`seal_scan_token`] under the same key and CID.
+///
+/// `None` for every failure class alike (wrong key, tampered, foreign CID, expired,
+/// malformed, wrong version), so the caller can treat all of them as "absent" without
+/// leaking which one occurred.
+pub fn open_scan_token(
+    key: &[u8; 32],
+    cid: &str,
+    token: &str,
+    now_unix: i64,
+) -> Option<ScanPosition> {
+    let raw = B64URL.decode(token).ok()?;
+    let plaintext = open_bytes(key, cid, &raw)?;
+    let (pos, expires_at) = decode_position(&plaintext)?;
+    if now_unix >= expires_at {
+        return None;
+    }
+    Some(pos)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// A production-shaped candidate: `git init --bare --object-format=sha1`, so 40 hex.
+    const OID_40: &str = "a1b2c3d4e5f60718293a4b5c6d7e8f9012345678";
+    /// A test-fixture-shaped candidate: the sha256 repos the suite creates mint 64 hex.
+    const OID_64: &str = "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef";
+
+    fn pos(created: &str, id: &str) -> ScanPosition {
+        pos_for(created, id, OID_40)
+    }
+
+    fn pos_for(created: &str, id: &str, sha256_hex: &str) -> ScanPosition {
+        ScanPosition {
+            created_at_key: created.to_string(),
+            id: id.to_string(),
+            sha256_hex: sha256_hex.to_string(),
+        }
+    }
+
+    /// Seal a hand-built plaintext through the AEAD half, so a test can frame bytes the
+    /// encoder would never mint (an old version byte, a zero-length oid) and still exercise
+    /// the real open path.
+    fn seal_raw(key: &[u8; 32], cid: &str, plaintext: &[u8]) -> String {
+        B64URL.encode(seal_bytes(key, cid, plaintext).unwrap())
+    }
+
+    /// Byte offset of the oid length prefix inside the plaintext, derived from the widths
+    /// rather than hardcoded so a width change moves it with the layout.
+    const OID_LEN_AT: usize = 1 + 2 + CREATED_WIDTH + 2 + ID_WIDTH;
+
+    /// Scenario 1: both of git's oid widths round trip. The 40-hex case is the PRODUCTION
+    /// shape (`git init --bare --object-format=sha1`), and an all-64 fixture suite would
+    /// never exercise it, which is exactly how a fixed-64 field would ship broken.
+    #[test]
+    fn round_trips_at_both_oid_widths() {
+        let key = new_key();
+        for hex in [OID_40, OID_64] {
+            let p = pos_for("2020-01-01T00:00:03+00:00", "z6MkOwner/private-repo", hex);
+            let t =
+                seal_scan_token(&key, "bafkcid", &p, 1 << 40).expect("both oid widths must seal");
+            assert_eq!(
+                open_scan_token(&key, "bafkcid", &t, 0),
+                Some(p),
+                "a {}-hex candidate must open to the identical position",
+                hex.len()
+            );
+        }
+    }
+
+    /// Scenario 2: a token framed under the OLD version opens to `None`, never a misparse.
+    /// Both legs matter: the version-2 plaintext was a different LENGTH, and a future
+    /// same-length layout would only be caught by the version byte itself.
+    #[test]
+    fn a_prior_layout_version_opens_to_none() {
+        let key = new_key();
+
+        // The version-2 layout verbatim: no oid field, so 461 bytes.
+        let mut old = vec![0u8; 1 + 2 + CREATED_WIDTH + 2 + ID_WIDTH + 8];
+        old[0] = 2;
+        assert_eq!(old.len(), 461, "the version-2 plaintext was 461 bytes");
+        assert_eq!(
+            open_scan_token(&key, "bafkcid", &seal_raw(&key, "bafkcid", &old), 0),
+            None,
+            "a version-2 token must open to None so the caller restarts at the front"
+        );
+
+        // Same length, stale version byte: only the version check can refuse this one.
+        let mut stamped = encode_position(
+            &pos("2020-01-01T00:00:03+00:00", "z6MkOwner/private-repo"),
+            1 << 40,
+        )
+        .unwrap();
+        stamped[0] = VERSION - 1;
+        assert_eq!(
+            open_scan_token(&key, "bafkcid", &seal_raw(&key, "bafkcid", &stamped), 0),
+            None,
+            "a stale version byte must open to None even at the current width"
+        );
+    }
+
+    /// Scenario 3: token length is invariant across candidate VALUE and candidate WIDTH.
+    /// The padding is what keeps the oid width off the wire: without it a 40-hex token is
+    /// 24 bytes shorter than a 64-hex one and the length says which repo format the
+    /// holder uses.
+    #[test]
+    fn token_length_is_invariant_across_oid_widths() {
+        let key = new_key();
+        let created = "2020-01-01T00:00:00+00:00";
+        let id = "z6MkOwner/private-repo";
+        let short =
+            seal_scan_token(&key, "bafkcid", &pos_for(created, id, OID_40), 1 << 40).unwrap();
+        let long =
+            seal_scan_token(&key, "bafkcid", &pos_for(created, id, OID_64), 1 << 40).unwrap();
+        assert_eq!(
+            short.len(),
+            long.len(),
+            "a 40-hex and a 64-hex candidate must mint tokens of identical length, or the \
+             oid width is a side channel"
+        );
+
+        // The absolute width, pinned by execution rather than by arithmetic on paper. The
+        // gl client's mock fixtures hardcode this number (`TOKEN_LEN` in
+        // crates/gl/src/ipfs_cmd.rs), and nothing in that crate seals a real token, so this
+        // assertion is the only executable check that the two agree.
+        assert_eq!(
+            short.len(),
+            756,
+            "24 nonce + {PLAINTEXT_LEN} plaintext + 16 tag, base64url no pad"
+        );
+    }
+
+    /// Scenario 4: the front-of-table sentinel. Empty row halves with a REAL candidate
+    /// round trip, which is what lets a seal say "this candidate, no row cursor yet".
+    #[test]
+    fn empty_row_fields_round_trip_with_a_real_candidate() {
+        let key = new_key();
+        let p = pos_for("", "", OID_40);
+        let t =
+            seal_scan_token(&key, "bafkcid", &p, 1 << 40).expect("the front sentinel must seal");
+        assert_eq!(open_scan_token(&key, "bafkcid", &t, 0), Some(p));
+    }
+
+    /// Scenario 6, third leg: a zero-length oid is a distinguishable third state that the
+    /// encoder never mints, and accepting it would hand the sentinel machinery a candidate
+    /// naming nothing. The decode path refuses it.
+    #[test]
+    fn a_zero_length_candidate_opens_to_none() {
+        let key = new_key();
+        let mut plaintext = encode_position(
+            &pos("2020-01-01T00:00:03+00:00", "z6MkOwner/private-repo"),
+            1 << 40,
+        )
+        .unwrap();
+        plaintext[OID_LEN_AT..OID_LEN_AT + 2].copy_from_slice(&0u16.to_le_bytes());
+        assert_eq!(
+            open_scan_token(&key, "bafkcid", &seal_raw(&key, "bafkcid", &plaintext), 0),
+            None,
+            "a zero-length candidate must open to None, not to a position naming nothing"
+        );
+    }
+
+    /// Scenario 6, second leg: an oid past the 64-byte width fails the seal loudly rather
+    /// than being clipped into a hex that names a different candidate.
+    #[test]
+    fn an_oid_over_the_fixed_width_fails_loudly() {
+        let key = new_key();
+        let p = pos_for(
+            "2020-01-01T00:00:03+00:00",
+            "z6MkOwner/private-repo",
+            &"a".repeat(OID_WIDTH + 1),
+        );
+        assert!(
+            seal_scan_token(&key, "bafkcid", &p, 1 << 40).is_err(),
+            "an over-wide candidate must fail the seal, never be truncated into a hex that \
+             resumes the wrong candidate"
+        );
+    }
+
+    #[test]
+    fn round_trips_under_the_same_key_and_cid() {
+        let key = new_key();
+        let p = pos("2020-01-01T00:00:03+00:00", "z6MkOwner/private-repo");
+        let t = seal_scan_token(&key, "bafkcid", &p, 1 << 40).unwrap();
+        assert_eq!(open_scan_token(&key, "bafkcid", &t, 0), Some(p));
+    }
+
+    #[test]
+    fn every_failure_class_opens_to_none() {
+        let key = new_key();
+        let other = new_key();
+        let p = pos("2020-01-01T00:00:03+00:00", "z6MkOwner/private-repo");
+        let t = seal_scan_token(&key, "bafkcid", &p, 1 << 40).unwrap();
+
+        assert_eq!(open_scan_token(&other, "bafkcid", &t, 0), None, "wrong key");
+        assert_eq!(
+            open_scan_token(&key, "bafkOTHER", &t, 0),
+            None,
+            "foreign CID"
+        );
+        assert_eq!(
+            open_scan_token(&key, "bafkcid", &t, 1 << 41),
+            None,
+            "expired"
+        );
+        assert_eq!(open_scan_token(&key, "bafkcid", "!!not b64", 0), None);
+        assert_eq!(open_scan_token(&key, "bafkcid", "", 0), None);
+        let mut flipped: Vec<u8> = t.bytes().collect();
+        let last = flipped.len() - 1;
+        flipped[last] = if flipped[last] == b'A' { b'B' } else { b'A' };
+        assert_eq!(
+            open_scan_token(&key, "bafkcid", &String::from_utf8(flipped).unwrap(), 0),
+            None,
+            "tampered"
+        );
+    }
+
+    /// The id half must clear the LARGEST repo id the node's own write path admits,
+    /// not merely a typical one. `upsert_mirror_repo` builds `repos.id` as
+    /// `{owner}/{name}`, and the slug validators in `repo_store` admit 255 bytes of
+    /// owner and 100 of name, so 356 is reachable. A width under that turns the
+    /// boundary row into a seal failure, which sheds a tokenless 503 that is
+    /// byte-identical to "your ladder is over" and strands every row past it forever.
+    #[test]
+    fn round_trips_a_repo_id_at_the_write_paths_maximum() {
+        let key = new_key();
+        let id = format!("{}/{}", "o".repeat(255), "n".repeat(100));
+        assert_eq!(id.len(), 356, "255 owner + '/' + 100 name");
+        let p = pos("2020-01-01T00:00:03+00:00", &id);
+        let t = seal_scan_token(&key, "bafkcid", &p, 1 << 40)
+            .expect("a repo id the write path admits must seal, never fail the width");
+        assert_eq!(open_scan_token(&key, "bafkcid", &t, 0), Some(p));
+    }
+
+    #[test]
+    fn a_field_over_the_fixed_width_fails_loudly() {
+        let key = new_key();
+        let p = pos("2020-01-01T00:00:03+00:00", &"x".repeat(ID_WIDTH + 1));
+        assert!(
+            seal_scan_token(&key, "bafkcid", &p, 1 << 40).is_err(),
+            "an over-wide field must fail the seal, never be truncated into a cursor \
+             that resumes at the wrong row"
+        );
+    }
+}
@@ -1,7 +1,7 @@
 //! UCAN (User Controlled Authorization Networks) — capability token types.
 //!
 //! UCANs let a DID delegate specific capabilities to another DID,
-//! with optional expiry and revocation. gitlawb uses UCANs for:
+//! with optional expiry and revocation. Twigpine uses UCANs for:
 //!   - Delegating push access to a branch to a CI agent
 //!   - Granting a reviewer the ability to approve PRs
 //!   - Bootstrap tokens issued at registration
@@ -57,7 +57,7 @@ impl Capability {
    }
 }

-/// Well-known gitlawb capability strings.
+/// Well-known Twigpine capability strings.
 pub mod caps {
    pub const GIT_PUSH: &str = "git/push";
    pub const GIT_FETCH: &str = "git/fetch";
@@ -1,6 +1,6 @@
 [package]
 name = "gitlawb-node"
-description = "The gitlawb node daemon — git hosting over HTTP with DID auth"
+description = "The Twigpine node daemon — git hosting over HTTP with DID auth"
 version = "0.7.1" # x-release-please-version
 edition.workspace = true
 rust-version.workspace = true
@@ -23,6 +23,8 @@ tracing = { workspace = true }
 tracing-subscriber = { workspace = true }
 chrono = { workspace = true }
 uuid = { workspace = true }
+url = { workspace = true }
+percent-encoding = { workspace = true }
 axum = { version = "0.8", features = ["http1", "http2", "json", "ws"] }
 async-graphql = { version = "7", features = ["chrono", "uuid", "tracing"] }
 async-graphql-axum = "7"
@@ -33,6 +35,7 @@ sqlx = { version = "0.8", features = ["postgres", "runtime-tokio-rustls", "chron
 clap = { version = "4", features = ["derive", "env"] }
 bytes = "1"
 libc = "0.2"
+zeroize = "1"
 cid = { workspace = true }
 hex = { workspace = true }
 sha2 = { workspace = true }
@@ -44,14 +47,14 @@ dirs-next = "2"
 # withholds a canonically-equivalent path regardless of NFC/NFD form (#101).
 unicode-normalization = "0.1"
 reqwest = { workspace = true }
-libp2p-core = "0.43"
-libp2p-gossipsub = "0.49.4"
-libp2p-identify = "0.47"
-libp2p-identity = { version = "0.2", features = ["ed25519", "peerid", "rand"] }
-libp2p-kad = "0.48"
-libp2p-quic = { version = "0.13", features = ["tokio"] }
-libp2p-swarm = { version = "0.47", features = ["macros", "tokio"] }
-multiaddr = "0.18"
+libp2p-core = "0.44"
+libp2p-gossipsub = "0.50.1"
+libp2p-identify = "0.48"
+libp2p-identity = { version = "0.3", features = ["ed25519", "peerid", "rand"] }
+libp2p-kad = "0.49"
+libp2p-quic = { version = "0.14", features = ["tokio"] }
+libp2p-swarm = { version = "0.48", features = ["macros", "tokio"] }
+multiaddr = "0.19"
 futures = "0.3"
 aws-sdk-s3 = { version = "1", default-features = false, features = ["sigv4a", "default-https-client", "rt-tokio"] }
 aws-config = { version = "1", features = ["behavior-version-latest"] }
@@ -72,7 +75,8 @@ alloy = { version = "1", default-features = false, features = [
    "network",
    "rpc-types-eth",
 ] }
-libp2p-dns = { version = "0.44.0", features = ["tokio"] }
+rand = { workspace = true }
+libp2p-dns = { version = "0.45.0", features = ["tokio"] }

 [dev-dependencies]
 mockito = "1"
@@ -1,12 +1,13 @@
 //! GET /api/v1/arweave/anchors — list Arweave ref-update anchors.

 use axum::{
-    extract::{Query, State},
+    extract::{Extension, Query, State},
    Json,
 };
 use serde::Deserialize;

-use crate::error::Result;
+use crate::db::normalize_owner_key;
+use crate::error::{AppError, Result};
 use crate::state::AppState;

 #[derive(Debug, Deserialize)]
@@ -21,17 +22,62 @@ fn default_limit() -> i64 {
 }

 /// GET /api/v1/arweave/anchors
+///
+/// `?repo=<owner>/<name>` returns anchors for one repository and binds the
+/// request to the canonical read-authorization path used by every other
+/// repo-scoped read: the same `authorize_repo_read` helper, with the caller
+/// identity attached. Missing repositories, quarantined mirrors, and signed
+/// non-readers all collapse to the standard `404` (no existence oracle). The
+/// unscoped listing is auth-only — the #121 contract — and does not admit
+/// visibility filtering, which is the #136 stale-index class and explicitly
+/// out of scope for this auth slice.
 pub async fn list_anchors(
    State(state): State<AppState>,
    Query(q): Query<ListAnchorsQuery>,
+    auth: Option<Extension<crate::auth::AuthenticatedDid>>,
 ) -> Result<Json<serde_json::Value>> {
-    let limit = q.limit.min(200);
-    // Bare `?` so connection-class sqlx failures downcast to `AppError::Db` and
-    // map to 503 `db_unavailable` (not 500 via `.map_err(AppError::Internal)`) (#251).
-    let anchors = state
-        .db
-        .list_arweave_anchors(q.repo.as_deref(), limit)
-        .await?;
+    // The route's `optional_signature` layer is permissive (it admits unsigned
+    // legacy callers) so this in-handler check is the actual gate. A signed
+    // non-reader still passes this check and proceeds to the authz step below.
+    let caller = auth.as_ref().map(|e| e.0 .0.as_str());
+    if caller.is_none() {
+        return Err(AppError::Unauthorized(
+            "authentication required for anchor listing".into(),
+        ));
+    }
+
+    // Clamp before the SQL query. A negative or non-numeric limit must not
+    // reach Postgres as `LIMIT -1` (which it rejects as a db error / 500) — the
+    // contract is the same default ceiling as the listing page itself.
+    let limit = q.limit.clamp(0, 200);
+
+    let anchors = if let Some(repo_slug) = q.repo.as_deref() {
+        // Parse the user-supplied "owner/name" through the same slug validator
+        // the sync path uses, so a malformed query is a 400, not a 500.
+        let (owner, name) = match crate::git::repo_store::validate_repo_slug(repo_slug) {
+            Ok(parts) => parts,
+            Err(e) => return Err(AppError::BadRequest(format!("invalid ?repo: {e}"))),
+        };
+        // Read gate. Returns `RepoNotFound` (→ 404) indistinguishably for
+        // missing repos, quarantined mirrors, and signed non-readers.
+        let (record, _rules) =
+            crate::api::authorize_repo_read(&state, owner, name, caller, "/").await?;
+        // Build the SQL filter from the canonical stored slug, NOT from the
+        // user-supplied `?repo=` string. Anchor rows are written as
+        // `{normalize_owner_key(owner_did)}/{name}` (the short form), so a
+        // request that passes authz with the full `did:key:…/name` form would
+        // match zero rows and return a false empty page. The gate already
+        // verified the repo is readable; `record.owner_did` is the canonical
+        // identity and `record.name` is the stored name.
+        let stored_slug = format!("{}/{}", normalize_owner_key(&record.owner_did), record.name);
+        state
+            .db
+            .list_arweave_anchors(Some(&stored_slug), limit)
+            .await?
+    } else {
+        // No scope → no repo-read decision. Auth-only.
+        state.db.list_arweave_anchors(None, limit).await?
+    };

    Ok(Json(serde_json::json!({
        "anchors": anchors,
@@ -60,6 +106,7 @@ mod closed_pool_tests {
            .oneshot(
                Request::builder()
                    .uri("/api/v1/arweave/anchors")
+                    .extension(crate::auth::AuthenticatedDid("did:key:test".into()))
                    .body(axum::body::Body::empty())
                    .unwrap(),
            )
@@ -285,11 +285,22 @@ pub async fn submit_bounty(
    Path(id): Path<String>,
    Json(req): Json<SubmitBountyRequest>,
 ) -> Result<Json<BountyRecord>> {
-    let bounty = state
-        .db
-        .get_bounty(&id)
-        .await?
-        .ok_or_else(|| AppError::NotFound(format!("bounty {id} not found")))?;
+    let not_found = || AppError::NotFound(format!("bounty {id} not found"));
+
+    let bounty = state.db.get_bounty(&id).await?.ok_or_else(not_found)?;
+
+    if crate::api::authorize_repo_read(
+        &state,
+        &bounty.repo_owner,
+        &bounty.repo_name,
+        Some(auth.0.as_str()),
+        "/",
+    )
+    .await
+    .is_err()
+    {
+        return Err(not_found());
+    }

    if bounty.status != "claimed" {
        return Err(AppError::BadRequest(format!(
@@ -325,11 +336,22 @@ pub async fn approve_bounty(
    Path(id): Path<String>,
    Json(req): Json<ApproveBountyRequest>,
 ) -> Result<Json<BountyRecord>> {
-    let bounty = state
-        .db
-        .get_bounty(&id)
-        .await?
-        .ok_or_else(|| AppError::NotFound(format!("bounty {id} not found")))?;
+    let not_found = || AppError::NotFound(format!("bounty {id} not found"));
+
+    let bounty = state.db.get_bounty(&id).await?.ok_or_else(not_found)?;
+
+    if crate::api::authorize_repo_read(
+        &state,
+        &bounty.repo_owner,
+        &bounty.repo_name,
+        Some(auth.0.as_str()),
+        "/",
+    )
+    .await
+    .is_err()
+    {
+        return Err(not_found());
+    }

    if bounty.status != "submitted" {
        return Err(AppError::BadRequest(format!(
@@ -372,11 +394,22 @@ pub async fn cancel_bounty(
    Extension(auth): Extension<AuthenticatedDid>,
    Path(id): Path<String>,
 ) -> Result<Json<BountyRecord>> {
-    let bounty = state
-        .db
-        .get_bounty(&id)
-        .await?
-        .ok_or_else(|| AppError::NotFound(format!("bounty {id} not found")))?;
+    let not_found = || AppError::NotFound(format!("bounty {id} not found"));
+
+    let bounty = state.db.get_bounty(&id).await?.ok_or_else(not_found)?;
+
+    if crate::api::authorize_repo_read(
+        &state,
+        &bounty.repo_owner,
+        &bounty.repo_name,
+        Some(auth.0.as_str()),
+        "/",
+    )
+    .await
+    .is_err()
+    {
+        return Err(not_found());
+    }

    if bounty.status != "open" {
        return Err(AppError::BadRequest(format!(
@@ -407,11 +440,22 @@ pub async fn dispute_bounty(
    Extension(auth): Extension<AuthenticatedDid>,
    Path(id): Path<String>,
 ) -> Result<Json<BountyRecord>> {
-    let bounty = state
-        .db
-        .get_bounty(&id)
-        .await?
-        .ok_or_else(|| AppError::NotFound(format!("bounty {id} not found")))?;
+    let not_found = || AppError::NotFound(format!("bounty {id} not found"));
+
+    let bounty = state.db.get_bounty(&id).await?.ok_or_else(not_found)?;
+
+    if crate::api::authorize_repo_read(
+        &state,
+        &bounty.repo_owner,
+        &bounty.repo_name,
+        Some(auth.0.as_str()),
+        "/",
+    )
+    .await
+    .is_err()
+    {
+        return Err(not_found());
+    }

    if bounty.status != "claimed" && bounty.status != "submitted" {
        return Err(AppError::BadRequest(format!(
@@ -456,21 +500,68 @@ pub async fn dispute_bounty(
    Ok(Json(updated))
 }

+/// Resolve the set of `(owner_did, name)` pairs for repos an anonymous caller
+/// can list.  Mirrors the `stats()` pattern in `server.rs` (#104): batch-load
+/// all deduped repos, batch-load their visibility rules, keep only those that
+/// pass `listable_at_root`.  Pure I/O after the two DB round-trips — no
+/// per-repo authorization queries.
+///
+/// Both bounty stats handlers mount behind `optional_signature` (see
+/// `bounty_read_routes` in server.rs), so the caller is always `None` for
+/// anonymous access.  Every caller, signed or not, currently receives the
+/// anonymous-scoped aggregates (matching the #104 pattern); per-caller
+/// visibility can be threaded through here when that work lands.
+///
+/// A database error is returned to the caller, never collapsed to an empty
+/// set: an empty set would render as HTTP 200 with zero counts, which a client
+/// reads as confirmed absence of activity.
+async fn visible_repo_pairs(state: &AppState) -> Result<Vec<(String, String)>> {
+    let rows = state.db.list_all_repos_deduped().await?;
+    let ids: Vec<String> = rows.iter().map(|r| r.id.clone()).collect();
+    let rules_by_repo = state.db.list_visibility_rules_for_repos(&ids).await?;
+    let pairs = rows
+        .iter()
+        .filter(|r| {
+            let rules = rules_by_repo.get(&r.id).map(Vec::as_slice).unwrap_or(&[]);
+            // mirror-rows-handled: a mirror with a canonical twin is deduped to the
+            // canonical row and its rules; a mirror-only repo counts as public, the
+            // same answer authorize_repo_read and /stats give. Pinned by
+            // bounty_stats_mirror_rows_follow_canonical_twin_or_count_when_alone.
+            crate::visibility::listable_at_root(rules, r.is_public, &r.owner_did, None)
+        })
+        .map(|r| {
+            (
+                crate::db::normalize_owner_key(&r.owner_did).to_string(),
+                r.name.clone(),
+            )
+        })
+        .collect();
+    Ok(pairs)
+}
+
 /// GET /api/v1/bounties/stats
+///
+/// Aggregates are restricted to anonymously-listable repos so private-repo
+/// bounty activity is not exposed (#477).  The visible-repo set is resolved
+/// once per request via `visible_repo_pairs` (two SQL round-trips), then
+/// passed into the filtered aggregate queries.
 pub async fn bounty_stats(State(state): State<AppState>) -> Result<Json<BountyStatsResponse>> {
-    let open = state.db.count_bounties_by_status("open").await.unwrap_or(0);
+    let visible = visible_repo_pairs(&state).await?;
+
+    let open = state
+        .db
+        .count_bounties_by_status_visible("open", &visible)
+        .await?;
    let claimed = state
        .db
-        .count_bounties_by_status("claimed")
-        .await
-        .unwrap_or(0);
+        .count_bounties_by_status_visible("claimed", &visible)
+        .await?;
    let completed = state
        .db
-        .count_bounties_by_status("completed")
-        .await
-        .unwrap_or(0);
+        .count_bounties_by_status_visible("completed", &visible)
+        .await?;

-    let leaders = state.db.bounty_leaderboard(10).await.unwrap_or_default();
+    let leaders = state.db.bounty_leaderboard_visible(10, &visible).await?;
    let leaderboard = leaders
        .into_iter()
        .map(|(did, cnt, total)| AgentBountyEntry {
@@ -489,11 +580,14 @@ pub async fn bounty_stats(State(state): State<AppState>) -> Result<Json<BountySt
 }

 /// GET /api/v1/agents/{did}/bounties
+///
+/// Per-agent earnings restricted to anonymously-listable repos (#477).
 pub async fn agent_bounty_stats(
    State(state): State<AppState>,
    Path(did): Path<String>,
 ) -> Result<Json<serde_json::Value>> {
-    let (count, total) = state.db.agent_bounty_stats(&did).await.unwrap_or((0, 0));
+    let visible = visible_repo_pairs(&state).await?;
+    let (count, total) = state.db.agent_bounty_stats_visible(&did, &visible).await?;
    Ok(Json(serde_json::json!({
        "did": did,
        "completed_bounties": count,
@@ -52,6 +52,7 @@ pub async fn list_certs(
                "node_did":   c.node_did,
                "signature":  c.signature,
                "issued_at":  c.issued_at,
+                "version":    c.version,
            })
        })
        .collect();
@@ -92,5 +93,6 @@ pub async fn get_cert(
        "node_did":   cert.node_did,
        "signature":  cert.signature,
        "issued_at":  cert.issued_at,
+        "version":    cert.version,
    })))
 }
@@ -11,7 +11,8 @@ use crate::state::AppState;

 /// Hard ceiling on rows any ref-update feed returns in one request. Shared by the
 /// shared collector's clamp and the per-handler request caps so they can't drift.
-const MAX_VISIBLE_REF_UPDATES: i64 = 200;
+/// `pub(crate)` so the GraphQL complexity meter can price against the same bound.
+pub(crate) const MAX_VISIBLE_REF_UPDATES: i64 = 200;

 /// Collect up to `limit` ref-update rows visible to `caller`, newest first,
 /// paging past rows the feed gate drops. Filtering after a plain SQL `LIMIT`
@@ -437,6 +438,7 @@ mod ref_updates_feed_tests {
            node_did: "did:key:z6MkNode".into(),
            signature: "sig".into(),
            issued_at: Utc::now().to_rfc3339(),
+            version: 1,
        }
    }

@@ -14,14 +14,13 @@ pub struct LabelRequest {
    pub label: String,
 }

-/// POST /api/v1/repos/:owner/:repo/labels
-pub async fn add_label(
-    State(state): State<AppState>,
-    Extension(auth): Extension<AuthenticatedDid>,
-    Path((owner, name)): Path<(String, String)>,
-    Json(req): Json<LabelRequest>,
-) -> Result<(StatusCode, Json<serde_json::Value>)> {
-    let label = req.label.trim().to_lowercase();
+/// Canonicalize and validate a label string. `add_label` and `remove_label`
+/// must agree on the form stored in the DB: `add_label` already trims and
+/// lowercases before storing, so `remove_label` has to do the same to find
+/// the row. Validation is shared so neither handler can drift on charset or
+/// length (#344).
+fn canonicalize_label(raw: &str) -> std::result::Result<String, AppError> {
+    let label = raw.trim().to_lowercase();
    if label.is_empty() || label.len() > 50 {
        return Err(AppError::BadRequest("label must be 1–50 characters".into()));
    }
@@ -33,6 +32,17 @@ pub async fn add_label(
            "label must contain only alphanumeric characters, hyphens, and colons".into(),
        ));
    }
+    Ok(label)
+}
+
+/// POST /api/v1/repos/:owner/:repo/labels
+pub async fn add_label(
+    State(state): State<AppState>,
+    Extension(auth): Extension<AuthenticatedDid>,
+    Path((owner, name)): Path<(String, String)>,
+    Json(req): Json<LabelRequest>,
+) -> Result<(StatusCode, Json<serde_json::Value>)> {
+    let label = canonicalize_label(&req.label)?;

    let record = state
        .db
@@ -59,6 +69,8 @@ pub async fn remove_label(
    Extension(auth): Extension<AuthenticatedDid>,
    Path((owner, name, label)): Path<(String, String, String)>,
 ) -> Result<Json<serde_json::Value>> {
+    let label = canonicalize_label(&label)?;
+
    let record = state
        .db
        .get_repo(&owner, &name)
@@ -66,7 +78,10 @@ pub async fn remove_label(
        .ok_or_else(|| AppError::RepoNotFound(format!("{owner}/{name}")))?;
    crate::api::require_repo_owner(&record, &auth.0)?;

-    state.db.remove_label(&record.id, &label).await?;
+    let removed = state.db.remove_label(&record.id, &label).await?;
+    if !removed {
+        return Err(AppError::NotFound(format!("label '{label}'")));
+    }
    Ok(Json(serde_json::json!({ "label": label, "removed": true })))
 }

@@ -22,6 +22,7 @@ pub mod replicas;
 pub mod repos;
 pub mod resolve;
 pub mod stars;
+pub mod task_cursor;
 pub mod tasks;
 pub mod visibility;
 pub mod webhooks;
@@ -80,15 +81,39 @@ pub(crate) fn did_matches(a: &str, b: &str) -> bool {
    !ka.contains(':') && !kb.contains(':') && ka == kb
 }

+/// Client-facing denial shared by every repo-owner gate copy. `require_repo_owner`
+/// renders it as `AppError::Forbidden`; a handler that emits its own
+/// `(StatusCode, Json)` envelope instead of `AppError` (e.g.
+/// `api::tasks::create_task`) must use this constant too, so two copies of one
+/// gate cannot drift in observable output.
+pub(crate) const REPO_OWNER_ONLY_MESSAGE: &str = "only the repo owner can perform this action";
+
 /// 403 unless `caller` is the repo owner. Uses [`did_matches`] so the owner check
 /// and the author check (close policy) share one normalization.
 pub(crate) fn require_repo_owner(record: &RepoRecord, caller: &str) -> Result<()> {
    if did_matches(caller, &record.owner_did) {
        Ok(())
    } else {
-        Err(AppError::Forbidden(
-            "only the repo owner can perform this action".into(),
-        ))
+        Err(AppError::Forbidden(REPO_OWNER_ONLY_MESSAGE.into()))
+    }
+}
+
+/// Validate a branch name using the configured git binary (storage boundaries).
+pub(crate) fn validate_git_ref_with_git(
+    git_bin: &str,
+    name: &str,
+) -> std::result::Result<(), crate::git::store::GitRefValidationError> {
+    crate::git::store::validate_git_ref_with_git(git_bin, name)
+}
+
+/// Map ref-validation failures to the correct HTTP surface: malformed caller
+/// input stays 400; an inability to run git stays on the git/server-error path.
+pub(crate) fn map_git_ref_validation_error(
+    err: crate::git::store::GitRefValidationError,
+) -> AppError {
+    match err {
+        crate::git::store::GitRefValidationError::Invalid(msg) => AppError::BadRequest(msg),
+        crate::git::store::GitRefValidationError::GitUnavailable(msg) => AppError::Git(msg),
    }
 }

@@ -208,17 +233,26 @@ mod authz_guard {
            (issues, "create_issue", "authorize_repo_read("),
            (bounties, "create_bounty", "authorize_repo_read("),
            (repos, "fork_repo", "authorize_repo_read("),
-            // get_by_cid gates each iterated repo row directly via visibility_check
-            // (KTD2a: it must NOT route through authorize_repo_read's fuzzy re-resolve).
-            (ipfs, "get_by_cid", "visibility_check("),
-            // #94 sibling read surfaces: gate private-repo metadata on read
-            // visibility (public repos stay anonymous; private repos 404).
+            // get_by_cid resolves each candidate (provenance path + legacy scan) through
+            // the shared `gate_and_serve`; the gate markers themselves are asserted
+            // below. This row proves the delegation is real, so the gate is actually
+            // reached rather than dead code. The delegated gate still calls
+            // `visibility_check` directly and never `authorize_repo_read`, so it keeps
+            // the property the pre-merge marker enforced: no fuzzy re-resolve.
+            (ipfs, "get_by_cid", "gate_and_serve("),
+            // Sibling read surfaces: gate private-repo metadata on read visibility
+            // (public repos stay anonymous; private repos 404).
            (replicas, "list_replicas", "authorize_repo_read("),
            (protect, "list_protected_branches", "authorize_repo_read("),
            (labels, "list_labels", "authorize_repo_read("),
            (events, "list_repo_events", "authorize_repo_read("),
            // Bucket C — signer-self: the acting DID is matched/bound to auth.0
            (tasks, "create_task", "did_matches("),
+            // #496: create_task is signer-self AND owner-when-repo-scoped — a
+            // repo_id naming a hosted repo admits tasks only from its owner.
+            // Both halves are pinned: the did_matches row guards the signer
+            // binding, this one the repo-ownership gate.
+            (tasks, "create_task", "require_repo_owner("),
            (tasks, "claim_task", "did_matches("),
            (tasks, "complete_task", "did_matches("),
            (tasks, "fail_task", "did_matches("),
@@ -250,6 +284,22 @@ mod authz_guard {
            "visibility::require_owner must use did_matches for DID-safe owner matching"
        );

+        // The CID read surface (#173) enforces its gate inside the shared
+        // `gate_and_serve`, which BOTH the provenance path and the legacy scan call, so
+        // the markers must live there (the get_by_cid row above only proves delegation).
+        // The repo's own "/" visibility check (KTD2a — never authorize_repo_read's fuzzy
+        // re-resolve) and the quarantine hard-drop BEFORE visibility (INV-11) are both
+        // load-bearing: removing either re-opens a leak on the provenance path.
+        let gate_body = fn_body(ipfs, "gate_and_serve");
+        assert!(
+            gate_body.contains("visibility_check("),
+            "gate_and_serve must gate the CID read surface via visibility_check (KTD2a)"
+        );
+        assert!(
+            gate_body.contains("if quarantined"),
+            "gate_and_serve must hard-drop a quarantined repo before the visibility gate (INV-11)"
+        );
+
        for (src, func, marker) in rows {
            let body = fn_body(src, func);
            assert!(
@@ -259,6 +309,577 @@ mod authz_guard {
        }
    }

+    /// Completeness fence over the GraphQL mutation surface: the mirror of
+    /// `every_in_scope_mutation_has_its_gate` for `graphql/mutation.rs`. Kept a
+    /// SEPARATE guard with its own table (never folded into the REST rows above):
+    /// a guard over a heterogeneous set must classify per member, and these
+    /// same-named task rows belong to the GraphQL surface, not `api/tasks.rs`.
+    ///
+    /// COMPLETENESS comes from the engine's own SDL (`schema.sdl()`), parsed as a
+    /// structured AST (#219), NOT introspection and NOT a text parse of source.
+    /// This matters: async-graphql OMITS `#[graphql(visible = false)]` fields from
+    /// introspection while leaving them executable, so an introspection-sourced fence
+    /// could be defeated by a hidden mutation. SDL includes those fields, so no
+    /// visibility flag (and, via the AST parse, no brace-in-description quirk) can hide
+    /// a live mutation. The SDL field set must equal the registered set: a new field with
+    /// no row fails (a repo-write cannot slip in unlisted); a registered name absent from
+    /// the schema fails (renamed or removed). The set is also cross-checked against
+    /// introspection, so any field present in SDL but hidden from introspection fails
+    /// named as a forbidden hidden mutation.
+    ///
+    /// The MARKER check is the only source-scraped part, and it is not the
+    /// completeness guarantee. Three honest limits, so a green check is not misread
+    /// as full authorization coverage:
+    ///   1. A present marker is not load-bearingness (the gate may not redden on a
+    ///      hostile probe: the #203 present-but-vacuous case). The adversarial tests
+    ///      in `graphql/mutation.rs` are the load-bearing layer.
+    ///   2. A present marker is not gate-correctness (it may be the wrong bucket for
+    ///      the operation's resource: a repo-write registered as signer-self would
+    ///      pass, so INV-1 review remains the defense).
+    ///   3. The marker check scrapes the comment-masked, brace-bounded resolver body;
+    ///      a `{`/`}` or a `did_matches(` inside a STRING literal is not masked, so it
+    ///      could mis-slice that body or satisfy the marker vacuously (the same
+    ///      vacuous-marker class as limit 1). Completeness is unaffected: it comes
+    ///      from the SDL AST, not from this scrape.
+    #[test]
+    fn every_graphql_mutation_has_its_gate() {
+        // (Rust snake name, schema field in camelCase, expected gate markers). The
+        // four are Bucket C signer-self: `/graphql` uses optional_signature, so each
+        // resolver must call `require_signer(ctx)?` and bind the acting DID with
+        // `did_matches(`. A repo-write would be Bucket A (`require_repo_owner(`).
+        // Same-named REST rows live in the guard above against `api/tasks.rs`.
+        let registered: &[(&str, &str, &[&str])] = &[
+            (
+                "create_task",
+                "createTask",
+                &[
+                    "require_signer(ctx)?",
+                    "did_matches(",
+                    "require_repo_owner(",
+                ],
+            ),
+            (
+                "claim_task",
+                "claimTask",
+                &["require_signer(ctx)?", "did_matches("],
+            ),
+            (
+                "complete_task",
+                "completeTask",
+                &["require_signer(ctx)?", "did_matches("],
+            ),
+            (
+                "fail_task",
+                "failTask",
+                &["require_signer(ctx)?", "did_matches("],
+            ),
+        ];
+
+        // Completeness + hidden-mutation cross-check against the REAL schema (#219),
+        // delegated to the generic `assert_mutation_fence` so the SDL sourcing and the
+        // cross-check are ALSO exercised by a synthetic hidden schema in
+        // `fence_rejects_a_hidden_mutation` (INV-21). Because the real MutationRoot has no
+        // hidden field today, that synthetic test — not a hypothetical production canary —
+        // is what turns reverting the SDL source or deleting the cross-check RED.
+        {
+            use crate::graphql::mutation::MutationRoot;
+            use crate::graphql::query::QueryRoot;
+            use crate::graphql::subscription::SubscriptionRoot;
+            let schema =
+                async_graphql::Schema::build(QueryRoot, MutationRoot, SubscriptionRoot).finish();
+            let registered_camel: Vec<&str> = registered.iter().map(|(_, c, _)| *c).collect();
+            assert_mutation_fence(&schema, &registered_camel);
+        }
+
+        // Marker present (source scrape; the vacuity limit, backstopped by the
+        // adversarial tests): each resolver body carries every required gate marker.
+        let masked = mask_comments(include_str!("../graphql/mutation.rs"));
+        for (snake, _, markers) in registered {
+            let body = resolver_body(&masked, snake);
+            for marker in *markers {
+                assert!(
+                    body.contains(marker),
+                    "GraphQL mutation `{snake}` is missing its gate marker `{marker}`: gate removed"
+                );
+            }
+        }
+    }
+
+    /// Version-safety self-check (#219, INV-21): the fence's completeness guarantee rests
+    /// on `schema.sdl()` INCLUDING `visible = false` fields while introspection omits
+    /// them. That is observed async-graphql behavior, not a documented contract, so pin
+    /// it here: a synthetic schema with a hidden mutation must have that field in its
+    /// SDL-derived set and NOT in its introspection-derived set. If a future
+    /// async-graphql upgrade ever filters SDL by visibility, the first assertion flips
+    /// RED — the assumption is load-bearing, not trusted.
+    #[test]
+    fn sdl_source_sees_hidden_mutations_that_introspection_omits() {
+        use async_graphql::{EmptySubscription, Object, Schema};
+        struct Q;
+        #[Object]
+        impl Q {
+            async fn ping(&self) -> i32 {
+                0
+            }
+        }
+        struct M;
+        #[Object]
+        impl M {
+            async fn visible_mut(&self) -> i32 {
+                1
+            }
+            #[graphql(visible = false)]
+            async fn hidden_mut(&self) -> i32 {
+                2
+            }
+        }
+        let schema = Schema::build(Q, M, EmptySubscription).finish();
+
+        // SDL source (the fix): includes the hidden field.
+        let sdl_fields = mutation_field_names_from_sdl(&schema.sdl());
+        assert!(
+            sdl_fields.iter().any(|f| f == "hiddenMut"),
+            "SDL-sourced set MUST include the visible=false mutation; if this is RED, \
+             async-graphql started filtering SDL by visibility and the fence's \
+             completeness guarantee is broken: {sdl_fields:?}"
+        );
+
+        // Introspection (the fence's own helper): OMITS the hidden field. This is the
+        // exact gap the fix closes — an introspection-sourced fence never sees `hiddenMut`.
+        let introspected = introspection_mutation_fields(&schema);
+        assert!(
+            !introspected.iter().any(|f| f == "hiddenMut"),
+            "introspection is expected to OMIT the visible=false mutation (that is the \
+             gap the SDL source closes); got {introspected:?}"
+        );
+
+        // The exact diff the fence computes: hidden ∈ (sdl \ introspection).
+        let hidden: Vec<&String> = sdl_fields
+            .iter()
+            .filter(|f| !introspected.iter().any(|i| i == *f))
+            .collect();
+        assert_eq!(
+            hidden,
+            vec!["hiddenMut"],
+            "the SDL-vs-introspection diff must surface exactly the hidden mutation"
+        );
+    }
+
+    /// INV-21 wiring guard (#219): drives the generic `assert_mutation_fence` — the SAME
+    /// helper the real fence calls — against a synthetic schema whose mutation is
+    /// `visible = false` and IS registered, so completeness passes and ONLY the
+    /// hidden-from-introspection cross-check can fail. This is what makes the fence's SDL
+    /// sourcing load-bearing without a production canary: reverting the source to
+    /// introspection (the hidden field then absent from the SDL set) or deleting the
+    /// cross-check removes this panic (or changes its message), turning the
+    /// `should_panic` test RED. The real MutationRoot has no hidden field, so nothing on
+    /// the real schema could catch that regression.
+    #[test]
+    #[should_panic(expected = "hidden from introspection")]
+    fn fence_rejects_a_hidden_mutation() {
+        use async_graphql::{EmptySubscription, Object, Schema};
+        struct Q;
+        #[Object]
+        impl Q {
+            async fn ping(&self) -> i32 {
+                0
+            }
+        }
+        struct M;
+        #[Object]
+        impl M {
+            async fn visible_mut(&self) -> i32 {
+                1
+            }
+            #[graphql(visible = false)]
+            async fn hidden_mut(&self) -> i32 {
+                2
+            }
+        }
+        let schema = Schema::build(Q, M, EmptySubscription).finish();
+        // Both fields registered, so completeness passes and only the cross-check fires.
+        assert_mutation_fence(&schema, &["visibleMut", "hiddenMut"]);
+    }
+
+    /// Deprecated-mutation guard (#219): a `#[graphql(deprecation = "...")]` mutation is
+    /// executable and present in SDL, but GraphQL introspection omits it by default. The
+    /// fence's introspection helper uses `includeDeprecated: true`, so a deprecated (but
+    /// visible) mutation is NOT mistaken for a hidden one. With it registered, the fence
+    /// must NOT panic. Remove `includeDeprecated: true` and this goes RED — the deprecated
+    /// field would look hidden and false-red as a forbidden visible=false mutation.
+    #[test]
+    fn fence_admits_a_deprecated_mutation() {
+        use async_graphql::{EmptySubscription, Object, Schema};
+        struct Q;
+        #[Object]
+        impl Q {
+            async fn ping(&self) -> i32 {
+                0
+            }
+        }
+        struct M;
+        #[Object]
+        impl M {
+            async fn visible_mut(&self) -> i32 {
+                1
+            }
+            #[graphql(deprecation = "use visibleMut")]
+            async fn deprecated_mut(&self) -> i32 {
+                2
+            }
+        }
+        let schema = Schema::build(Q, M, EmptySubscription).finish();
+        // Must not panic: the deprecated field is visible (present in introspection with
+        // includeDeprecated) and registered, so it is neither hidden nor unclassified.
+        assert_mutation_fence(&schema, &["visibleMut", "deprecatedMut"]);
+    }
+
+    /// Completeness + hidden-mutation cross-check for a BUILT schema (#219). Generic over
+    /// the schema type so the real fence AND the synthetic `fence_rejects_a_hidden_mutation`
+    /// / `fence_admits_a_deprecated_mutation` tests drive the SAME sourcing and cross-check.
+    /// Completeness is sourced from SDL (which includes `visible = false` fields); the
+    /// cross-check fails any SDL field hidden from introspection. This genericity is the
+    /// INV-21 fix: because the real MutationRoot has no hidden field, only a synthetic
+    /// hidden schema through THIS helper makes reverting the SDL source or removing the
+    /// cross-check RED, rather than leaving the class latent behind a production canary.
+    fn assert_mutation_fence<Q, M, S>(
+        schema: &async_graphql::Schema<Q, M, S>,
+        registered_camel: &[&str],
+    ) where
+        Q: async_graphql::ObjectType + 'static,
+        M: async_graphql::ObjectType + 'static,
+        S: async_graphql::SubscriptionType + 'static,
+    {
+        let sdl_fields = mutation_field_names_from_sdl(&schema.sdl());
+        let introspected = introspection_mutation_fields(schema);
+        // No SDL field may be hidden from introspection (a visible=false mutation).
+        for f in &sdl_fields {
+            assert!(
+                introspected.iter().any(|i| i == f),
+                "GraphQL mutation `{f}` is in the schema but hidden from introspection \
+                 (a `#[graphql(visible = false)]` field): hidden mutations are forbidden \
+                 by this fence — make it visible so its classification is enforceable, or \
+                 remove it"
+            );
+        }
+        // Completeness both directions against the registered (camelCase) set.
+        for f in &sdl_fields {
+            assert!(
+                registered_camel.contains(&f.as_str()),
+                "unclassified GraphQL mutation `{f}`: add it to \
+                 every_graphql_mutation_has_its_gate with a deliberate gate \
+                 (`require_repo_owner(` for a repo-write, `did_matches(` for a task op)"
+            );
+        }
+        for c in registered_camel {
+            assert!(
+                sdl_fields.iter().any(|f| f == c),
+                "registered GraphQL mutation `{c}` is not in the schema \
+                 (renamed or removed); update the table"
+            );
+        }
+    }
+
+    /// The mutation field names via the engine's introspection, run against a passed
+    /// schema. STRUCTURED (names arrive as data). Uses `fields(includeDeprecated: true)`:
+    /// GraphQL's default omits `@deprecated` fields, but a `#[graphql(deprecation = "...")]`
+    /// mutation is still executable and present in SDL, so without this the cross-check
+    /// above would false-red a deprecated mutation as a hidden (visible=false) one. Only a
+    /// real visibility filter should trip that arm; deprecation is a lifecycle marker.
+    fn introspection_mutation_fields<Q, M, S>(
+        schema: &async_graphql::Schema<Q, M, S>,
+    ) -> Vec<String>
+    where
+        Q: async_graphql::ObjectType + 'static,
+        M: async_graphql::ObjectType + 'static,
+        S: async_graphql::SubscriptionType + 'static,
+    {
+        use async_graphql::Value;
+        let resp = tokio::runtime::Runtime::new()
+            .expect("tokio runtime")
+            .block_on(schema.execute(
+                "{ __schema { mutationType { fields(includeDeprecated: true) { name } } } }",
+            ));
+        assert!(
+            resp.errors.is_empty(),
+            "schema introspection failed: {:?}",
+            resp.errors
+        );
+        let mut names = Vec::new();
+        if let Value::Object(root) = &resp.data {
+            if let Some(Value::Object(sc)) = root.get("__schema") {
+                if let Some(Value::Object(mt)) = sc.get("mutationType") {
+                    if let Some(Value::List(field_list)) = mt.get("fields") {
+                        for field in field_list {
+                            if let Value::Object(fo) = field {
+                                if let Some(Value::String(n)) = fo.get("name") {
+                                    names.push(n.clone());
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        names
+    }
+
+    /// The mutation field names sourced from the built schema's SDL — the completeness
+    /// source (#219). Builds the same DB-free schema as the introspection helper and runs
+    /// its `.sdl()` (sync, no runtime needed) through [`mutation_field_names_from_sdl`].
+    /// SDL includes `#[graphql(visible = false)]` fields, so this source cannot be
+    /// defeated by hiding a mutation from introspection.
+    fn graphql_mutation_fields_from_sdl() -> Vec<String> {
+        use crate::graphql::mutation::MutationRoot;
+        use crate::graphql::query::QueryRoot;
+        use crate::graphql::subscription::SubscriptionRoot;
+        let schema =
+            async_graphql::Schema::build(QueryRoot, MutationRoot, SubscriptionRoot).finish();
+        mutation_field_names_from_sdl(&schema.sdl())
+    }
+
+    /// The mutation type's field names extracted from an async-graphql SDL string via
+    /// the engine's OWN parser AST (`async_graphql::parser::parse_schema`), NOT
+    /// introspection. This is the completeness source (#219): unlike introspection,
+    /// which async-graphql filters by `visible` (a `#[graphql(visible = false)]`
+    /// mutation is executable but ABSENT from introspection), `schema.sdl()` includes
+    /// every registered field, so a hidden mutation still appears here and must carry a
+    /// classified row. Parsing the AST (not a brace/line text scan) means a `}` inside a
+    /// field description cannot mis-slice the type block. camelCase, as async-graphql
+    /// emits. An empty result on a navigation miss is safe: the set-equality in
+    /// `every_graphql_mutation_has_its_gate` turns "no fields" into a loud failure for
+    /// every registered name, never a vacuous pass.
+    fn mutation_field_names_from_sdl(sdl: &str) -> Vec<String> {
+        use async_graphql::parser::parse_schema;
+        use async_graphql::parser::types::{TypeKind, TypeSystemDefinition};
+
+        let doc = parse_schema(sdl).expect("async-graphql SDL must parse");
+        // The mutation root type name: from the `schema { mutation: X }` block if
+        // present, else the GraphQL default "Mutation".
+        let mutation_type = doc
+            .definitions
+            .iter()
+            .find_map(|def| match def {
+                TypeSystemDefinition::Schema(s) => s
+                    .node
+                    .mutation
+                    .as_ref()
+                    .map(|m| m.node.as_str().to_string()),
+                _ => None,
+            })
+            .unwrap_or_else(|| "Mutation".to_string());
+        // The fields of that object type.
+        doc.definitions
+            .iter()
+            .find_map(|def| match def {
+                TypeSystemDefinition::Type(t) if t.node.name.node.as_str() == mutation_type => {
+                    match &t.node.kind {
+                        TypeKind::Object(obj) => Some(
+                            obj.fields
+                                .iter()
+                                .map(|f| f.node.name.node.as_str().to_string())
+                                .collect::<Vec<_>>(),
+                        ),
+                        _ => None,
+                    }
+                }
+                _ => None,
+            })
+            .unwrap_or_default()
+    }
+
+    #[test]
+    fn sdl_extractor_includes_a_hidden_visible_false_field() {
+        // A field that would be hidden from introspection must still be listed from SDL.
+        let sdl = "schema { query: QueryRoot mutation: MutationRoot }\n\
+                   type QueryRoot { ping: Int! }\n\
+                   type MutationRoot { createTask: Int! hiddenMut: Int! }\n";
+        let fields = mutation_field_names_from_sdl(sdl);
+        assert!(
+            fields.iter().any(|f| f == "createTask"),
+            "visible field must be listed: {fields:?}"
+        );
+        assert!(
+            fields.iter().any(|f| f == "hiddenMut"),
+            "a visible=false field (present in SDL, hidden from introspection) MUST be \
+             listed by the completeness source: {fields:?}"
+        );
+    }
+
+    #[test]
+    fn sdl_extractor_is_not_fooled_by_a_brace_in_a_field_description() {
+        // A `}` inside a field description must not truncate the type block (the AST
+        // parse is immune to the text-parse quirk the old design feared).
+        let sdl = "schema { query: QueryRoot mutation: MutationRoot }\n\
+                   type QueryRoot { ping: Int! }\n\
+                   type MutationRoot {\n\
+                     \"\"\"a description with a } brace and { another\"\"\"\n\
+                     createTask: Int!\n\
+                     claimTask: Int!\n\
+                   }\n";
+        let fields = mutation_field_names_from_sdl(sdl);
+        assert!(
+            fields.iter().any(|f| f == "createTask") && fields.iter().any(|f| f == "claimTask"),
+            "both fields must be extracted despite a brace in a description: {fields:?}"
+        );
+    }
+
+    #[test]
+    fn sdl_extractor_returns_empty_for_a_query_only_schema() {
+        let sdl = "schema { query: QueryRoot }\ntype QueryRoot { ping: Int! }\n";
+        assert!(
+            mutation_field_names_from_sdl(sdl).is_empty(),
+            "a schema with no mutation root yields no mutation fields (no panic)"
+        );
+    }
+
+    /// The comment-masked body of `async fn <snake>` in `graphql/mutation.rs`, bounded
+    /// to the resolver's OWN braces (its `{ .. }`), for the marker check. Brace-matching
+    /// on the comment-masked source means a `}` in a trailing helper cannot extend the
+    /// body and a comment marker does not count. A `{`/`}` inside a STRING literal is
+    /// not masked, a documented marker-check-only residual (limit 3); completeness does
+    /// not use this.
+    fn resolver_body(masked_src: &str, snake: &str) -> String {
+        let needle = format!("async fn {snake}(");
+        let start = masked_src.find(&needle).unwrap_or_else(|| {
+            panic!("resolver `async fn {snake}(` not found in graphql/mutation.rs")
+        });
+        let open = start
+            + masked_src[start..]
+                .find('{')
+                .expect("resolver body opening brace");
+        let mut depth = 0i32;
+        let mut end = masked_src.len();
+        for (i, c) in masked_src[open..].char_indices() {
+            match c {
+                '{' => depth += 1,
+                '}' => {
+                    depth -= 1;
+                    if depth == 0 {
+                        end = open + i + 1;
+                        break;
+                    }
+                }
+                _ => {}
+            }
+        }
+        masked_src[open..end].to_string()
+    }
+
+    /// `src` with `//` line comments and `/* */` block comments (nested) blanked to
+    /// spaces, WITHOUT opening a comment inside a `"..."` string (so a `"https://"` does
+    /// not erase the rest of its line). String CONTENTS are left intact, so a brace or
+    /// a gate marker inside a string is not masked; for the marker check that is a
+    /// documented vacuity residual (limit 3), and completeness does not use this at
+    /// all. Char literals and raw/byte strings are not tracked (absent in resolver
+    /// bodies); a `'"'` char or a raw string is the remaining narrow residual.
+    fn mask_comments(src: &str) -> String {
+        let chars: Vec<char> = src.chars().collect();
+        let mut out = String::with_capacity(src.len());
+        let mut i = 0;
+        let mut block_depth = 0usize;
+        let mut in_string = false;
+        while i < chars.len() {
+            let c = chars[i];
+            let next = chars.get(i + 1).copied();
+            if block_depth > 0 {
+                if c == '/' && next == Some('*') {
+                    block_depth += 1;
+                    out.push_str("  ");
+                    i += 2;
+                } else if c == '*' && next == Some('/') {
+                    block_depth -= 1;
+                    out.push_str("  ");
+                    i += 2;
+                } else {
+                    out.push(if c == '\n' { '\n' } else { ' ' });
+                    i += 1;
+                }
+            } else if in_string {
+                out.push(c);
+                if c == '\\' {
+                    if let Some(n) = next {
+                        out.push(n);
+                        i += 2;
+                    } else {
+                        i += 1;
+                    }
+                } else {
+                    if c == '"' {
+                        in_string = false;
+                    }
+                    i += 1;
+                }
+            } else if c == '"' {
+                in_string = true;
+                out.push('"');
+                i += 1;
+            } else if c == '/' && next == Some('/') {
+                while i < chars.len() && chars[i] != '\n' {
+                    out.push(' ');
+                    i += 1;
+                }
+            } else if c == '/' && next == Some('*') {
+                block_depth = 1;
+                out.push_str("  ");
+                i += 2;
+            } else {
+                out.push(c);
+                i += 1;
+            }
+        }
+        out
+    }
+
+    /// The SDL field set (the completeness source of truth, #219) is exactly the
+    /// registered mutations. If async-graphql adds or renames a field, this reddens
+    /// and forces the `every_graphql_mutation_has_its_gate` table to be updated. Uses
+    /// SDL, not introspection: introspection omits `visible = false` fields, so an
+    /// introspection-based pin here would go green while a hidden mutation existed.
+    #[test]
+    fn schema_lists_exactly_the_registered_mutations() {
+        let mut fields = graphql_mutation_fields_from_sdl();
+        fields.sort();
+        assert_eq!(
+            fields,
+            vec!["claimTask", "completeTask", "createTask", "failTask"],
+            "the SDL mutation field set is the authoritative discovered set"
+        );
+    }
+
+    /// The marker check is bounded to a resolver's OWN body and is comment- and
+    /// string-aware: a `//` inside a `"https://"` string does not erase a real gate on
+    /// that line (no false-red), and a trailing module helper's `did_matches(` does not
+    /// rescue a resolver whose own gate was removed (no false-green).
+    #[test]
+    fn resolver_body_is_bounded_and_comment_and_string_aware() {
+        let src = "\
+impl MutationRoot {
+    async fn ok(&self) -> Result<u32> {
+        let _ = (\"see https://x/y\", did_matches(a, b)); // URL string then gate, one line
+        Ok(0)
+    }
+    async fn last(&self) -> Result<u32> {
+        // did_matches( only in a comment
+        Ok(0)
+    }
+}
+fn helper() { let _ = did_matches(z, z); }
+#[cfg(test)]
+mod tests { async fn t() { let _ = did_matches(x, y); } }
+";
+        let masked = mask_comments(src);
+        assert!(
+            resolver_body(&masked, "ok").contains("did_matches("),
+            "a `//` inside a string must not erase the real gate on that line"
+        );
+        assert!(
+            !resolver_body(&masked, "last").contains("did_matches("),
+            "the marker check must be bounded to the resolver's own body (a trailing \
+             helper's marker must not rescue a resolver whose gate was removed)"
+        );
+    }
+
    /// Proves the comment-stripping that GUARD-1 added: a marker that appears only
    /// in a full-line comment (the real `replicas.rs` false-pass shape) must NOT
    /// satisfy a row.
@@ -70,6 +70,8 @@ pub struct PeerResponse {
    pub reachable: bool,
 }

+pub(crate) const PUBLIC_HTTP_URL_REQUIREMENT: &str = "must be a public http(s) URL (no loopback, private, localhost, .localhost, .internal, or .local hosts)";
+
 /// Extract an IPv4 address embedded in an IPv6 literal across the transition
 /// formats that carry one: IPv4-mapped (`::ffff:a.b.c.d`), IPv4-compatible
 /// (`::a.b.c.d`), 6to4 (`2002:WWXX:YYZZ::/16`), and the NAT64 well-known prefix
@@ -82,7 +84,7 @@ pub struct PeerResponse {
 /// sits at a prefix-length-dependent offset (RFC 6052 §2.2) — so they return
 /// `None`. Any caller that needs them blocked must reject the wider
 /// `64:ff9b::/32` itself; `is_public_http_url` does this in its native-v6 arm.
-fn embedded_ipv4(v6: std::net::Ipv6Addr) -> Option<std::net::Ipv4Addr> {
+pub(crate) fn embedded_ipv4(v6: std::net::Ipv6Addr) -> Option<std::net::Ipv4Addr> {
    if let Some(v4) = v6.to_ipv4_mapped().or_else(|| v6.to_ipv4()) {
        return Some(v4);
    }
@@ -102,10 +104,120 @@ fn embedded_ipv4(v6: std::net::Ipv6Addr) -> Option<std::net::Ipv4Addr> {
    None
 }

+/// Whether an IP literal is outside the local/private ranges that an
+/// attacker-controlled outbound URL must never reach. Kept separate from URL
+/// parsing so callers that resolve a DNS name themselves can apply the same
+/// classification to every answer and then pin the approved addresses.
+pub(crate) fn is_public_ip(ip: std::net::IpAddr) -> bool {
+    // Reject loopback/unspecified on the literal as given — catches `::1` and
+    // `::` before the IPv4-folding below (`::1`.to_ipv4() would otherwise map
+    // to a non-loopback `0.0.0.1`).
+    if ip.is_loopback() || ip.is_unspecified() {
+        return false;
+    }
+    // Fold any IPv6 literal that embeds an IPv4 address (mapped, compatible,
+    // 6to4, NAT64) down to that IPv4 so the v4 range checks catch
+    // loopback/private addresses smuggled in via an IPv6 encoding, then
+    // re-check loopback/unspecified.
+    let ip = match ip {
+        std::net::IpAddr::V6(v6) => embedded_ipv4(v6).map(std::net::IpAddr::V4).unwrap_or(ip),
+        v4 => v4,
+    };
+    if ip.is_loopback() || ip.is_unspecified() {
+        return false;
+    }
+    match ip {
+        std::net::IpAddr::V4(v4) => {
+            let o = v4.octets();
+            // RFC1918 private, link-local, CGNAT (100.64.0.0/10), or the
+            // RFC1122 "this host" block 0.0.0.0/8 (never a valid destination;
+            // 0.0.0.0 itself is already caught by the is_unspecified check).
+            if v4.is_private()
+                || v4.is_link_local()
+                || (o[0] == 100 && (o[1] & 0xc0) == 64)
+                || o[0] == 0
+            {
+                return false;
+            }
+        }
+        std::net::IpAddr::V6(v6) => {
+            let s = v6.segments();
+            // fc00::/7 (unique-local), fe80::/10 (link-local), or fec0::/10
+            // (site-local, deprecated by RFC 3879; the link-local mask does
+            // not cover it).
+            if (s[0] & 0xfe00) == 0xfc00 || (s[0] & 0xffc0) == 0xfe80 || (s[0] & 0xffc0) == 0xfec0 {
+                return false;
+            }
+            // Any NAT64 address (64:ff9b::/32) that is not the cleanly
+            // decodable well-known /96 — e.g. the RFC 8215 local-use
+            // 64:ff9b:1::/48 — carries a translated target whose embedded v4
+            // sits at a prefix-length-dependent offset (RFC 6052 §2.2).
+            // Rather than risk a wrong decode across every prefix length we
+            // reject the whole NAT64 space here. The well-known /96 was already
+            // folded to its v4 above and never reaches this arm.
+            if s[0] == 0x0064 && s[1] == 0xff9b {
+                return false;
+            }
+        }
+    }
+    true
+}
+
+/// Mirror-upstream variant of [`is_public_ip`]: in addition to the shared
+/// peer classification, rejects IANA special-purpose ranges that are
+/// non-global, locally meaningful, or unsafe as owner-controlled outbound
+/// fetch endpoints. For IP-literal upstreams, configure-time URL validation
+/// and the fetch-time egress policy share this classifier, so a stored
+/// literal can never fail the worker it is scheduled for. Hostname upstreams
+/// are only shape-checked at configure time: each fetch resolves DNS and
+/// applies this classifier to the answers, so a hostname whose answers turn
+/// non-public still fails at fetch time.
+pub(crate) fn mirror_upstream_ip_permitted(ip: std::net::IpAddr) -> bool {
+    if !is_public_ip(ip) {
+        return false;
+    }
+    let classified = match ip {
+        std::net::IpAddr::V6(v6) => embedded_ipv4(v6).map(std::net::IpAddr::V4).unwrap_or(ip),
+        _ => ip,
+    };
+    match classified {
+        std::net::IpAddr::V4(v4) => {
+            let [a, b, c, _] = v4.octets();
+            // IANA special-purpose ranges that are not globally reachable but
+            // may be routed inside labs, clouds, or operator networks. They are
+            // unsuitable for an owner-controlled outbound fetch even though
+            // `Ipv4Addr::is_private` does not classify them as RFC1918.
+            !(a >= 224
+                || (a == 192 && b == 0 && c == 0)
+                || (a == 192 && b == 0 && c == 2)
+                || (a == 192 && b == 88 && c == 99)
+                || (a == 198 && matches!(b, 18 | 19))
+                || (a == 198 && b == 51 && c == 100)
+                || (a == 203 && b == 0 && c == 113))
+        }
+        std::net::IpAddr::V6(v6) => {
+            let s = v6.segments();
+            // IANA special-purpose ranges that are non-global, locally
+            // meaningful, or unsafe to treat as ordinary forge endpoints.
+            // Reject the complete IETF protocol-assignment /23 rather than
+            // trying to maintain its narrow globally reachable exceptions;
+            // a public forge has no reason to depend on those anycast ranges.
+            !(v6.is_multicast()
+                || (s[0] & 0xffc0) == 0xfec0
+                || (s[0] == 0x0100 && s[1] == 0 && s[2] == 0 && s[3] == 0)
+                || (s[0] == 0x0100 && s[1] == 0 && s[2] == 0 && s[3] == 1)
+                || (s[0] == 0x2001 && (s[1] & 0xfe00) == 0)
+                || (s[0] == 0x2001 && s[1] == 0x0db8)
+                || (s[0] == 0x3fff && (s[1] & 0xf000) == 0)
+                || s[0] == 0x5f00)
+        }
+    }
+}
+
 /// Whether a peer `http_url` is a public http(s) endpoint safe to register.
 /// Rejects non-http(s) schemes, loopback/unspecified/private/link-local IPs,
-/// and `localhost` / `.local` / `.internal` hostnames. Used at announce time
-/// and by the boot-time prune of already-poisoned rows.
+/// and `localhost` / `.localhost` / `.local` / `.internal` hostnames. Used at
+/// announce time and by the boot-time prune of already-poisoned rows.
 pub fn is_public_http_url(raw: &str) -> bool {
    let url = match reqwest::Url::parse(raw) {
        Ok(u) => u,
@@ -118,13 +230,14 @@ pub fn is_public_http_url(raw: &str) -> bool {
        Some(h) => h.to_ascii_lowercase(),
        None => return false,
    };
-    // Drop a single trailing dot (FQDN root): `localhost.` resolves the same as
+    // Drop trailing dots (FQDN root): `localhost.` resolves the same as
    // `localhost`, so normalize before the suffix/equality checks below.
-    if let Some(stripped) = host.strip_suffix('.') {
-        host = stripped.to_string();
-    }
+    // Strip to fixation rather than once, or `localhost..` reduces to
+    // `localhost.` and matches neither the equality nor the suffix check.
+    host.truncate(host.trim_end_matches('.').len());
    if host.is_empty()
        || host == "localhost"
+        || host.ends_with(".localhost")
        || host.ends_with(".local")
        || host.ends_with(".internal")
    {
@@ -134,55 +247,16 @@ pub fn is_public_http_url(raw: &str) -> bool {
    // before parsing as an IP.
    let ip_candidate = host.trim_start_matches('[').trim_end_matches(']');
    if let Ok(ip) = ip_candidate.parse::<std::net::IpAddr>() {
-        // Reject loopback/unspecified on the literal as given — catches `::1`
-        // and `::` before the IPv4-folding below (`::1`.to_ipv4() would
-        // otherwise map to a non-loopback `0.0.0.1`).
-        if ip.is_loopback() || ip.is_unspecified() {
-            return false;
-        }
-        // Fold any IPv6 literal that embeds an IPv4 address (mapped, compatible,
-        // 6to4, NAT64) down to that IPv4 so the v4 range checks catch
-        // loopback/private addresses smuggled in via an IPv6 encoding, then
-        // re-check loopback/unspecified.
-        let ip = match ip {
-            std::net::IpAddr::V6(v6) => embedded_ipv4(v6).map(std::net::IpAddr::V4).unwrap_or(ip),
-            v4 => v4,
-        };
-        if ip.is_loopback() || ip.is_unspecified() {
-            return false;
-        }
-        match ip {
-            std::net::IpAddr::V4(v4) => {
-                let o = v4.octets();
-                // RFC1918 private, link-local, CGNAT (100.64.0.0/10), or the
-                // RFC1122 "this host" block 0.0.0.0/8 (never a valid destination;
-                // 0.0.0.0 itself is already caught by the is_unspecified check).
-                if v4.is_private()
-                    || v4.is_link_local()
-                    || (o[0] == 100 && (o[1] & 0xc0) == 64)
-                    || o[0] == 0
-                {
-                    return false;
-                }
-            }
-            std::net::IpAddr::V6(v6) => {
-                let s = v6.segments();
-                // fc00::/7 (unique-local) or fe80::/10 (link-local)
-                if (s[0] & 0xfe00) == 0xfc00 || (s[0] & 0xffc0) == 0xfe80 {
-                    return false;
-                }
-                // Any NAT64 address (64:ff9b::/32) that is not the cleanly
-                // decodable well-known /96 — e.g. the RFC 8215 local-use
-                // 64:ff9b:1::/48 — carries a translated target whose embedded v4
-                // sits at a prefix-length-dependent offset (RFC 6052 §2.2).
-                // Rather than risk a wrong decode across every prefix length we
-                // reject the whole NAT64 space here. The well-known /96 was
-                // already folded to its v4 above and never reaches this arm.
-                if s[0] == 0x0064 && s[1] == 0xff9b {
-                    return false;
-                }
-            }
-        }
+        // IP literals share the single DNS/literal classification in
+        // `is_public_ip`; hostname-only rules stay local to this function.
+        return is_public_ip(ip);
+    } else if !host.contains('.') {
+        // A name with no dot at all. The `.local` / `.internal` rules above are
+        // suffix checks, so a single label slips both, and the resolver decides
+        // what it means by appending a search domain we cannot see from here.
+        // Only reachable when the host did not parse as an IP literal, which is
+        // what keeps bracketed IPv6 (also dotless) on the accepting path.
+        return false;
    }
    true
 }
@@ -248,9 +322,9 @@ pub async fn announce(
    // and turn our outbound sync-notify fan-out into an SSRF probe — and bury
    // the real peers under junk so node-origin repos stop replicating.
    if !is_public_http_url(&req.http_url) {
-        return Err(AppError::BadRequest(
-            "http_url must be a public http(s) URL (no loopback, private, or .internal/.local hosts)".into(),
-        ));
+        return Err(AppError::BadRequest(format!(
+            "http_url {PUBLIC_HTTP_URL_REQUIREMENT}"
+        )));
    }

    // Reject self-announcements: a peer row whose http_url is our own public
@@ -520,7 +594,7 @@ mod tests {

    #[test]
    fn accepts_public_https_and_http() {
-        assert!(is_public_http_url("https://node.gitlawb.com"));
+        assert!(is_public_http_url("https://node.twigpine.com"));
        assert!(is_public_http_url("https://manila.gitlawb.com/"));
        assert!(is_public_http_url("http://203.0.113.10:7545"));
    }
@@ -529,6 +603,7 @@ mod tests {
    fn rejects_loopback_private_and_internal() {
        for bad in [
            "http://localhost:7545",
+            "http://node.localhost:7545",
            "http://127.0.0.1:5432/",
            "http://localhost:22/",
            "http://0.0.0.0:7545",
@@ -544,7 +619,7 @@ mod tests {
            "http://[::1]:7545",
            "http://gitlawb-node.internal:7545",
            "http://my-node.local/",
-            "ftp://node.gitlawb.com",
+            "ftp://node.twigpine.com",
            "not-a-url",
            "",
            // Trailing-dot FQDN of an internal host.
@@ -577,6 +652,51 @@ mod tests {
        }
    }

+    #[test]
+    fn rejects_dotless_single_label_hosts() {
+        // `.local` and `.internal` are suffix checks, so a name with no dot at
+        // all never matches one. A single-label name is completed by whatever
+        // search domain the resolver is configured with, which is not a
+        // destination we can reason about from the URL.
+        for bad in [
+            "http://local/",
+            "http://internal/",
+            "http://intranet/",
+            "http://wpad/",
+            "http://metadata/",
+            "http://local./",
+            "http://internal.:7545",
+        ] {
+            assert!(!is_public_http_url(bad), "{bad:?} must be rejected");
+        }
+    }
+
+    #[test]
+    fn rejects_deprecated_site_local_v6() {
+        // fec0::/10, deprecated by RFC 3879. The link-local mask (0xffc0 ==
+        // 0xfe80) does not cover it: 0xfec0 & 0xffc0 is 0xfec0.
+        assert!(!super::is_public_ip("fec0::1".parse().unwrap()));
+        assert!(!super::is_public_ip("feff:ffff::1".parse().unwrap()));
+        for bad in ["http://[fec0::1]/", "http://[feff:ffff::1]:7545"] {
+            assert!(!is_public_http_url(bad), "{bad:?} must be rejected");
+        }
+    }
+
+    #[test]
+    fn rejects_repeated_trailing_root_dots() {
+        // Stripping a single root dot leaves `localhost.`, which matches
+        // neither the equality nor the suffix check. The subdomain form now
+        // exercises the `.localhost` suffix rejection together with the extra
+        // trailing dot.
+        for bad in [
+            "http://localhost../",
+            "http://localhost...:7545",
+            "http://node.localhost../",
+        ] {
+            assert!(!is_public_http_url(bad), "{bad:?} must be rejected");
+        }
+    }
+
    #[test]
    fn accepts_6to4_and_nat64_wrapping_public_v4() {
        // Fold-and-recheck stays consistent with the mapped/compatible handling:
@@ -1530,6 +1650,31 @@ mod tests {
        )
    }

+    #[sqlx::test]
+    async fn announce_rejects_localhost_subdomains_with_an_accurate_error(pool: PgPool) {
+        let state = test_state(pool).await;
+        let did = Keypair::generate().did().to_string();
+        let resp = announce_only(state.clone())
+            .oneshot(announce_as(
+                &did,
+                &announce_body(&did, "https://node.localhost:7545"),
+            ))
+            .await
+            .unwrap();
+
+        let (status, error, message) = status_and_error(resp).await;
+        assert_eq!(status, StatusCode::BAD_REQUEST);
+        assert_eq!(error, "bad_request");
+        assert_eq!(
+            message,
+            "http_url must be a public http(s) URL (no loopback, private, localhost, .localhost, .internal, or .local hosts)"
+        );
+        assert!(
+            snapshot(&state.db, &did).await.is_none(),
+            "a rejected .localhost announce must leave no row behind"
+        );
+    }
+
    /// U4 scenario 1, and the first test the keyid branch has ever had: a caller
    /// who proved control of one DID must not announce another's. Kills
    /// neutralizing the handler's keyid comparison, and kills demoting this
@@ -50,6 +50,17 @@ pub async fn create_pr(
    let target_branch = req
        .target_branch
        .unwrap_or_else(|| record.default_branch.clone());
+
+    // Validate both refs before they are stored, since both are later
+    // interpolated into git argv (git diff / worktree add / merge). Validate the
+    // RESOLVED target, not just a caller-supplied one: create_repo also gates
+    // default_branch, but validating here as well means a PR can never feed the
+    // git sink an unchecked ref even if a default was poisoned by an older row
+    // or a future path that skips create_repo's gate.
+    crate::api::validate_git_ref_with_git(&state.git_bin, &req.source_branch)
+        .map_err(crate::api::map_git_ref_validation_error)?;
+    crate::api::validate_git_ref_with_git(&state.git_bin, &target_branch)
+        .map_err(crate::api::map_git_ref_validation_error)?;
    let number = state.db.next_pr_number(&record.id).await?;
    let now = Utc::now().to_rfc3339();

@@ -213,7 +224,7 @@ pub async fn merge_pr(
        .repo_store
        .acquire_write(&record.owner_did, &record.name)
        .await
-        .map_err(|e| AppError::Git(e.to_string()))?;
+        .map_err(|e| crate::api::repos::acquire_write_app_error(&e, &name))?;
    let disk_path = guard.path().to_path_buf();
    let merger_did = auth.0;
    let merge_result = store::merge_branch(
@@ -225,7 +236,10 @@ pub async fn merge_pr(
    );

    // Always release the advisory lock — even on error; upload to Tigris only on success.
-    guard.release(merge_result.is_ok()).await;
+    // Short-circuit on a refused publish before the PR is marked merged and
+    // before the webhook fires. Both are irreversible announcements of a merge
+    // commit that only exists on this node's disk.
+    guard.release(merge_result.is_ok()).await.into_result()?;

    let merge_sha = merge_result.map_err(|e| AppError::Git(e.to_string()))?;

@@ -424,3 +438,105 @@ pub async fn list_comments(
    let comments = state.db.list_pr_comments(&pr.id).await?;
    Ok(Json(serde_json::json!({ "comments": comments })))
 }
+
+/// #173 F1 follow-up: merge holds no admission permit, so an exhausted write-lock
+/// pool must shed 503 + Retry-After.
+#[cfg(test)]
+mod lock_pool_shed_tests {
+    use super::*;
+    use axum::response::IntoResponse;
+    use sqlx::PgPool;
+
+    fn seed_repo(owner_did: &str, name: &str) -> crate::db::RepoRecord {
+        let now = Utc::now();
+        crate::db::RepoRecord {
+            id: Uuid::new_v4().to_string(),
+            name: name.to_string(),
+            owner_did: owner_did.to_string(),
+            description: None,
+            is_public: true,
+            default_branch: "main".to_string(),
+            created_at: now,
+            updated_at: now,
+            disk_path: format!("/tmp/{name}"),
+            forked_from: None,
+            machine_id: None,
+        }
+    }
+
+    #[sqlx::test]
+    async fn merge_pr_lock_pool_exhaustion_sheds_503_not_500(pool: PgPool) {
+        let owner = "did:key:zMERGELOCKPOOLOWNERAAAAAAAAAAAAAAAAAAAAA";
+        let mut state = crate::test_support::test_state(pool.clone()).await;
+        state.repo_store = crate::git::repo_store::RepoStore::new(
+            std::path::PathBuf::from("/tmp/gitlawb-pulls-lockpool"),
+            None,
+            crate::git::repo_store::build_lock_pool(&pool, 1, std::time::Duration::from_secs(1)),
+            std::time::Duration::from_secs(300),
+        );
+
+        let repo = seed_repo(owner, "lp-merge");
+        let repo_id = repo.id.clone();
+        state.db.create_repo(&repo).await.expect("seed repo");
+        let now = Utc::now().to_rfc3339();
+        state
+            .db
+            .create_pr(&PullRequest {
+                id: Uuid::new_v4().to_string(),
+                repo_id: repo_id.clone(),
+                number: 1,
+                title: "lp".to_string(),
+                body: None,
+                author_did: owner.to_string(),
+                source_branch: "feature".to_string(),
+                target_branch: "main".to_string(),
+                status: "open".to_string(),
+                merged_by_did: None,
+                merged_at: None,
+                created_at: now.clone(),
+                updated_at: now,
+            })
+            .await
+            .expect("seed open PR");
+
+        let held = state
+            .repo_store
+            .acquire_write(owner, "other-repo")
+            .await
+            .expect("the first write takes the only lock-pool connection");
+
+        let shed = merge_pr(
+            State(state.clone()),
+            Extension(AuthenticatedDid(owner.to_string())),
+            Path((owner.to_string(), "lp-merge".to_string(), 1)),
+        )
+        .await;
+        let err = shed.expect_err("an exhausted lock pool must fail the call");
+        let resp = err.into_response();
+        assert_eq!(
+            resp.status(),
+            StatusCode::SERVICE_UNAVAILABLE,
+            "merge_pr: an exhausted write-lock pool must shed 503, not a 500 git error"
+        );
+        assert_eq!(
+            resp.headers()
+                .get("retry-after")
+                .map(|v| v.to_str().unwrap()),
+            Some("1"),
+            "merge_pr: a capacity shed must tell the client when to retry"
+        );
+
+        let _ = held.release(false).await;
+        let admitted = merge_pr(
+            State(state.clone()),
+            Extension(AuthenticatedDid(owner.to_string())),
+            Path((owner.to_string(), "lp-merge".to_string(), 1)),
+        )
+        .await;
+        assert!(
+            !matches!(admitted, Err(AppError::Overloaded(_))),
+            "with the lock pool free, merge_pr must not be shed as capacity; got {:?}",
+            admitted.err()
+        );
+    }
+}
@@ -39,11 +39,11 @@ pub async fn register_replica(
 ) -> Result<(StatusCode, Json<serde_json::Value>)> {
    validate_replica_url(&req.url)?;

-    let record = state
-        .db
-        .get_repo(&owner, &repo)
-        .await?
-        .ok_or_else(|| AppError::RepoNotFound(format!("{owner}/{repo}")))?;
+    // The mutation response carries repo metadata (name, replica count), so
+    // registration applies the same read-visibility decision as the listing:
+    // a caller who may not read the repo gets the same 404 as a missing one.
+    let (record, _rules) =
+        crate::api::authorize_repo_read(&state, &owner, &repo, Some(auth.0.as_str()), "/").await?;

    let replica_did = &auth.0;

@@ -93,11 +93,10 @@ pub async fn unregister_replica(
    Extension(auth): Extension<AuthenticatedDid>,
    Path((owner, repo)): Path<(String, String)>,
 ) -> Result<Json<serde_json::Value>> {
-    let record = state
-        .db
-        .get_repo(&owner, &repo)
-        .await?
-        .ok_or_else(|| AppError::RepoNotFound(format!("{owner}/{repo}")))?;
+    // Same gate as register_replica: removal mutates repo-scoped metadata and
+    // its response discloses the replica count.
+    let (record, _rules) =
+        crate::api::authorize_repo_read(&state, &owner, &repo, Some(auth.0.as_str()), "/").await?;

    let replica_did = &auth.0;
    state.db.unregister_replica(&record.id, replica_did).await?;
@@ -1,4 +1,4 @@
-//! DID resolution — look up a node by its gitlawb DID.
+//! DID resolution — look up a node by its Twigpine DID.
 //!
 //! First checks the local peers table (HTTP-layer peer discovery),
 //! then queries the Kademlia DHT via the libp2p swarm.
@@ -14,7 +14,7 @@ use crate::state::AppState;

 /// GET /api/v1/resolve/{did}
 ///
-/// Resolves a gitlawb DID to an HTTP URL and p2p multiaddr.
+/// Resolves a Twigpine DID to an HTTP URL and p2p multiaddr.
 /// Checks local peer cache first, then falls back to Kademlia DHT lookup.
 pub async fn resolve_did(
    State(state): State<AppState>,
@@ -0,0 +1,782 @@
+//! Opaque, integrity-protected continuation tokens for the task read surfaces.
+//!
+//! The task list pages by keyset over `(created_at, id)`, but the rows a
+//! caller may *see* are a filtered subset of the rows the query has to
+//! *examine*. Handing the caller a raw `(created_at, id)` cursor therefore
+//! forced a choice between two broken options (#327 review): anchor the cursor
+//! on the last visible row, and a window of denied tasks longer than the scan
+//! budget stalls paging forever; or anchor it on the last examined row, and a
+//! denied read leaks the id and timestamp of a task `GET /tasks/{id}` other-
+//! wise 404s.
+//!
+//! A server-issued token removes the choice. The position it carries is the
+//! last *examined* candidate, so paging always advances, and the payload is
+//! opaque and MAC'd, so the caller learns nothing from it and cannot forge one
+//! naming a row of their choosing.
+//!
+//! The position must be *confidential*, not merely unforgeable: a token that
+//! merely signed a base64 payload would still let its holder read the id and
+//! timestamp of the denied row it names, which is the disclosure the token
+//! exists to prevent. The payload is therefore encrypted under a
+//! synthetic-IV construction (SIV): the tag is an HMAC over the filter and the
+//! plaintext, and it doubles as the IV seeding the keystream the plaintext is
+//! XORed with. That needs no randomness source and no dependency beyond the
+//! `hmac`/`sha2` pair already used for webhook signatures and blob recipient
+//! tags, and it is decrypt-last: nothing is parsed until the tag verifies.
+//!
+//! Making the token the *only* accepted cursor also fixes the ordering-domain
+//! bug the same review found. `agent_tasks.created_at` is TEXT and compared as
+//! TEXT, so `...Z` and `...+00:00` denote one instant but sort differently. A
+//! caller-typed timestamp could silently skip or repeat same-time rows. The
+//! token instead carries the stored string verbatim, so the value compared is
+//! always a value the server wrote.
+//!
+//! A token is bound to the *caller* as well as the filter. The position it
+//! names is the last examined candidate, not the last visible one, so it
+//! encodes how far a scan got under one caller's visibility. Resuming it as a
+//! different caller would start the scan past rows that caller is entitled to
+//! read, silently dropping them from the answer (#327 review). The presenting
+//! caller's normalized DID (empty and flagged absent when anonymous) is
+//! therefore part of the MAC input, so a token presented by anyone else fails
+//! to verify rather than under-reporting.
+//!
+//! Wire form: `v1.<iv>.<body>`, both parts base64url unpadded: the 16-byte
+//! synthetic IV (the truncated MAC, which also seeds the keystream) and the
+//! padded, encrypted JSON payload.
+
+use base64::engine::general_purpose::URL_SAFE_NO_PAD;
+use base64::Engine;
+use hmac::{Hmac, Mac};
+use serde::{Deserialize, Serialize};
+use sha2::Sha256;
+
+use crate::error::AppError;
+
+type HmacSha256 = Hmac<Sha256>;
+
+const CURSOR_PREFIX: &str = "v1";
+const KEY_DERIVATION_LABEL: &[u8] = b"gitlawb/tasks-cursor-key/v1";
+const TAG_DOMAIN: &[u8] = b"gitlawb/tasks-cursor-tag/v1";
+const STREAM_DOMAIN: &[u8] = b"gitlawb/tasks-cursor-stream/v1";
+/// Truncated MAC length, and the synthetic IV width. 128 bits is far beyond
+/// forgery reach for a token that carries no authority of its own (every page
+/// re-runs the visibility gate against the presenting caller), and keeps the
+/// token short enough to sit in a query string.
+const TAG_LEN: usize = 16;
+
+/// Padding bucket for the serialized cursor payload, in bytes. The keystream
+/// XORs the plaintext byte for byte, so token length would otherwise track
+/// payload length and leak the named row's field widths to a holder who
+/// cannot decrypt anything. 128 comfortably covers the widest realistic
+/// payload: a 35-char RFC3339 timestamp (`to_rfc3339()` renders 29 chars at
+/// millisecond precision and 32 at microsecond, 35 only with all nine
+/// fractional digits), a 36-char server-minted UUID, and the fixed JSON
+/// envelope plus a 10-digit expiry, which together total 101 bytes. Should a future payload ever exceed the bucket, the length
+/// rounds up to the next multiple instead of failing, so `encode` never
+/// breaks — it just mints a longer token.
+const PAYLOAD_BUCKET: usize = 128;
+
+/// How long an issued cursor stays acceptable. Keyset positions never go stale
+/// on their own — `created_at`/`id` are immutable — so this is not a
+/// correctness bound. It bounds how long a token stays valid across a node
+/// restart-and-rotate and keeps an abandoned page from being resumed
+/// indefinitely.
+const CURSOR_TTL_SECS: i64 = 24 * 60 * 60;
+
+/// Node-keyed MAC key for continuation tokens, derived from the node keypair
+/// seed so it needs no configuration and survives restarts of the same node.
+/// Derived rather than used directly so a token forgery oracle could not
+/// bear on the signing key itself.
+#[derive(Clone)]
+pub struct TaskCursorKey([u8; 32]);
+
+impl std::fmt::Debug for TaskCursorKey {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("TaskCursorKey(<redacted>)")
+    }
+}
+
+impl TaskCursorKey {
+    pub fn derive(node_seed: &[u8; 32]) -> Self {
+        let mut mac = HmacSha256::new_from_slice(node_seed).expect("HMAC accepts any key length");
+        mac.update(KEY_DERIVATION_LABEL);
+        let mut key = [0u8; 32];
+        key.copy_from_slice(&mac.finalize().into_bytes());
+        Self(key)
+    }
+}
+
+/// The keyset position a token carries: the last candidate row the previous
+/// request examined, whether or not the caller was allowed to see it.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct TaskPosition {
+    pub created_at: String,
+    pub id: String,
+}
+
+impl TaskPosition {
+    pub fn new(created_at: impl Into<String>, id: impl Into<String>) -> Self {
+        Self {
+            created_at: created_at.into(),
+            id: id.into(),
+        }
+    }
+
+    pub fn as_pair(&self) -> (&str, &str) {
+        (self.created_at.as_str(), self.id.as_str())
+    }
+}
+
+#[derive(Serialize, Deserialize)]
+struct CursorPayload<'a> {
+    /// `created_at` of the last examined candidate, verbatim as stored.
+    t: &'a str,
+    /// `id` of the last examined candidate.
+    i: &'a str,
+    /// Unix expiry.
+    e: i64,
+}
+
+/// The filter a page was issued under. A cursor is only meaningful against the
+/// same `status`/`assignee_did` filter that produced it: resuming a filtered
+/// scan from an unfiltered position (or the reverse) silently skips rows.
+/// Bound into the MAC rather than stored in the payload, so it costs no token
+/// length and cannot be edited without invalidating the tag.
+#[derive(Debug, Clone, Copy)]
+pub struct TaskFilter<'a> {
+    pub status: Option<&'a str>,
+    pub assignee_did: Option<&'a str>,
+}
+
+/// The visibility identity a token was issued to. Normalized through
+/// `normalize_owner_key` so the two spellings of one `did:key` identity
+/// (`did:key:X` and bare `X`) bind identically, matching `did_matches` on the
+/// read path: a caller who authenticates with the other spelling of their own
+/// DID must be able to resume their own page.
+fn caller_binding(caller: Option<&str>) -> &str {
+    caller.map(crate::db::normalize_owner_key).unwrap_or("")
+}
+
+/// The assignee filter a token was issued under. Normalized through
+/// `normalize_owner_key` so the two spellings of one `did:key` identity
+/// (`did:key:X` and bare `X`) bind identically, matching `list_tasks_keyset`
+/// in SQL: minting under `did:key:X` and resuming with bare `X` (or the reverse)
+/// hits the same rows and must verify under the MAC.
+fn assignee_binding(assignee: Option<&str>) -> &str {
+    assignee.map(crate::db::normalize_owner_key).unwrap_or("")
+}
+
+fn cursor_mac(
+    key: &TaskCursorKey,
+    filter: TaskFilter<'_>,
+    caller: Option<&str>,
+    plaintext: &[u8],
+) -> HmacSha256 {
+    let mut mac = HmacSha256::new_from_slice(&key.0).expect("HMAC accepts any key length");
+    mac.update(TAG_DOMAIN);
+    // Length-prefix every field so no two distinct filter/caller/plaintext
+    // tuples can produce the same MAC input.
+    for field in [
+        filter.status.unwrap_or("").as_bytes(),
+        assignee_binding(filter.assignee_did).as_bytes(),
+        caller_binding(caller).as_bytes(),
+        plaintext,
+    ] {
+        mac.update(&(field.len() as u64).to_be_bytes());
+        mac.update(field);
+    }
+    // Presence is distinct from emptiness for all three optional fields, so an
+    // anonymous token cannot collide with one issued to a caller whose
+    // normalized DID is the empty string.
+    mac.update(&[
+        u8::from(filter.status.is_some()),
+        u8::from(filter.assignee_did.is_some()),
+        u8::from(caller.is_some()),
+    ]);
+    mac
+}
+
+/// Synthetic IV: the authentication tag over the filter and plaintext, which
+/// also seeds the keystream. Deterministic by construction, so no randomness
+/// source is needed and two tokens for the same page are byte-identical.
+fn siv(
+    key: &TaskCursorKey,
+    filter: TaskFilter<'_>,
+    caller: Option<&str>,
+    plaintext: &[u8],
+) -> [u8; TAG_LEN] {
+    let mut out = [0u8; TAG_LEN];
+    out.copy_from_slice(
+        &cursor_mac(key, filter, caller, plaintext)
+            .finalize()
+            .into_bytes()[..TAG_LEN],
+    );
+    out
+}
+
+/// XOR `buf` with the keystream for `iv`. Its own inverse, so encrypt and
+/// decrypt are the same call.
+fn apply_keystream(key: &TaskCursorKey, iv: &[u8; TAG_LEN], buf: &mut [u8]) {
+    for (block_index, chunk) in buf.chunks_mut(32).enumerate() {
+        let mut mac = HmacSha256::new_from_slice(&key.0).expect("HMAC accepts any key length");
+        mac.update(STREAM_DOMAIN);
+        mac.update(iv);
+        mac.update(&(block_index as u64).to_be_bytes());
+        let block = mac.finalize().into_bytes();
+        for (byte, k) in chunk.iter_mut().zip(block.iter()) {
+            *byte ^= k;
+        }
+    }
+}
+
+/// Mint a token resuming at `position` for `filter`, usable only by `caller`.
+///
+/// The serialized payload is padded with trailing whitespace to a fixed
+/// bucket before the SIV tag is computed: the keystream preserves plaintext
+/// length, so an unpadded payload would let the token's length leak the
+/// named row's field widths (a 25-char `to_rfc3339()` rendering mints a
+/// shorter token than a 35-char one). `decode` tolerates the padding because
+/// `serde_json` ignores trailing whitespace, and the tag covers the padded
+/// bytes, so a token minted before padding landed still verifies.
+pub fn encode(
+    key: &TaskCursorKey,
+    filter: TaskFilter<'_>,
+    caller: Option<&str>,
+    position: &TaskPosition,
+) -> String {
+    let mut payload = serde_json::to_vec(&CursorPayload {
+        t: &position.created_at,
+        i: &position.id,
+        e: chrono::Utc::now().timestamp() + CURSOR_TTL_SECS,
+    })
+    .expect("cursor payload is plain strings and an integer");
+    let padded_len = payload.len().next_multiple_of(PAYLOAD_BUCKET);
+    payload.resize(padded_len, b' ');
+    let iv = siv(key, filter, caller, &payload);
+    apply_keystream(key, &iv, &mut payload);
+    format!(
+        "{CURSOR_PREFIX}.{}.{}",
+        URL_SAFE_NO_PAD.encode(iv),
+        URL_SAFE_NO_PAD.encode(&payload)
+    )
+}
+
+/// One rejection message for every way a token can fail to verify. A caller
+/// who mangled a token, replayed an expired one, or tried to move one to a
+/// different filter or a different presenting identity learns only that the
+/// cursor is not usable — never which of those it was, and never anything
+/// about the row it named.
+const INVALID_CURSOR: &str = "invalid or expired cursor";
+
+fn reject() -> AppError {
+    AppError::BadRequest(INVALID_CURSOR.into())
+}
+
+/// Verify a token against `filter` and the presenting `caller`, and return the
+/// position it carries.
+pub fn decode(
+    key: &TaskCursorKey,
+    filter: TaskFilter<'_>,
+    caller: Option<&str>,
+    token: &str,
+) -> crate::error::Result<TaskPosition> {
+    let mut parts = token.split('.');
+    let (Some(version), Some(iv_b64), Some(body_b64), None) =
+        (parts.next(), parts.next(), parts.next(), parts.next())
+    else {
+        return Err(reject());
+    };
+    if version != CURSOR_PREFIX {
+        return Err(reject());
+    }
+    let iv_bytes = URL_SAFE_NO_PAD.decode(iv_b64).map_err(|_| reject())?;
+    let mut plaintext = URL_SAFE_NO_PAD.decode(body_b64).map_err(|_| reject())?;
+    let iv: [u8; TAG_LEN] = iv_bytes.try_into().map_err(|_| reject())?;
+
+    apply_keystream(key, &iv, &mut plaintext);
+    // Authenticate before parsing: until the tag matches, `plaintext` is just
+    // attacker-chosen bytes run through a keystream. `verify_truncated_left`
+    // is the constant-time compare, so a forgery attempt cannot be steered by
+    // timing the first differing byte.
+    cursor_mac(key, filter, caller, &plaintext)
+        .verify_truncated_left(&iv)
+        .map_err(|_| reject())?;
+
+    let decoded: CursorPayload<'_> = serde_json::from_slice(&plaintext).map_err(|_| reject())?;
+    if decoded.e < chrono::Utc::now().timestamp() {
+        return Err(reject());
+    }
+    Ok(TaskPosition::new(decoded.t, decoded.i))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn key() -> TaskCursorKey {
+        TaskCursorKey::derive(&[7u8; 32])
+    }
+
+    fn unfiltered() -> TaskFilter<'static> {
+        TaskFilter {
+            status: None,
+            assignee_did: None,
+        }
+    }
+
+    #[test]
+    fn round_trips_position_verbatim() {
+        let k = key();
+        // A stored timestamp the server wrote, fractional digits and all.
+        let pos = TaskPosition::new("2026-01-03T00:00:00.123456789+00:00", "task-a");
+        let token = encode(&k, unfiltered(), None, &pos);
+        assert_eq!(decode(&k, unfiltered(), None, &token).unwrap(), pos);
+        // A sub-bucket payload pads to exactly one bucket: the padding
+        // claim is an exact width, and every other width assertion here is
+        // bucket-relative, so only an absolute check catches a pad that
+        // skips "next multiple" or a drifted bucket constant.
+        let body = token.split('.').nth(2).expect("token has a body part");
+        let raw = URL_SAFE_NO_PAD.decode(body).unwrap();
+        assert_eq!(
+            raw.len(),
+            PAYLOAD_BUCKET,
+            "normal token must pad to exactly one bucket"
+        );
+    }
+
+    /// The whole point of the token is that the caller may hold it without
+    /// learning the denied row it names. A signed-but-plaintext payload would
+    /// pass every other test here and still fail this one.
+    #[test]
+    fn token_does_not_expose_the_row_it_names() {
+        let k = key();
+        let token = encode(
+            &k,
+            unfiltered(),
+            None,
+            &TaskPosition::new("2026-01-03T00:00:00+00:00", "denied-task-id"),
+        );
+        let body = token.split('.').nth(2).expect("token has a body part");
+        let raw = URL_SAFE_NO_PAD.decode(body).unwrap();
+        let as_text = String::from_utf8_lossy(&raw);
+        for secret in ["denied-task-id", "2026-01-03", "\"t\"", "\"i\""] {
+            assert!(
+                !as_text.contains(secret),
+                "token body must not carry {secret:?} in the clear: {as_text:?}"
+            );
+        }
+        // Also assert it is not merely reordered or whitespace-mangled JSON.
+        assert!(
+            serde_json::from_slice::<serde_json::Value>(&raw).is_err(),
+            "token body must not be parseable as JSON"
+        );
+    }
+
+    /// The keystream preserves plaintext length, so without the fixed-width
+    /// padding a token naming a 35-char timestamp would be measurably longer
+    /// than one naming a 32-char one: the holder would learn the withheld
+    /// row's field widths without decrypting anything. Mint the same shape
+    /// both ways and require identical token length.
+    #[test]
+    fn token_length_does_not_vary_with_position_width() {
+        let k = key();
+        // Ids are server-minted fixed-length UUIDs; the timestamp is the
+        // width that varies in production (`to_rfc3339()` renders 35 chars
+        // with nanoseconds, 25 without fractional seconds).
+        let id = "b3f7a1c2-4d5e-4f6a-8b9c-0d1e2f3a4b5c";
+        let narrow = encode(
+            &k,
+            unfiltered(),
+            None,
+            &TaskPosition::new("2026-01-03T00:00:00+00:00", id),
+        );
+        let wide = encode(
+            &k,
+            unfiltered(),
+            None,
+            &TaskPosition::new("2026-01-03T00:00:00.123456789+00:00", id),
+        );
+        assert_eq!(
+            narrow.len(),
+            wide.len(),
+            "tokens must not leak the named row's field widths: {narrow} vs {wide}"
+        );
+        // The padded token still decodes to the exact stored string.
+        assert_eq!(
+            decode(&k, unfiltered(), None, &wide).unwrap(),
+            TaskPosition::new("2026-01-03T00:00:00.123456789+00:00", id)
+        );
+    }
+
+    /// A token minted before the fixed-width padding landed (unpadded payload,
+    /// tag over the raw bytes) must still verify: `decode` MACs whatever
+    /// plaintext the keystream recovers and `serde_json` tolerates the
+    /// missing trailing whitespace. This pins the back-compat claim in the
+    /// `encode` docstring with a frozen v1 artifact, so a change to the v1
+    /// recipe pins (domains, MAC input, keystream, segment layout, field
+    /// names) breaks this test instead of silently re-deriving under the
+    /// new internals. It does not pin the padding: `decode` never consults
+    /// `PAYLOAD_BUCKET` and the fixture is unpadded, so the padding-width
+    /// claim belongs to `oversized_payload_rounds_up_to_next_bucket`.
+    ///
+    /// Fixture expiry is fixed at 2100-01-01 (`4102444800`): it must stay in
+    /// the future whenever the test runs.
+    #[test]
+    fn unpadded_legacy_token_still_verifies() {
+        let k = key();
+        let filter = unfiltered();
+        let position = TaskPosition::new("2026-01-03T00:00:00+00:00", "legacy-task");
+        // Pre-padding recipe minted once and frozen: serialize with no
+        // padding, tag, and encrypt. Deterministic by construction (SIV needs
+        // no randomness), so this literal is stable across runs.
+        let token = "v1.ILJyy7JZScm_h-HM1-CggA.l4kJUoHTbqSGTVKL0DP4hORnbOS1_gHYV2fwJYXd2QtM-6TBMCKh-wvluIiaEYC9rUSUPVSteAKkp7jBCmK_bR2D";
+        assert_eq!(
+            decode(&k, filter, None, token).unwrap(),
+            position,
+            "pre-padding token must verify under the current decode"
+        );
+        // The fixture is a genuine pre-padding artifact, so its decoded
+        // body must stay below one bucket: a literal re-minted under the
+        // padded recipe would retire this back-compat pin silently.
+        let body = token.split('.').nth(2).expect("token has a body part");
+        let raw = URL_SAFE_NO_PAD.decode(body).unwrap();
+        assert!(
+            raw.len() < PAYLOAD_BUCKET,
+            "legacy fixture must be the unpadded shape: {} bytes",
+            raw.len()
+        );
+    }
+
+    /// The `PAYLOAD_BUCKET` docstring promises an oversized payload rounds up
+    /// to the next multiple instead of failing. Mint a position whose id
+    /// pushes the serialized payload past one bucket and require the token
+    /// to round-trip (at two buckets).
+    #[test]
+    fn oversized_payload_rounds_up_to_next_bucket() {
+        let k = key();
+        // ~140-char id: payload crosses the 128-byte bucket.
+        let long_id = "a".repeat(140);
+        let position = TaskPosition::new("2026-01-03T00:00:00+00:00", long_id.clone());
+        let probe = serde_json::to_vec(&CursorPayload {
+            t: &position.created_at,
+            i: &position.id,
+            e: 1_700_000_000,
+        })
+        .expect("cursor payload is plain strings and an integer");
+        assert!(
+            probe.len() > PAYLOAD_BUCKET,
+            "fixture payload must exceed one bucket: {} bytes",
+            probe.len()
+        );
+        assert!(
+            probe.len() <= 2 * PAYLOAD_BUCKET,
+            "fixture payload must fit in two buckets: {} bytes",
+            probe.len()
+        );
+        let token = encode(&k, unfiltered(), None, &position);
+        assert_eq!(
+            decode(&k, unfiltered(), None, &token).unwrap(),
+            position,
+            "oversized payload must round-trip at two buckets"
+        );
+        // The padded body must be exactly two buckets, not merely longer
+        // than one: the padding claim is an exact width, not a direction.
+        let body = token.split('.').nth(2).expect("token has a body part");
+        let raw = URL_SAFE_NO_PAD.decode(body).unwrap();
+        assert_eq!(
+            raw.len(),
+            2 * PAYLOAD_BUCKET,
+            "oversized payload must pad to exactly two buckets"
+        );
+    }
+
+    /// Callers paste the token straight into a query string, so it must carry
+    /// no character that needs percent-encoding.
+    #[test]
+    fn token_is_url_safe_verbatim() {
+        let token = encode(
+            &key(),
+            TaskFilter {
+                status: Some("pending"),
+                assignee_did: Some("did:key:z6MkAssignee"),
+            },
+            None,
+            &TaskPosition::new("2026-01-03T00:00:00.123456789+00:00", "task-a"),
+        );
+        assert!(
+            token
+                .chars()
+                .all(|c| c.is_ascii_alphanumeric() || matches!(c, '-' | '_' | '.')),
+            "token must be query-safe without encoding: {token}"
+        );
+    }
+
+    /// Flipping any byte of a token must fail the tag, not decode to a
+    /// different position: the keystream is malleable on its own, the tag is
+    /// what stops a chosen-position forgery.
+    #[test]
+    fn rejects_any_tampered_byte() {
+        let k = key();
+        let token = encode(
+            &k,
+            unfiltered(),
+            None,
+            &TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a"),
+        );
+        let parts: Vec<&str> = token.split('.').collect();
+        for part in [1usize, 2] {
+            for position in [0usize, 3] {
+                let mut bytes = parts[part].as_bytes().to_vec();
+                bytes[position] = if bytes[position] == b'A' { b'B' } else { b'A' };
+                let mut mangled: Vec<String> = parts.iter().map(|p| p.to_string()).collect();
+                mangled[part] = String::from_utf8(bytes).unwrap();
+                assert!(
+                    decode(&k, unfiltered(), None, &mangled.join(".")).is_err(),
+                    "byte {position} of part {part} must not be malleable"
+                );
+            }
+        }
+        // Sanity: the untouched token still decodes, so the loop above is not
+        // passing because every token is rejected.
+        assert!(decode(&k, unfiltered(), None, &token).is_ok());
+    }
+
+    #[test]
+    fn rejects_token_from_another_node() {
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let token = encode(&TaskCursorKey::derive(&[1u8; 32]), unfiltered(), None, &pos);
+        assert!(decode(
+            &TaskCursorKey::derive(&[2u8; 32]),
+            unfiltered(),
+            None,
+            &token
+        )
+        .is_err());
+    }
+
+    #[test]
+    fn rejects_cursor_moved_to_a_different_filter() {
+        let k = key();
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let token = encode(
+            &k,
+            TaskFilter {
+                status: Some("pending"),
+                assignee_did: None,
+            },
+            None,
+            &pos,
+        );
+        assert!(decode(&k, unfiltered(), None, &token).is_err());
+        assert!(decode(
+            &k,
+            TaskFilter {
+                status: Some("claimed"),
+                assignee_did: None
+            },
+            None,
+            &token
+        )
+        .is_err());
+        assert!(decode(
+            &k,
+            TaskFilter {
+                status: Some("pending"),
+                assignee_did: None
+            },
+            None,
+            &token
+        )
+        .is_ok());
+    }
+
+    /// A filter of `Some("")` must not verify a token minted with `None`.
+    #[test]
+    fn distinguishes_absent_filter_from_empty_filter() {
+        let k = key();
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let token = encode(&k, unfiltered(), None, &pos);
+        assert!(decode(
+            &k,
+            TaskFilter {
+                status: Some(""),
+                assignee_did: None
+            },
+            None,
+            &token
+        )
+        .is_err());
+    }
+
+    /// A cursor names how far a scan got under one caller's visibility. Moved
+    /// to a different presenting identity it would start that caller's scan
+    /// past rows they are entitled to read, so it must fail to verify rather
+    /// than silently under-report (#327 review).
+    #[test]
+    fn rejects_cursor_moved_to_a_different_caller() {
+        let k = key();
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let alice = Some("did:key:z6MkAlice");
+        let bob = Some("did:key:z6MkBob");
+
+        let anon_token = encode(&k, unfiltered(), None, &pos);
+        assert!(
+            decode(&k, unfiltered(), alice, &anon_token).is_err(),
+            "an anonymously minted cursor must not resume an authenticated scan"
+        );
+        assert!(decode(&k, unfiltered(), None, &anon_token).is_ok());
+
+        let alice_token = encode(&k, unfiltered(), alice, &pos);
+        assert!(
+            decode(&k, unfiltered(), bob, &alice_token).is_err(),
+            "one caller's cursor must not resume another caller's scan"
+        );
+        assert!(
+            decode(&k, unfiltered(), None, &alice_token).is_err(),
+            "an authenticated cursor must not resume an anonymous scan"
+        );
+        assert!(decode(&k, unfiltered(), alice, &alice_token).is_ok());
+    }
+
+    /// The binding is on identity, not spelling: `did:key:X` and bare `X` are
+    /// one caller everywhere else on the read path (`did_matches`), so a caller
+    /// who presents the other form of their own DID must keep their own page.
+    #[test]
+    fn caller_binding_is_normalized_across_did_key_spellings() {
+        let k = key();
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let token = encode(&k, unfiltered(), Some("did:key:z6MkAlice"), &pos);
+        assert_eq!(
+            decode(&k, unfiltered(), Some("z6MkAlice"), &token).unwrap(),
+            pos
+        );
+        // A different method sharing the base58 tail is a different principal.
+        assert!(decode(&k, unfiltered(), Some("did:web:z6MkAlice"), &token).is_err());
+    }
+
+    /// The filter binding is on identity, not spelling: `assignee_did` matches
+    /// normalized keys in SQL (`normalize_owner_key`), so minting under
+    /// `did:key:X` and resuming under bare `X` (or vice versa) must succeed,
+    /// while `did:web:X` remains distinct.
+    #[test]
+    fn assignee_filter_binding_is_normalized_across_did_key_spellings() {
+        let k = key();
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let filter_did_key = TaskFilter {
+            status: Some("pending"),
+            assignee_did: Some("did:key:z6MkAssignee"),
+        };
+        let filter_bare = TaskFilter {
+            status: Some("pending"),
+            assignee_did: Some("z6MkAssignee"),
+        };
+        let filter_web = TaskFilter {
+            status: Some("pending"),
+            assignee_did: Some("did:web:z6MkAssignee"),
+        };
+
+        let token_from_did_key = encode(&k, filter_did_key, None, &pos);
+        assert_eq!(
+            decode(&k, filter_bare, None, &token_from_did_key).unwrap(),
+            pos
+        );
+        assert!(decode(&k, filter_web, None, &token_from_did_key).is_err());
+
+        let token_from_bare = encode(&k, filter_bare, None, &pos);
+        assert_eq!(
+            decode(&k, filter_did_key, None, &token_from_bare).unwrap(),
+            pos
+        );
+        assert!(decode(&k, filter_web, None, &token_from_bare).is_err());
+    }
+
+    /// Anonymous is a distinct binding from a caller whose normalized DID is
+    /// the empty string, the same way `Some("")` is distinct from `None` for
+    /// the filter fields.
+    #[test]
+    fn distinguishes_anonymous_from_empty_caller() {
+        let k = key();
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let token = encode(&k, unfiltered(), None, &pos);
+        assert!(decode(&k, unfiltered(), Some(""), &token).is_err());
+    }
+
+    /// Length-prefixing must stop a status/assignee pair from being re-split.
+    #[test]
+    fn rejects_field_boundary_shift() {
+        let k = key();
+        let pos = TaskPosition::new("2026-01-03T00:00:00+00:00", "task-a");
+        let token = encode(
+            &k,
+            TaskFilter {
+                status: Some("pend"),
+                assignee_did: Some("ing"),
+            },
+            None,
+            &pos,
+        );
+        assert!(decode(
+            &k,
+            TaskFilter {
+                status: Some("pending"),
+                assignee_did: Some("")
+            },
+            None,
+            &token
+        )
+        .is_err());
+    }
+
+    #[test]
+    fn rejects_expired_token() {
+        let k = key();
+        // Minted the same way `encode` does, but already past its expiry.
+        let mut payload = serde_json::to_vec(&CursorPayload {
+            t: "2026-01-03T00:00:00+00:00",
+            i: "task-a",
+            e: chrono::Utc::now().timestamp() - 1,
+        })
+        .unwrap();
+        let iv = siv(&k, unfiltered(), None, &payload);
+        apply_keystream(&k, &iv, &mut payload);
+        let expired = format!(
+            "v1.{}.{}",
+            URL_SAFE_NO_PAD.encode(iv),
+            URL_SAFE_NO_PAD.encode(&payload)
+        );
+        let err = decode(&k, unfiltered(), None, &expired).unwrap_err();
+        assert!(err.to_string().contains(INVALID_CURSOR));
+    }
+
+    #[test]
+    fn rejects_malformed_shapes() {
+        let k = key();
+        for bad in [
+            "",
+            "v1",
+            "v1.",
+            "v1.abc",
+            "v2.abc.def",
+            "v1.abc.def.ghi",
+            "v1.!!!.def",
+        ] {
+            assert!(
+                decode(&k, unfiltered(), None, bad).is_err(),
+                "must reject {bad:?}"
+            );
+        }
+    }
+
+    /// Every rejection path renders the same text, so a caller cannot tell a
+    /// forged token from an expired one from a filter mismatch.
+    #[test]
+    fn every_rejection_is_indistinguishable() {
+        let k = key();
+        for bad in ["v1.abc.def", "not-a-cursor", "v1..", "v9.a.b"] {
+            assert_eq!(
+                decode(&k, unfiltered(), None, bad).unwrap_err().to_string(),
+                format!("invalid request: {INVALID_CURSOR}")
+            );
+        }
+    }
+}
@@ -46,9 +46,10 @@ pub async fn create_webhook(
    // endpoints (SSRF). Delivery runs on the shared no-redirect client
    // (main.rs), which closes the 3xx-to-internal bounce.
    if !crate::api::peers::is_public_http_url(&req.url) {
-        return Err(AppError::BadRequest(
-            "webhook URL must be a public http(s) URL (no loopback, private, or .internal/.local hosts)".into(),
-        ));
+        return Err(AppError::BadRequest(format!(
+            "webhook URL {}",
+            crate::api::peers::PUBLIC_HTTP_URL_REQUIREMENT
+        )));
    }

    let events = req.events.unwrap_or_else(|| vec!["*".into()]);
@@ -141,7 +142,19 @@ pub async fn delete_webhook(

 #[cfg(test)]
 mod tests {
-    use crate::api::peers::is_public_http_url;
+    use axum::extract::{Extension, Path, State};
+    use axum::Json;
+    use chrono::Utc;
+    use gitlawb_core::identity::Keypair;
+    use sqlx::PgPool;
+    use uuid::Uuid;
+
+    use super::{create_webhook, CreateWebhookRequest};
+    use crate::api::peers::{is_public_http_url, PUBLIC_HTTP_URL_REQUIREMENT};
+    use crate::auth::AuthenticatedDid;
+    use crate::db::RepoRecord;
+    use crate::error::AppError;
+    use crate::test_support::test_state;

    // create_webhook gates req.url through is_public_http_url. Pin the exact
    // SSRF targets from issue #81 so the webhook path can never regress to the
@@ -152,6 +165,7 @@ mod tests {
            "http://127.0.0.1:5432/",
            "http://169.254.169.254/latest/meta-data/",
            "http://localhost/",
+            "http://node.localhost/",
            "http://10.0.0.5/",
            "http://[::1]/",
            // IPv6 transition encodings smuggling loopback v4 (6to4 / NAT64).
@@ -169,4 +183,53 @@ mod tests {
        assert!(is_public_http_url("https://hooks.example.com/gitlawb"));
        assert!(is_public_http_url("http://203.0.113.10:7545/"));
    }
+
+    #[sqlx::test]
+    async fn webhook_localhost_rejection_uses_shared_public_url_contract(pool: PgPool) {
+        let state = test_state(pool).await;
+        let owner = Keypair::generate().did().to_string();
+        let repo_id = Uuid::new_v4().to_string();
+        state
+            .db
+            .create_repo(&RepoRecord {
+                id: repo_id.clone(),
+                name: "webhook-contract".to_string(),
+                owner_did: owner.clone(),
+                description: None,
+                is_public: true,
+                default_branch: "main".to_string(),
+                created_at: Utc::now(),
+                updated_at: Utc::now(),
+                disk_path: "/tmp/webhook-contract.git".to_string(),
+                forked_from: None,
+                machine_id: None,
+            })
+            .await
+            .unwrap();
+
+        let result = create_webhook(
+            State(state.clone()),
+            Extension(AuthenticatedDid(owner.clone())),
+            Path((owner, "webhook-contract".to_string())),
+            Json(CreateWebhookRequest {
+                url: "https://node.localhost/hook".to_string(),
+                secret: None,
+                events: None,
+            }),
+        )
+        .await;
+
+        match result {
+            Err(AppError::BadRequest(message)) => assert_eq!(
+                message,
+                format!("webhook URL {PUBLIC_HTTP_URL_REQUIREMENT}")
+            ),
+            Err(error) => panic!("unexpected webhook rejection: {error}"),
+            Ok(_) => panic!("a .localhost webhook must be rejected"),
+        }
+        assert!(
+            state.db.list_webhooks(&repo_id).await.unwrap().is_empty(),
+            "a rejected .localhost webhook must leave no row behind"
+        );
+    }
 }
@@ -0,0 +1,62 @@
+//! #26 Split PR 2 — three-outcome recovery policy types.
+//!
+//! This module owns the recovery classification policy the reviewer
+//! demanded: only a trustworthy, protocol-defined absence authorizes
+//! a paid re-upload. Everything else stays non-terminal.
+//!
+//! Deliberately narrow: the ANS-104 codec lives in `ans104`, and the
+//! gateway probe, the `verify_anchor` path, and the public verify
+//! endpoint are deferred until the uploader, retrieval proof, and
+//! recovery consumer can be proven as one vertical slice (reviewer 2,
+//! round 5). A probe against a real provider contract — same-network
+//! upload/read pair, redirect-safe client, envelope-supplying read
+//! API, provider-documented absence evidence — belongs in that slice,
+//! not here. What ships here is the policy type plus the rule that
+//! only `DefinitivelyAbsent` permits spending another upload.
+
+/// Outcome of a gateway probe for a persisted `item_id`.
+// Policy surface for the recovery slice: no construction on this
+// head (the probe arrives with the provider contract); the unit
+// test pins the rule meanwhile.
+#[allow(dead_code)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ProbeOutcome {
+    /// The persisted item was served back with a verifiable envelope
+    /// bound to the requested id. No re-upload allowed.
+    Present,
+    /// Trustworthy, protocol-defined proof the item was never
+    /// accepted. Authorizes re-upload. The exact evidence that
+    /// establishes this is defined by the provider contract in the
+    /// vertical slice that introduces the probe — never by a
+    /// mock-only body shape.
+    DefinitivelyAbsent,
+    /// Anything else: transport failure, oversized body, bad
+    /// signature, id mismatch, ambiguous gateway response. The
+    /// outbox stays non-terminal.
+    Indeterminate,
+}
+
+impl ProbeOutcome {
+    /// True iff the recovery code is allowed to spend another paid
+    /// upload request. Only `DefinitivelyAbsent` qualifies.
+    #[allow(dead_code)] // the recovery consumer lives in a later slice
+    pub fn permits_reupload(self) -> bool {
+        matches!(self, ProbeOutcome::DefinitivelyAbsent)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Only `DefinitivelyAbsent` authorizes a paid re-upload. This is
+    /// the policy the reviewer named: collapsing any ambiguous outcome
+    /// to absent spends a second immutable upload for the same
+    /// transition.
+    #[test]
+    fn only_definitively_absent_authorizes_reupload() {
+        assert!(!ProbeOutcome::Present.permits_reupload());
+        assert!(ProbeOutcome::DefinitivelyAbsent.permits_reupload());
+        assert!(!ProbeOutcome::Indeterminate.permits_reupload());
+    }
+}
@@ -34,6 +34,38 @@ use gitlawb_core::http_sig::{
 };
 use gitlawb_core::identity::verify;

+/// Per-route-group ceiling on how much request body [`require_signature`] will
+/// buffer BEFORE it has authenticated anything.
+///
+/// This is middleware state rather than a constant because a single number cannot
+/// serve both a 2 GB git push and a 2 MB profile update: the git write group
+/// passes `GITLAWB_MAX_PACK_BYTES`, every other signed group passes
+/// [`crate::server::SIGNED_BODY_LIMIT`]. Making it a required piece of state means
+/// a new signed route group cannot be wired up without choosing a limit.
+#[derive(Clone, Copy, Debug)]
+pub struct SignatureBodyLimit(pub usize);
+
+/// The hint for an unresolvable DID, chosen by method. The hint is only true
+/// for a method we have no resolver for: a did:key that parsed and then failed
+/// on its key material (a small-order key, wrong multicodec, wrong length)
+/// would be sent looking for the wrong problem by the alpha-support wording.
+fn unresolvable_did_hint(key_id: &Did) -> &'static str {
+    if key_id.is_did_key() {
+        "the DID is a did:key whose key material did not resolve"
+    } else {
+        "only did:key is supported in alpha"
+    }
+}
+
+/// Bound the echoed keyid in the unresolvable-DID message. `key_id` is
+/// unauthenticated request input, so reflecting it back at whatever length the
+/// caller chose would amplify attacker input into the response body. The bound
+/// is slack (a did:key method-id is a fixed 48 characters) and only defends
+/// the echo, not the resolution.
+fn bounded_did_echo(key_id: &str) -> String {
+    key_id.chars().take(96).collect()
+}
+
 /// Axum middleware that enforces HTTP Signature authentication (RFC 9421).
 ///
 /// Every write request must carry:
@@ -48,20 +80,52 @@ use gitlawb_core::identity::verify;
 ///   4. Resolves the did:key to an Ed25519 VerifyingKey
 ///   5. Rebuilds the signing string and verifies the Ed25519 signature
 ///   6. Verifies Content-Digest matches the request body
-pub async fn require_signature(request: Request, next: Next) -> Response {
-    // Buffer the body so we can verify content-digest and pass it downstream
+pub async fn require_signature(
+    State(SignatureBodyLimit(limit)): State<SignatureBodyLimit>,
+    request: Request,
+    next: Next,
+) -> Response {
+    verify_signature(limit, request, next).await
+}
+
+/// The signature check itself, shared by [`require_signature`] and the optional
+/// variant. `limit` bounds the pre-auth buffer; see [`SignatureBodyLimit`].
+async fn verify_signature(limit: usize, request: Request, next: Next) -> Response {
+    // Buffer the body so we can verify content-digest and pass it downstream.
+    //
+    // The buffer is bounded, because this runs BEFORE any identity check: an
+    // unauthenticated caller with no DID, no signature and no repo reaches this
+    // line. Nothing upstream supplies the bound. On the git write routes
+    // `RequestBodyLimitLayer` sits on the inner router, and a later `.layer()` is
+    // the OUTER service, so this middleware runs first and sees the raw body. On
+    // every other signed group axum's 2 MB `DefaultBodyLimit` is enforced by the
+    // `Bytes`/`Json` extractors through a request extension, which collecting the
+    // body here bypasses. `Limited` stops reading at the limit rather than reading
+    // everything and then measuring it, which is the property that matters.
    let (parts, body) = request.into_parts();
-    let body_bytes =
-        match body.collect().await {
-            Ok(collected) => collected.to_bytes(),
-            Err(_) => return (
+    let body_bytes = match http_body_util::Limited::new(body, limit).collect().await {
+        Ok(collected) => collected.to_bytes(),
+        // Over the limit: a distinct status from the 401 an unsigned request gets
+        // and the 400 below, so a client can tell "too big" from "not authenticated"
+        // and from "your body did not arrive".
+        Err(e) if e.is::<http_body_util::LengthLimitError>() => return (
+            StatusCode::PAYLOAD_TOO_LARGE,
+            Json(json!({
+                "error": "body_too_large",
+                "message": format!("request body exceeds the {limit}-byte limit for this route"),
+            })),
+        )
+            .into_response(),
+        Err(_) => {
+            return (
                StatusCode::BAD_REQUEST,
                Json(
                    json!({ "error": "unreadable_body", "message": "could not read request body" }),
                ),
            )
-                .into_response(),
-        };
+                .into_response()
+        }
+    };

    let sig_input = parts
        .headers
@@ -120,7 +184,7 @@ pub async fn require_signature(request: Request, next: Next) -> Response {
                    COVERED_COMPONENTS.join(", "),
                    missing.join(", ")
                ),
-                "hint": "See https://gitlawb.com/agents#authentication",
+                "hint": "See https://twigpine.com/agents#authentication",
            })),
        )
            .into_response();
@@ -145,11 +209,14 @@ pub async fn require_signature(request: Request, next: Next) -> Response {
                StatusCode::BAD_REQUEST,
                Json(json!({
                    "error": "unresolvable_did",
-                    "message": format!("cannot resolve DID '{}': {e}", sig.key_id),
-                    "hint": "only did:key is supported in alpha",
+                    "message": format!(
+                        "cannot resolve DID '{}': {e}",
+                        bounded_did_echo(&sig.key_id.to_string())
+                    ),
+                    "hint": unresolvable_did_hint(&sig.key_id),
                })),
            )
-                .into_response()
+                .into_response();
        }
    };

@@ -162,17 +229,34 @@ pub async fn require_signature(request: Request, next: Next) -> Response {
        .unwrap_or("/")
        .to_string();

-    let content_digest = parts
+    // Required, never defaulted. The signing string is rebuilt from the request,
+    // so an absent header would make the covered `content-digest` empty on both
+    // sides: verification would pass and the digest-versus-body comparison below
+    // would have nothing to compare, leaving a signed request free to carry a
+    // body its signature never covered. `sign_request` always emits the header,
+    // so refusing here costs no conforming client anything.
+    let content_digest = match parts
        .headers
        .get("content-digest")
        .and_then(|v| v.to_str().ok())
-        .unwrap_or("")
-        .to_string();
+    {
+        Some(v) => v.to_string(),
+        None => {
+            return (
+                StatusCode::BAD_REQUEST,
+                Json(json!({
+                    "error": "missing_content_digest",
+                    "message": "Content-Digest header is required on a signed request",
+                })),
+            )
+                .into_response()
+        }
+    };

    let mut request_values: HashMap<String, String> = HashMap::new();
    request_values.insert("@method".to_string(), method);
    request_values.insert("@path".to_string(), path_and_query);
-    request_values.insert("content-digest".to_string(), content_digest);
+    request_values.insert("content-digest".to_string(), content_digest.clone());

    // The @signature-params value is the part of Signature-Input after "sig1="
    let sig_params_value = sig_input.strip_prefix("sig1=").unwrap_or(&sig_input);
@@ -217,23 +301,18 @@ pub async fn require_signature(request: Request, next: Next) -> Response {
            .into_response();
    }

-    // Verify Content-Digest matches the actual request body
-    if let Some(claimed) = parts
-        .headers
-        .get("content-digest")
-        .and_then(|v| v.to_str().ok())
-    {
-        let actual = compute_content_digest(&body_bytes);
-        if claimed != actual {
-            return (
-                StatusCode::BAD_REQUEST,
-                Json(json!({
-                    "error": "content_digest_mismatch",
-                    "message": "Content-Digest does not match request body",
-                })),
-            )
-                .into_response();
-        }
+    // Verify Content-Digest matches the actual request body. Unconditional: the
+    // header's presence was established above, so there is no branch here that
+    // skips the comparison.
+    if content_digest != compute_content_digest(&body_bytes) {
+        return (
+            StatusCode::BAD_REQUEST,
+            Json(json!({
+                "error": "content_digest_mismatch",
+                "message": "Content-Digest does not match request body",
+            })),
+        )
+            .into_response();
    }

    tracing::info!(did = %sig.key_id, "✓ authenticated request");
@@ -248,11 +327,17 @@ pub async fn require_signature(request: Request, next: Next) -> Response {
 /// Optional variant for rolling upgrades: verify and inject `AuthenticatedDid` when
 /// RFC 9421 signature headers are present, but allow legacy unsigned requests to
 /// continue when no signature attempt was made.
-pub async fn optional_signature(request: Request, next: Next) -> Response {
+pub async fn optional_signature(
+    State(SignatureBodyLimit(limit)): State<SignatureBodyLimit>,
+    request: Request,
+    next: Next,
+) -> Response {
    let has_signature_headers = request.headers().contains_key("signature-input")
        || request.headers().contains_key("signature");
    if has_signature_headers {
-        return require_signature(request, next).await;
+        // Same pre-auth buffer, same bound: presenting signature headers is not
+        // authentication, so this path is as reachable as the required one.
+        return verify_signature(limit, request, next).await;
    }
    next.run(request).await
 }
@@ -380,7 +465,7 @@ fn human_detected(message: &str) -> impl IntoResponse {
            "error": "not_an_agent",
            "message": message,
            "hint": "gl identity new && gl register",
-            "docs": "https://gitlawb.com/agents",
+            "docs": "https://twigpine.com/agents",
        })),
    )
 }
@@ -423,6 +508,58 @@ mod tests {
        assert!(validate_ucan_chain(&token, &node_did, &agent_did).is_ok());
    }

+    /// The two-branch hint selection is the point of the unresolvable-DID
+    /// change: a did:key that failed on its key material must be told the key
+    /// material did not resolve, and any other method must be told only did:key
+    /// is supported. A regression that drops the is_did_key() check would send
+    /// a did:key caller after the wrong problem, so both branches are pinned.
+    #[test]
+    fn unresolvable_did_hint_distinguishes_key_material_from_method() {
+        // A real did:key whose key material fails to resolve (small-order):
+        // key-material hint.
+        let mut weak = [0u8; 32];
+        weak[0] = 1; // compressed identity point
+        let weak_vk = ed25519_dalek::VerifyingKey::from_bytes(&weak).unwrap();
+        let weak_did = Did::from_verifying_key(&weak_vk);
+        assert!(
+            weak_did.to_verifying_key().is_err(),
+            "fixture precondition: small-order did:key must not resolve"
+        );
+        assert_eq!(
+            unresolvable_did_hint(&weak_did),
+            "the DID is a did:key whose key material did not resolve"
+        );
+
+        // A method with no resolver: alpha-support hint.
+        let web = Did::web("example.com");
+        assert!(
+            web.to_verifying_key().is_err(),
+            "fixture precondition: did:web must not resolve locally"
+        );
+        assert_eq!(
+            unresolvable_did_hint(&web),
+            "only did:key is supported in alpha"
+        );
+    }
+
+    /// The unresolvable-DID message echoes the keyid, which is unauthenticated
+    /// request input. The echoed value must be bounded so a hostile caller
+    /// cannot amplify arbitrary-length input into the response body.
+    #[test]
+    fn unresolvable_did_message_bounds_the_echoed_keyid() {
+        let long_keyid = format!("did:key:z{}", "A".repeat(2000));
+        let shown = bounded_did_echo(&long_keyid);
+        assert!(
+            shown.len() <= 96,
+            "echoed keyid must be bounded, got {} chars",
+            shown.len()
+        );
+        assert!(
+            !shown.contains(&"A".repeat(1000)),
+            "the full keyid must not survive into the message"
+        );
+    }
+
    #[test]
    fn validate_ucan_chain_wrong_issuer() {
        let node = Keypair::generate();
@@ -490,16 +627,19 @@ mod tests {
        use clap::Parser;

        let keypair = Keypair::generate();
+        let scan_token_key = crate::state::AppState::derive_scan_token_key(&keypair);
        let (ref_tx, _) = tokio::sync::broadcast::channel(1);
        let (task_tx, _) = tokio::sync::broadcast::channel(1);
        let pool = sqlx::postgres::PgPoolOptions::new()
            .connect_lazy("postgres://localhost/gitlawb_test_placeholder")
            .expect("lazy pool creation should not fail");
        let db = Arc::new(crate::db::Db::for_testing(pool.clone()));
+        let task_cursor_key = crate::api::task_cursor::TaskCursorKey::derive(&keypair.to_seed());
        let schema = Arc::new(graphql::build_schema(
            db.clone(),
            ref_tx.clone(),
            task_tx.clone(),
+            task_cursor_key.clone(),
        ));
        crate::state::AppState {
            config: Arc::new(Config::parse_from(["gitlawb-node"])),
@@ -511,14 +651,27 @@ mod tests {
            ref_update_tx: ref_tx,
            task_event_tx: task_tx,
            graphql_schema: schema,
+            task_cursor_key,
            machine_id: None,
            repo_store: crate::git::repo_store::RepoStore::for_testing(PathBuf::from("/tmp"), pool),
            rate_limiter: RateLimiter::new(100, Duration::from_secs(60)),
            create_ip_rate_limiter: RateLimiter::new(1000, Duration::from_secs(3600)),
            push_rate_limiter: RateLimiter::new(600, Duration::from_secs(3600)),
+            close_issue_rate_limiter: RateLimiter::new(120, Duration::from_secs(3600)),
+            ipfs_rate_limiter: RateLimiter::new(600, Duration::from_secs(3600)),
+            ipfs_work_rate_limiter: RateLimiter::new(600, Duration::from_secs(3600)),
+            ipfs_max_history_walks: crate::api::ipfs::MAX_HISTORY_WALKS_PER_REQUEST,
+            ipfs_max_legacy_probes: crate::api::ipfs::MAX_LEGACY_PROBES_PER_REQUEST,
+            ipfs_legacy_scan_page_rows: crate::api::ipfs::LEGACY_SCAN_PAGE_ROWS,
+            ipfs_max_legacy_scan_rows: crate::api::ipfs::MAX_LEGACY_SCAN_ROWS_PER_REQUEST,
+            ipfs_max_legacy_scan_rule_bytes:
+                crate::api::ipfs::MAX_LEGACY_SCAN_RULE_BYTES_PER_REQUEST,
+            ipfs_scan_token_key: Arc::new(scan_token_key),
+            ipfs_max_served_object_bytes: crate::api::ipfs::MAX_SERVED_OBJECT_BYTES,
            push_limiter_trust: crate::rate_limit::TrustedProxy::None,
            sync_trigger_rate_limiter: RateLimiter::new(60, Duration::from_secs(3600)),
            peer_write_rate_limiter: RateLimiter::new(600, Duration::from_secs(3600)),
+            federated_rate_limiter: RateLimiter::new_bounded(12, Duration::from_secs(60), 10_000),
            shutdown_tx: tokio::sync::watch::channel(false).0,
            git_read_semaphore: Arc::new(tokio::sync::Semaphore::new(64)),
            git_write_semaphore: Arc::new(tokio::sync::Semaphore::new(64)),
@@ -534,7 +687,7 @@ mod tests {
            git_ipfs_walk_semaphore: Arc::new(tokio::sync::Semaphore::new(64)),
            git_ipfs_walk_per_caller:
                crate::rate_limit::PerCallerConcurrency::with_default_max_keys(16),
-            ipfs_rate_limiter: RateLimiter::new(600, Duration::from_secs(3600)),
+            task_read_rate_limiter: RateLimiter::new(1200, Duration::from_secs(3600)),
            git_bin: "git".to_string(),
        }
    }
@@ -626,4 +779,78 @@ mod tests {
        let body_json: serde_json::Value = serde_json::from_slice(&body_bytes).unwrap();
        assert_eq!(body_json["error"], "invalid_ucan");
    }
+
+    /// A request that signs `content-digest` as the EMPTY string and then sends
+    /// no Content-Digest header at all must be refused before any handler runs.
+    ///
+    /// This is the whole point of requiring the header rather than defaulting it:
+    /// the middleware rebuilds the signing string from the request, so an absent
+    /// header makes the covered digest empty on both sides. The Ed25519 check
+    /// passes, the body-versus-digest comparison is skipped for want of a header,
+    /// and a signed request carries a body the signature never covered.
+    #[tokio::test]
+    async fn signed_request_without_content_digest_is_rejected() {
+        use base64::{engine::general_purpose::STANDARD, Engine};
+        use std::sync::atomic::{AtomicBool, Ordering};
+
+        static REACHED: AtomicBool = AtomicBool::new(false);
+
+        let kp = Keypair::generate();
+        let did = kp.did();
+        let created = chrono::Utc::now().timestamp();
+        let signature_input = format!(
+            r#"sig1=("@method" "@path" "content-digest");keyid="{did}";alg="ed25519";created={created}"#
+        );
+        let sig_params_value = &signature_input["sig1=".len()..];
+
+        let mut values: HashMap<String, String> = HashMap::new();
+        values.insert("@method".to_string(), "POST".to_string());
+        values.insert("@path".to_string(), "/x".to_string());
+        // Exactly what the middleware derives when the header is absent.
+        values.insert("content-digest".to_string(), String::new());
+        let signing_string =
+            build_signing_string(COVERED_COMPONENTS, sig_params_value, &values).unwrap();
+        let signature = format!(
+            "sig1=:{}:",
+            STANDARD.encode(kp.sign(signing_string.as_bytes()).to_bytes())
+        );
+
+        let app = Router::new()
+            .route(
+                "/x",
+                axum::routing::post(|| async {
+                    REACHED.store(true, Ordering::SeqCst);
+                    StatusCode::OK
+                }),
+            )
+            .layer(middleware::from_fn_with_state(
+                SignatureBodyLimit(crate::server::SIGNED_BODY_LIMIT),
+                require_signature,
+            ));
+
+        let req = Request::builder()
+            .method("POST")
+            .uri("/x")
+            .header("signature-input", signature_input)
+            .header("signature", signature)
+            .body(Body::from("a body no signature covers"))
+            .unwrap();
+
+        let resp = app.oneshot(req).await.unwrap();
+        assert_eq!(
+            resp.status(),
+            StatusCode::BAD_REQUEST,
+            "a signed request with no Content-Digest header must be refused"
+        );
+        let bytes = axum::body::to_bytes(resp.into_body(), 2048).await.unwrap();
+        let body: serde_json::Value = serde_json::from_slice(&bytes).unwrap();
+        assert_eq!(
+            body["error"], "missing_content_digest",
+            "the refusal must name the absent Content-Digest header"
+        );
+        assert!(
+            !REACHED.load(Ordering::SeqCst),
+            "the handler must never see a request whose body no signature covers"
+        );
+    }
 }
@@ -1,4 +1,4 @@
-//! Embedded seed list of public Gitlawb network nodes.
+//! Embedded seed list of public Twigpine network nodes.
 //!
 //! This module parses `bootstrap-peers.json` (embedded at compile time) and
 //! merges its contents into the runtime config so a fresh `docker compose up`
@@ -11,23 +11,34 @@ use uuid::Uuid;
 use crate::db::RefCertificate;
 use crate::state::AppState;

-/// Issue a signed ref-update certificate for a successful push.
+/// Issue a signed ref-update certificate for a successful push. The
+/// live receive-pack handler calls this on every successful push.
+/// Build the v1 signing payload — the EXACT bytes the node signs and
+/// gl's `verify_signature` reconstructs. This is the single source
+/// of truth shared by `issue_ref_certificate` and the frozen
+/// canonical-form tests below; drift here is what would break every
+/// existing cert.
 ///
-/// Builds a canonical JSON payload, signs it with the node's Ed25519 key,
-/// persists the certificate, and returns it.
-pub async fn issue_ref_certificate(
-    state: &AppState,
+/// Field order and key set are fixed: a default `serde_json::Value`
+/// serializes `Map<String, Value>` with sorted keys, so any change to
+/// the literal is observable as a different byte sequence on the
+/// wire.
+///
+/// v1 is the unversioned legacy shape: no `version` key in the signed
+/// JSON. From v2 onward the version belongs INSIDE the signed bytes
+/// (see `v2_signing_payload`): leaving it as an unsigned sibling
+/// would let a future v2 cert be downgraded to v1 and still verify
+/// cleanly, since the bytes would be identical.
+pub(crate) fn v1_signing_payload(
    repo_id: &str,
    ref_name: &str,
    old_sha: &str,
    new_sha: &str,
    pusher_did: &str,
-) -> Result<RefCertificate> {
-    let node_did = state.node_did.to_string();
-    let issued_at = Utc::now().to_rfc3339();
-
-    // Build the canonical signing payload.
-    let payload = serde_json::json!({
+    node_did: &str,
+    issued_at: &str,
+) -> serde_json::Value {
+    serde_json::json!({
        "repo_id": repo_id,
        "ref":     ref_name,
        "old":     old_sha,
@@ -35,13 +46,180 @@ pub async fn issue_ref_certificate(
        "pusher":  pusher_did,
        "node":    node_did,
        "ts":      issued_at,
-    });
+    })
+}
+
+/// Build the v2 signing payload — the future shape for versioned
+/// certs. Identical to v1 except the wire `version` is part of the
+/// signed bytes, so stripping or flipping the `version` column
+/// invalidates the signature instead of verifying cleanly under the
+/// other version's path.
+///
+/// NOT yet issued: `issue_ref_certificate` still stamps v1 because
+/// the shipped `gl` verifier only supports v1. This builder exists
+/// to pin the downgrade-resistant shape now, before any v2 cert is
+/// ever signed.
+#[allow(dead_code)]
+pub(crate) fn v2_signing_payload(
+    repo_id: &str,
+    ref_name: &str,
+    old_sha: &str,
+    new_sha: &str,
+    pusher_did: &str,
+    node_did: &str,
+    issued_at: &str,
+) -> serde_json::Value {
+    serde_json::json!({
+        "repo_id": repo_id,
+        "ref":     ref_name,
+        "old":     old_sha,
+        "new":     new_sha,
+        "pusher":  pusher_did,
+        "node":    node_did,
+        "ts":      issued_at,
+        "version": 2,
+    })
+}
+
+/// Issue a signed ref-update certificate for a successful push.
+///
+/// `cert_id` is the deterministic id derived from `(request_id,
+/// ref_name)` (see [`crate::db::ref_cert_id_for`]). It is required so
+/// the recovery drain and the live handler produce the same primary
+/// key: a live push followed by a recovery pass collapses to a
+/// single cert row, and a re-push to the same `(repo, ref)` updates
+/// the existing row's `old_sha` / `new_sha` / `pusher_did` /
+/// `issued_at` / `signature` to the new transition while preserving
+/// the original `id` (the `insert_ref_certificate` upsert is
+/// keyed on `(repo_id, ref_name)` and only updates fields when the
+/// new `issued_at` is strictly greater).
+///
+/// #26 Split PR 1 P1-B: the live handler routes through this
+/// function (the upsert). After the reviewer-1 round-2 fix, the
+/// recovery drain also routes through this function (P1: refresh a
+/// stale cert), so both paths use the same deterministic `cert_id`
+/// and the same upsert. A re-pass is always safe:
+///
+/// - Live handler → live upsert: re-push updates the row, preserves
+///   the original `id`. The contract pinned by
+///   `insert_ref_certificate_upserts_on_repo_ref` is restored.
+/// - Live handler → recovery: live's `ON CONFLICT (id) DO UPDATE`
+///   preserves the original `id`; the recovery's same upsert is
+///   a no-op for an equal-`issued_at` re-run and a refresh for a
+///   strictly-newer one.
+/// - Recovery → live handler: the recovery wrote a row with the
+///   deterministic `id`; the live upsert (which preserves `id` and
+///   only updates other fields when `issued_at` is strictly newer)
+///   is a no-op for an equal-`issued_at` re-run and a refresh for
+///   a strictly-newer one.
+#[allow(dead_code)] // round-trip test in db/mod.rs pins the upsert contract; the live path and the drain use issue_ref_certificate_with_issued_at
+pub async fn issue_ref_certificate(
+    state: &AppState,
+    repo_id: &str,
+    ref_name: &str,
+    old_sha: &str,
+    new_sha: &str,
+    pusher_did: &str,
+    cert_id: &str,
+) -> Result<RefCertificate> {
+    issue_ref_certificate_with_issued_at(
+        state, repo_id, ref_name, old_sha, new_sha, pusher_did, cert_id, None,
+    )
+    .await
+}
+
+/// #26 Split PR 1 round 4 — variant that lets the caller stamp the
+/// cert's `issued_at` with a transition-time timestamp instead of
+/// `Utc::now()`. The recovery drain passes the persisted
+/// `row.created_at` so a replay after a later live cert does not
+/// outrank the live cert in the `EXCLUDED.issued_at >
+/// ref_certificates.issued_at` upsert guard.
+///
+/// The live handler uses the default `issue_ref_certificate` (no
+/// override), which keeps `Utc::now()` — the reviewer's invariant
+/// is that `issued_at` reflects the transition time, and for a
+/// live push the transition time and the wall-clock are the same.
+///
+/// `issued_at_override` is honored verbatim; passing a value not in
+/// RFC 3339 form is a logic bug (the upsert will mis-order), so
+/// callers must use the row's persisted `created_at`.
+///
+/// # Clippy allow — too many arguments
+/// This is the explicit "stamp a transition-time `issued_at`"
+/// variant of `issue_ref_certificate`. The drain
+/// (`durable_outbox::derive_one`) is the in-crate caller; the
+/// test `replay_of_stale_row_does_not_overwrite_live_cert_b` pins
+/// the contract that a recovery replay's `issued_at` does NOT
+/// outrank a later live cert. Adding a struct-arg would be a
+/// larger refactor for two callers (live + drain) and obscure the
+/// parallel to `issue_ref_certificate` (which is `#[allow]`'d for
+/// the same reason historically).
+#[allow(clippy::too_many_arguments)]
+pub async fn issue_ref_certificate_with_issued_at(
+    state: &AppState,
+    repo_id: &str,
+    ref_name: &str,
+    old_sha: &str,
+    new_sha: &str,
+    pusher_did: &str,
+    cert_id: &str,
+    issued_at_override: Option<String>,
+) -> Result<RefCertificate> {
+    let cert = build_ref_certificate(
+        state,
+        repo_id,
+        ref_name,
+        old_sha,
+        new_sha,
+        pusher_did,
+        Some(cert_id.to_string()),
+        issued_at_override,
+    )
+    .await?;
+    state.db.insert_ref_certificate(&cert).await
+}
+
+/// Shared cert construction: build the JSON payload, sign it with the
+/// node key, and assemble the `RefCertificate` row. `cert_id_override`
+/// lets the recovery path plug in a deterministic id; the live path
+/// passes `None` and gets a fresh UUID. `issued_at_override` lets
+/// the recovery path stamp the cert with the original transition
+/// time so the upsert's `issued_at > issued_at` guard correctly
+/// orders transitions regardless of write order.
+#[allow(clippy::too_many_arguments)]
+async fn build_ref_certificate(
+    state: &AppState,
+    repo_id: &str,
+    ref_name: &str,
+    old_sha: &str,
+    new_sha: &str,
+    pusher_did: &str,
+    cert_id_override: Option<String>,
+    issued_at_override: Option<String>,
+) -> Result<RefCertificate> {
+    let node_did = state.node_did.to_string();
+    // P1 (reviewer-1 round 4): when the caller passes a transition-
+    // time `issued_at` (the recovery drain passes `row.created_at`),
+    // use it verbatim so the upsert's per-column guard
+    // `EXCLUDED.issued_at > ref_certificates.issued_at` correctly
+    // orders transitions regardless of write order. The live handler
+    // passes `None` and gets `Utc::now()` — for a live push the
+    // transition time and the wall-clock are the same.
+    let issued_at = issued_at_override.unwrap_or_else(|| Utc::now().to_rfc3339());
+
+    // Build the canonical signing payload via the shared builder so
+    // the frozen-vector test below and the live signer cannot drift
+    // apart — a regression in either side fails both.
+    let payload = v1_signing_payload(
+        repo_id, ref_name, old_sha, new_sha, pusher_did, &node_did, &issued_at,
+    );
    let payload_bytes = serde_json::to_vec(&payload)?;

    let signature = state.node_keypair.sign_b64(&payload_bytes);

+    let id = cert_id_override.unwrap_or_else(|| Uuid::new_v4().to_string());
    let cert = RefCertificate {
-        id: Uuid::new_v4().to_string(),
+        id,
        repo_id: repo_id.to_string(),
        ref_name: ref_name.to_string(),
        old_sha: old_sha.to_string(),
@@ -50,9 +228,353 @@ pub async fn issue_ref_certificate(
        node_did,
        signature,
        issued_at,
+        // #26 Split PR 3: the wire-format version. New certs stamp
+        // v1 — the pre-versioning 7-field payload with no `version`
+        // key in the signed JSON — because the shipped `gl`
+        // verifier only supports v1. Stamping v2 here would issue
+        // certificates the shipped client refuses (round-3
+        // finding: node pinned `version == 2` while gl pinned
+        // "2 is refused", with nothing exercising the
+        // composition). v2 is reserved for the future shape
+        // defined by `v2_signing_payload`, which carries the
+        // version INSIDE the signed bytes; land the v2 verify
+        // path before stamping v2.
+        version: 1,
    };

    // Persist and return the row as it exists in the database (on a
    // conflict the existing row survives when it is newer).
    state.db.insert_ref_certificate(&cert).await
 }
+
+#[cfg(test)]
+mod v1_payload_tests {
+    //! #26 Split PR 3 — cert payload version compat.
+    //!
+    //! The v1 cert payload is the frozen canonical byte form the
+    //! node signs. gl's `verify_signature` reconstructs the same
+    //! payload and verifies. Any drift in field order, key set, or
+    //! whitespace breaks every existing cert. The tests here pin
+    //! the v1 payload shape, the round-trip, the cert id
+    //! determinism per the v1 version, and the v2 read forward
+    //! compat (an unknown version refuses to verify rather than
+    //! guessing).
+    //!
+    //! Reviewer 1 finding: the live signer (`issue_ref_certificate`)
+    //! and the frozen-vector test (`v1_payload_matches_frozen_canonical_form`)
+    //! previously held two separate `serde_json::json!` literals,
+    //! so a regression in the signer could not fail this test.
+    //! Both now go through the shared `v1_signing_payload` builder.
+    use super::{v1_signing_payload, v2_signing_payload};
+    use crate::db::RefCertificate;
+    use gitlawb_core::identity::Keypair;
+
+    /// v1 payload is the same shape the legacy code signed, byte
+    /// for byte. Reverting this assertion to the v2 shape (which
+    /// includes a `version` key) is what the versioned format
+    /// forbids: an old client reading a v1 cert must not see a
+    /// `version` key in the signed JSON.
+    ///
+    /// The literal is built via the shared `v1_signing_payload`
+    /// helper that `issue_ref_certificate` also uses, so the live
+    /// signer and the frozen vector share one source of truth.
+    #[test]
+    fn v1_payload_matches_frozen_canonical_form() {
+        let payload = v1_signing_payload(
+            "repo-1",
+            "refs/heads/main",
+            &"0".repeat(40),
+            &"a".repeat(40),
+            "did:key:z6MkPusher",
+            "did:key:z6MkNode",
+            "2026-07-22T00:00:00+00:00",
+        );
+        let frozen = format!(
+            r#"{{"new":"{}","node":"did:key:z6MkNode","old":"{}","pusher":"did:key:z6MkPusher","ref":"refs/heads/main","repo_id":"repo-1","ts":"2026-07-22T00:00:00+00:00"}}"#,
+            "a".repeat(40),
+            "0".repeat(40),
+        );
+        assert_eq!(serde_json::to_string(&payload).unwrap(), frozen);
+    }
+
+    /// Round-trip: sign the v1 payload with a keypair, build a
+    /// RefCertificate with version: 1, and verify the structure.
+    /// This is the construction the live code does on every push.
+    ///
+    /// Uses the shared builder for both halves so a regression in
+    /// the live signer (adding/removing/renaming a field) cannot
+    /// pass this test by accident.
+    #[test]
+    fn v1_ref_certificate_structure_is_well_formed() {
+        let kp = Keypair::generate();
+        let node_did = kp.did().as_str().to_string();
+        let payload = v1_signing_payload(
+            "repo-1",
+            "refs/heads/main",
+            &"0".repeat(40),
+            &"a".repeat(40),
+            "did:key:z6MkPusher",
+            &node_did,
+            "2026-07-22T00:00:00+00:00",
+        );
+        let payload_bytes = serde_json::to_vec(&payload).unwrap();
+        let sig = kp.sign_b64(&payload_bytes);
+        let cert = RefCertificate {
+            id: "cert-id".into(),
+            repo_id: "repo-1".into(),
+            ref_name: "refs/heads/main".into(),
+            old_sha: "0".repeat(40),
+            new_sha: "a".repeat(40),
+            pusher_did: "did:key:z6MkPusher".into(),
+            node_did: node_did.clone(),
+            signature: sig.clone(),
+            issued_at: "2026-07-22T00:00:00+00:00".into(),
+            version: 1,
+        };
+        assert_eq!(cert.version, 1, "new certs carry version: 1");
+        // The signed payload reconstructs identically: gl's
+        // verify_signature would build the same JSON, hash the
+        // same bytes, and verify the same signature.
+        let reconstructed = v1_signing_payload(
+            &cert.repo_id,
+            &cert.ref_name,
+            &cert.old_sha,
+            &cert.new_sha,
+            &cert.pusher_did,
+            &cert.node_did,
+            &cert.issued_at,
+        );
+        let reconstructed_bytes = serde_json::to_vec(&reconstructed).unwrap();
+        assert_eq!(
+            reconstructed_bytes, payload_bytes,
+            "the round-trip serialization must be byte-identical; \
+             this is what makes gl's verify_signature succeed"
+        );
+        // Round 3 (P3 reviewer): byte-equality alone never calls
+        // `identity::verify`, so a bad signature helper would pass.
+        // Verify the computed signature against the reconstructed
+        // bytes — the same check gl performs.
+        let vk = node_did
+            .parse::<gitlawb_core::did::Did>()
+            .unwrap()
+            .to_verifying_key()
+            .unwrap();
+        use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine};
+        let sig_bytes: [u8; 64] = URL_SAFE_NO_PAD
+            .decode(&cert.signature)
+            .unwrap()
+            .try_into()
+            .unwrap();
+        gitlawb_core::identity::verify(&vk, &reconstructed_bytes, &sig_bytes)
+            .expect("well-formed v1 cert signature must verify");
+    }
+
+    /// Reviewer 1 finding: the live signer must share the
+    /// canonical-form literal so a regression in either side
+    /// fails both. This test pins the shared builder's output
+    /// against the literal a future v2 byte stream would have
+    /// to break — if someone adds a `version` key to
+    /// `v1_signing_payload`, every existing cert breaks verify,
+    /// and this test is one of the canaries that fires.
+    #[test]
+    fn v1_signing_payload_has_no_version_key() {
+        let payload = v1_signing_payload(
+            "repo-1",
+            "refs/heads/main",
+            "oldsha",
+            "newsha",
+            "did:key:z6MkPusher",
+            "did:key:z6MkNode",
+            "2026-07-22T00:00:00+00:00",
+        );
+        let obj = payload.as_object().expect("payload is a JSON object");
+        assert!(
+            !obj.contains_key("version"),
+            "v1 signing payload must not carry a `version` key — adding \
+             one changes the bytes and breaks every existing cert's \
+             signature. v2 certs build a different payload, not this one"
+        );
+        // Pin the exact set so a future field addition is caught.
+        let expected: std::collections::BTreeSet<&str> =
+            ["new", "node", "old", "pusher", "ref", "repo_id", "ts"]
+                .into_iter()
+                .collect();
+        let actual: std::collections::BTreeSet<&str> = obj.keys().map(String::as_str).collect();
+        assert_eq!(actual, expected, "v1 payload key set is frozen");
+    }
+
+    /// Round 3 (P2 reviewer): from v2 onward the version belongs
+    /// INSIDE the signed bytes. v1 stays the unversioned legacy
+    /// shape; v2 carries `"version": 2` in the payload so a
+    /// version flip without re-signing invalidates the signature
+    /// instead of verifying cleanly under the other path.
+    #[test]
+    fn v2_signing_payload_binds_version_and_resists_downgrade() {
+        let kp = Keypair::generate();
+        let node_did = kp.did().as_str().to_string();
+        let ts = "2026-07-22T00:00:00+00:00";
+        let v1 = v1_signing_payload(
+            "repo-1",
+            "refs/heads/main",
+            "oldsha",
+            "newsha",
+            "did:key:z6MkPusher",
+            &node_did,
+            ts,
+        );
+        let v2 = v2_signing_payload(
+            "repo-1",
+            "refs/heads/main",
+            "oldsha",
+            "newsha",
+            "did:key:z6MkPusher",
+            &node_did,
+            ts,
+        );
+        let v2_obj = v2.as_object().expect("v2 payload is a JSON object");
+        assert_eq!(
+            v2_obj.get("version"),
+            Some(&serde_json::json!(2)),
+            "v2 payload must carry the version inside the signed bytes"
+        );
+        let v1_bytes = serde_json::to_vec(&v1).unwrap();
+        let v2_bytes = serde_json::to_vec(&v2).unwrap();
+        assert_ne!(
+            v1_bytes, v2_bytes,
+            "v1 and v2 payloads must differ, otherwise a version flip is cryptographically invisible"
+        );
+
+        // A signature over v1 must NOT verify as v2 and vice versa.
+        let vk = node_did
+            .parse::<gitlawb_core::did::Did>()
+            .unwrap()
+            .to_verifying_key()
+            .unwrap();
+        use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine};
+        let sig_v1 = kp.sign_b64(&v1_bytes);
+        let sig_v1_bytes: [u8; 64] = URL_SAFE_NO_PAD.decode(&sig_v1).unwrap().try_into().unwrap();
+        assert!(
+            gitlawb_core::identity::verify(&vk, &v1_bytes, &sig_v1_bytes).is_ok(),
+            "v1 signature must verify under v1 bytes"
+        );
+        assert!(
+            gitlawb_core::identity::verify(&vk, &v2_bytes, &sig_v1_bytes).is_err(),
+            "a v1 signature presented as v2 (downgrade/upgrade flip) must not verify"
+        );
+    }
+
+    /// The live issuer stamps the version the shipped client can
+    /// verify. The gl client refuses to verify v2 certs explicitly;
+    /// this test pins that the version the ISSUING PATH claims
+    /// agrees with the payload shape it actually signs.
+    ///
+    /// Round 2: the prior form built a `RefCertificate` literal with
+    /// `version: 1` and asserted `version == 1` — true no matter what the
+    /// signer does, so flipping the issuer's version (or its payload shape)
+    /// left it green. This one drives the actual
+    /// `issue_ref_certificate` so a regression in the issuer flips the test.
+    /// The DB row is created against `test_support::test_state`, which
+    /// gives us a real `AppState` with a real node keypair.
+    ///
+    /// Round 3: the issuer stamps v1 because the shipped `gl`
+    /// verifier only supports v1 (stamping v2 shipped certs the
+    /// client refuses). The binding to the live issuer is kept —
+    /// flipping the stamp still turns this test red — without
+    /// changing what production issues.
+    #[sqlx::test]
+    #[allow(clippy::async_yields_async)]
+    async fn issuer_stamps_v1_over_v1_payload(pool: sqlx::PgPool) {
+        let state = crate::test_support::test_state(pool.clone()).await;
+        let repo_id = "repo-issuer-observe";
+        let ref_name = "refs/heads/main";
+        let old = "0".repeat(40);
+        let new = "a".repeat(40);
+        let pusher = "did:key:z6MkPusher";
+        let cert_id = "cert-issuer-observe";
+
+        // Drive the live issuer and read the persisted row back by
+        // id, so the test covers the stored `version` column and
+        // not just the in-process return value.
+        let cert = crate::cert::issue_ref_certificate(
+            &state, repo_id, ref_name, &old, &new, pusher, cert_id,
+        )
+        .await
+        .expect("issue_ref_certificate must succeed");
+
+        // The cert returned by the issuer is the source of truth
+        // for both the version claim and the signed bytes.
+        assert_eq!(cert.repo_id, repo_id);
+        assert_eq!(cert.ref_name, ref_name);
+
+        // Cross-check 1: the issuer's claimed version must match
+        // what the live function stamps. This assertion is bound
+        // to the live function, not a hand-built literal:
+        // flipping the stamp in `issue_ref_certificate` breaks it
+        // through the actual call path.
+        assert_eq!(
+            cert.version, 1,
+            "the live issuer must stamp v1 until a v2 verifier ships; \
+             flipping the stamp breaks this assertion through the actual call path"
+        );
+
+        // Cross-check 1b: the persisted row carries the same stamp.
+        let stored = state
+            .db
+            .get_ref_certificate(&cert.id)
+            .await
+            .expect("persisted cert must be readable")
+            .expect("persisted cert must exist");
+        assert_eq!(
+            stored.version, cert.version,
+            "the stored version column must match the issued claim"
+        );
+
+        // Cross-check 2: the signature on the cert must verify
+        // against the v1 payload builder's bytes, not against some
+        // other shape the issuer might have introduced. This is
+        // what gl's `verify_signature` does for a v1 cert: rebuild
+        // the payload from the cert's own fields and verify the
+        // signature. If the issuer ever signs a different shape
+        // while still claiming v1, every shipped client breaks —
+        // and so does this.
+        let rebuilt = v1_signing_payload(
+            &cert.repo_id,
+            &cert.ref_name,
+            &cert.old_sha,
+            &cert.new_sha,
+            &cert.pusher_did,
+            &cert.node_did,
+            &cert.issued_at,
+        );
+        let vk = cert
+            .node_did
+            .parse::<gitlawb_core::did::Did>()
+            .unwrap()
+            .to_verifying_key()
+            .unwrap();
+        use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine};
+        let sig_bytes: [u8; 64] = URL_SAFE_NO_PAD
+            .decode(&cert.signature)
+            .unwrap()
+            .try_into()
+            .unwrap();
+        gitlawb_core::identity::verify(&vk, &serde_json::to_vec(&rebuilt).unwrap(), &sig_bytes)
+            .expect("a v1 cert's signature must verify against the v1 payload shape");
+
+        // Cross-check 3: the v1 signed payload must NOT contain a
+        // version key — v2 is the future shape that adds one inside
+        // the signed bytes (see `v2_signing_payload`). The v1
+        // payload (which the cert is signed over) is frozen at 7
+        // fields. A change that adds `version` to the v1 payload
+        // breaks every existing cert and this test.
+        let obj = rebuilt
+            .as_object()
+            .expect("rebuilt v1 payload is a JSON object");
+        assert!(
+            !obj.contains_key("version"),
+            "the v1 signing payload must not carry a version key — \
+             adding one changes the bytes and breaks every existing \
+             cert's signature"
+        );
+    }
+}
@@ -6,6 +6,7 @@
 use std::collections::{BTreeSet, HashMap};
 use std::path::Path;
 use std::str::FromStr;
+use std::time::Duration;

 use ed25519_dalek::VerifyingKey;
 use gitlawb_core::did::Did;
@@ -69,6 +70,12 @@ fn resolve_all_recipients(dids: &BTreeSet<String>) -> Result<Vec<VerifyingKey>,
 enum SealPlan {
    /// An existing envelope already covers exactly this recipient set; nothing to do.
    SkipUnchanged,
+    /// The stored tag matches, but the stored reader set contains a DID that
+    /// must not resolve (e.g. a small-order did:key sealed before the choke
+    /// point landed). The old envelope is already public and cannot be
+    /// withdrawn, but it must be surfaced for operator action rather than
+    /// silently skipped as unchanged.
+    SkipUnresolvableStored(Vec<String>),
    /// No recipient DID resolved to a key, so there is nothing to seal to.
    SkipNoRecipients,
    /// At least one recipient DID is unresolvable. Fail closed: never seal to a
@@ -90,15 +97,30 @@ enum SealPlan {
 /// recover the blob; reader removal is not retroactive (the old envelope is
 /// already public). The comparison is on the opaque node-keyed tag, never the
 /// DID list.
+///
+/// A matching stored tag is normally a skip, but the set is still resolved
+/// before skipping: a reader that used to resolve and now must not (a
+/// small-order did:key sealed before the choke point landed) is surfaced as
+/// [`SealPlan::SkipUnresolvableStored`] instead of being silently treated as
+/// unchanged. The old envelope is already public and cannot be withdrawn, but
+/// the operator must be able to see that a persisted envelope was sealed to a
+/// weak recipient.
 fn plan_seal(node_seed: &[u8; 32], dids: &BTreeSet<String>, stored_tag: Option<&str>) -> SealPlan {
    let tag = recipients_tag(node_seed, dids);
    if stored_tag == Some(tag.as_str()) {
-        return SealPlan::SkipUnchanged;
-    }
-    match resolve_all_recipients(dids) {
-        Ok(keys) if keys.is_empty() => SealPlan::SkipNoRecipients,
-        Ok(keys) => SealPlan::Seal { keys, tag },
-        Err(unresolved) => SealPlan::SkipUnresolvable(unresolved),
+        // The stored envelope already covers exactly this set, so there is
+        // nothing to re-seal. Resolve anyway so a now-unresolvable reader in an
+        // unchanged set is surfaced rather than hidden by the tag match.
+        match resolve_all_recipients(dids) {
+            Ok(_) => SealPlan::SkipUnchanged,
+            Err(unresolved) => SealPlan::SkipUnresolvableStored(unresolved),
+        }
+    } else {
+        match resolve_all_recipients(dids) {
+            Ok(keys) if keys.is_empty() => SealPlan::SkipNoRecipients,
+            Ok(keys) => SealPlan::Seal { keys, tag },
+            Err(unresolved) => SealPlan::SkipUnresolvable(unresolved),
+        }
    }
 }

@@ -106,17 +128,60 @@ fn plan_seal(node_seed: &[u8; 32], dids: &BTreeSet<String>, stored_tag: Option<&
 /// `node_seed` keys the opaque recipients tag. Returns `(oid, cid)` for each blob
 /// actually sealed and recorded this call (the per-push delta), used by Option B3
 /// to anchor a manifest. Recipient identities are never stored or returned.
+///
+/// Nine args (the fence joins the seal's eight) but grouping them would churn
+/// both callers and the race/hung-git tests for no behavioral gain.
+#[allow(clippy::too_many_arguments)]
 pub async fn encrypt_and_pin(
    ipfs_api: &str,
    repo_path: &Path,
    db: &Db,
    repo_id: &str,
    node_seed: &[u8; 32],
+    git_bin: &str,
+    batch_budget: Duration,
    recipients: &HashMap<String, BTreeSet<String>>,
+    fence: Option<&crate::ipfs_pin::PolicyFence>,
 ) -> Vec<(String, String)> {
    let mut sealed = Vec::new();
    let mut skipped_unresolvable = 0usize;
-    for (oid, dids) in recipients {
+    // One shared read deadline for the whole batch, like `pin_new_objects`: a
+    // hung git child is watchdog-reaped at this bound, so the outer
+    // `PIN_PHASE_DEADLINE` timeout cannot be held open by a blocking read
+    // (R1-P2). Each read runs under `spawn_blocking` — it is synchronous child
+    // spawn + pipe drain + watchdog join.
+    let read_deadline = std::time::Instant::now() + batch_budget;
+    let total = recipients.len();
+    for (attempted, (oid, dids)) in recipients.iter().enumerate() {
+        // Batch budget gate (R2-P3), mirroring the public pin loops: an object
+        // is never started with a remainder too small to cover a bounded read's
+        // teardown. This is consistency (the seal is bounded by the outer
+        // `PIN_PHASE_DEADLINE` either way), but it keeps the three loops from
+        // drifting apart in how they report a truncated batch.
+        if crate::ipfs_pin::batch_budget_gate(
+            "encrypted-seal",
+            read_deadline,
+            sealed.len(),
+            total - attempted,
+        )
+        .is_none()
+        {
+            break;
+        }
+        // Policy fence (R1-P1): the recipients snapshot was derived before the
+        // long withheld-blob walk; if a visibility rule moved while that walk
+        // ran (a reader added or removed), stop sealing instead of pinning to a
+        // stale recipient set. Checked FIRST so a changed policy costs nothing.
+        if let Some(f) = fence {
+            if !f.is_current().await {
+                tracing::warn!(
+                    repo = %f.repo_id(),
+                    oid = %oid,
+                    "visibility policy changed after the recipients snapshot; stopping the seal loop"
+                );
+                break;
+            }
+        }
        // A DB read failure is not a cache miss: re-sealing here would do an
        // avoidable IPFS write during a partial outage. Skip and retry next push.
        let stored_tag = match db.encrypted_blob_recipients_tag(repo_id, oid).await {
@@ -131,6 +196,21 @@ pub async fn encrypt_and_pin(
        // covered (which would permanently lock out the dropped readers, #47).
        let (keys, tag) = match plan_seal(node_seed, dids, stored_tag.as_deref()) {
            SealPlan::SkipUnchanged => continue,
+            SealPlan::SkipUnresolvableStored(unresolved) => {
+                // A stored envelope already covers this exact reader set, and
+                // the set now contains a DID that must not resolve (a
+                // small-order did:key sealed before the choke point landed).
+                // The old envelope is already public and cannot be withdrawn
+                // or re-sealed; surface it so the operator can act.
+                let sample: Vec<&String> = unresolved.iter().take(3).collect();
+                tracing::warn!(
+                    oid = %oid,
+                    unresolved_count = unresolved.len(),
+                    unresolved_sample = ?sample,
+                    "stored envelope was sealed to a recipient that must not resolve; operator action may be required"
+                );
+                continue;
+            }
            SealPlan::SkipNoRecipients => {
                tracing::warn!(oid = %oid, "no resolvable recipient keys; skipping encrypted pin");
                continue;
@@ -152,7 +232,9 @@ pub async fn encrypt_and_pin(
            }
            SealPlan::Seal { keys, tag } => (keys, tag),
        };
-        let data = match crate::git::store::read_object(repo_path, oid) {
+        let data = match read_object_bounded_spawn_blocking(git_bin, repo_path, oid, read_deadline)
+            .await
+        {
            Ok(Some((_t, bytes))) => bytes,
            Ok(None) => {
                tracing::warn!(oid = %oid, "git object not found; skipping encrypted pin");
@@ -170,6 +252,22 @@ pub async fn encrypt_and_pin(
                continue;
            }
        };
+        // Dispatch fence (R1-P1): re-read the policy epoch immediately before
+        // the irreversible HTTP POST. The iteration-top check catches a narrow
+        // that landed before work began; THIS check catches a narrow that landed
+        // during the tag lookup, recipient resolution, git read, or seal — all
+        // of which can take seconds. Without this, a reader removed during
+        // preparation can still receive a newly published envelope.
+        if let Some(f) = fence {
+            if !f.is_current().await {
+                tracing::warn!(
+                    repo = %f.repo_id(),
+                    oid = %oid,
+                    "visibility policy changed during encrypted seal preparation; aborting upload"
+                );
+                break;
+            }
+        }
        let cid = match crate::ipfs_pin::pin_git_object(ipfs_api, oid, &envelope, None).await {
            Ok(c) if !c.is_empty() => c,
            Ok(_) => {
@@ -201,10 +299,76 @@ pub async fn encrypt_and_pin(
    sealed
 }

+/// Bounded, reaped git object read for the seal loop, run off the async thread:
+/// `read_object_bounded` is synchronous child spawn + pipe drain + watchdog
+/// join, so blocking the runtime task on it would let a hung git hold a worker
+/// thread (R1-P2). The `deadline` is the batch's shared read deadline; a child
+/// still alive at it is SIGTERM/SIGKILL group-reaped by the watchdog.
+async fn read_object_bounded_spawn_blocking(
+    git_bin: &str,
+    repo_path: &Path,
+    sha256_hex: &str,
+    deadline: std::time::Instant,
+) -> anyhow::Result<Option<(String, Vec<u8>)>> {
+    let git_bin = git_bin.to_string();
+    let repo_path = repo_path.to_path_buf();
+    let sha256_hex = sha256_hex.to_string();
+    tokio::task::spawn_blocking(move || {
+        crate::git::store::read_object_bounded(&git_bin, &repo_path, &sha256_hex, deadline)
+            .map_err(anyhow::Error::from)
+    })
+    .await
+    .map_err(|e| anyhow::anyhow!("read_object spawn_blocking join failed: {e}"))?
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;
    use ed25519_dalek::SigningKey;
+    use std::time::Duration;
+
+    /// A did:key encoding a small-order point. Well-formed, indistinguishable
+    /// from a real one by eye, and refused at resolution since the choke-point
+    /// guard landed.
+    fn weak_did_key() -> String {
+        let mut weak = [0u8; 32];
+        weak[0] = 1; // compressed identity point
+        let vk = ed25519_dalek::VerifyingKey::from_bytes(&weak).expect("decompresses");
+        Did::from_verifying_key(&vk).to_string()
+    }
+
+    /// The #47 fail-closed arm already refuses to seal to a partial recipient
+    /// set. Because a small-order did:key no longer RESOLVES, it lands in the
+    /// unresolved set and that existing arm catches it: nothing is sealed, and
+    /// no new code path was added to make that happen.
+    #[test]
+    fn weak_recipient_falls_into_the_existing_fail_closed_arm() {
+        let weak = weak_did_key();
+        let mut dids = BTreeSet::new();
+        dids.insert(did_key(9));
+        dids.insert(weak.clone());
+
+        match plan_seal(&[0u8; 32], &dids, None) {
+            SealPlan::SkipUnresolvable(unresolved) => assert!(
+                unresolved.contains(&weak),
+                "the weak DID must be named in the unresolved set, got {unresolved:?}"
+            ),
+            other => panic!("expected SkipUnresolvable, got {other:?}"),
+        }
+    }
+
+    /// Must-not-over-reject control: an all-legitimate recipient set still seals.
+    #[test]
+    fn legitimate_recipient_set_still_seals() {
+        let mut dids = BTreeSet::new();
+        dids.insert(did_key(9));
+        dids.insert(did_key(11));
+
+        match plan_seal(&[0u8; 32], &dids, None) {
+            SealPlan::Seal { keys, .. } => assert_eq!(keys.len(), 2, "both recipients resolve"),
+            other => panic!("expected Seal, got {other:?}"),
+        }
+    }

    fn did_key(seed: u8) -> String {
        let vk = SigningKey::from_bytes(&[seed; 32]).verifying_key();
@@ -348,6 +512,43 @@ mod tests {
        ));
    }

+    /// P1 remediation (review): an unchanged reader set is normally a skip, but
+    /// the skip must not hide a recipient that now fails to resolve. A blob
+    /// sealed to a small-order did:key before the choke-point guard landed keeps
+    /// its old CID and stays decryptable-without-a-private-key unless the
+    /// matching-tag path surfaces it. This pins that the stored-tag match still
+    /// resolves the set instead of short-circuiting.
+    #[test]
+    fn plan_seal_surfaces_unresolvable_recipient_in_unchanged_reader_set() {
+        let weak = weak_did_key();
+        let mut dids = BTreeSet::new();
+        dids.insert(did_key(9));
+        dids.insert(weak.clone());
+        let stored = recipients_tag(&SEED, &dids);
+
+        match plan_seal(&SEED, &dids, Some(&stored)) {
+            SealPlan::SkipUnresolvableStored(unresolved) => assert!(
+                unresolved.contains(&weak),
+                "the weak DID must be named in the unresolved set, got {unresolved:?}"
+            ),
+            other => panic!("expected SkipUnresolvableStored, got {other:?}"),
+        }
+    }
+
+    /// Degenerate control: an empty set with a matching stored tag stays a
+    /// silent skip. The state is unreachable in practice (a seal is never
+    /// recorded for an empty reader set), so this pins only that the
+    /// resolve-on-match path must not turn it into a re-seal or a warning.
+    #[test]
+    fn plan_seal_unchanged_empty_set_still_skips() {
+        let empty = BTreeSet::new();
+        let empty_tag = recipients_tag(&SEED, &empty);
+        assert!(matches!(
+            plan_seal(&SEED, &empty, Some(&empty_tag)),
+            SealPlan::SkipUnchanged
+        ));
+    }
+
    #[test]
    fn plan_seal_reseals_when_recipient_set_changed() {
        // A stored tag for a DIFFERENT set is a miss: a newly added reader must
@@ -359,4 +560,304 @@ mod tests {
            other => panic!("changed recipient set must re-seal; got {other:?}"),
        }
    }
+
+    /// A reader removed mid-seal must stop the seal loop (R1-P1 "race test for
+    /// reader removal"): `encrypt_and_pin` re-checks the policy fence before
+    /// each blob, so a `remove_visibility_rule` landing while the first seal is
+    /// in flight aborts before a later blob is pinned to a stale recipient set.
+    #[sqlx::test]
+    async fn encrypt_and_pin_stops_sealing_when_reader_removed_mid_batch(pool: sqlx::PgPool) {
+        let db = crate::db::Db::for_testing(pool);
+        db.run_migrations().await.expect("migrations");
+        let tmp = tempfile::TempDir::new().unwrap();
+        let repo_path = tmp.path().join("seal-race.git");
+
+        // Three loose blobs, each withheld (path-scoped deny exists so the sweep
+        // would have derived recipients for them).
+        let oids: Vec<String> = {
+            crate::git::store::init_bare(&repo_path).expect("init bare repo");
+            (0..3)
+                .map(|i| {
+                    let mut cmd = std::process::Command::new("git");
+                    cmd.args(["hash-object", "-w", "--stdin"])
+                        .current_dir(&repo_path)
+                        .stdin(std::process::Stdio::piped())
+                        .stdout(std::process::Stdio::piped());
+                    let mut child = cmd.spawn().expect("spawn git hash-object");
+                    {
+                        use std::io::Write;
+                        child
+                            .stdin
+                            .as_mut()
+                            .expect("stdin")
+                            .write_all(format!("secret blob {i}\n").as_bytes())
+                            .expect("write stdin");
+                    }
+                    let out = child.wait_with_output().expect("hash-object output");
+                    assert!(out.status.success());
+                    String::from_utf8_lossy(&out.stdout).trim().to_string()
+                })
+                .collect()
+        };
+
+        // A real repos row so the fence has an epoch and a reader can be removed.
+        let now = chrono::Utc::now();
+        let repo_id = uuid::Uuid::new_v4().to_string();
+        db.create_repo(&crate::db::RepoRecord {
+            id: repo_id.clone(),
+            name: "seal-race-repo".into(),
+            owner_did: "did:key:zSealRaceOwner".into(),
+            description: None,
+            is_public: true,
+            default_branch: "main".into(),
+            created_at: now,
+            updated_at: now,
+            disk_path: repo_path.display().to_string(),
+            forked_from: None,
+            machine_id: None,
+        })
+        .await
+        .expect("create repo");
+        // A rule whose removal is the "reader removed" mutation: one reader per
+        // blob, all under the same path glob.
+        let reader = did_key(1);
+        db.set_visibility_rule(
+            &repo_id,
+            "**/secret/*",
+            crate::db::VisibilityMode::B,
+            std::slice::from_ref(&reader),
+            "did:key:zSealRaceOwner",
+        )
+        .await
+        .expect("set rule");
+
+        // IPFS endpoint that delays the FIRST add 2s so the removal lands while
+        // that seal is in flight, then answers immediately.
+        let endpoint = delaying_cid_endpoint(vec![Duration::from_secs(2)]).await;
+
+        let recipients: HashMap<String, BTreeSet<String>> = oids
+            .iter()
+            .cloned()
+            .map(|oid| {
+                let mut s = BTreeSet::new();
+                s.insert(reader.clone());
+                (oid, s)
+            })
+            .collect();
+
+        let fence = crate::ipfs_pin::PolicyFence::capture(&db, &repo_id)
+            .await
+            .expect("fence captures");
+
+        let sealed = tokio::time::timeout(Duration::from_secs(30), async {
+            let seal_db = db.clone();
+            let seal_repo = repo_path.clone();
+            let seal_endpoint = endpoint.clone();
+            let seal_repo_id = repo_id.clone();
+            let handle = tokio::spawn(async move {
+                encrypt_and_pin(
+                    &seal_endpoint,
+                    &seal_repo,
+                    &seal_db,
+                    &seal_repo_id,
+                    &SEED,
+                    "git",
+                    Duration::from_secs(60),
+                    &recipients,
+                    Some(&fence),
+                )
+                .await
+            });
+            // Let the first add start (endpoint sleeps 2s), then remove the
+            // reader so the fence is stale before the loop checks again.
+            // The window has to cover blob 1's encrypt phase: if the removal
+            // lands before the loop reaches dispatch, the fence aborts with
+            // nothing sealed, so give the in-flight add room to start under
+            // a loaded box while staying inside the 2s endpoint delay.
+            tokio::time::sleep(Duration::from_millis(1200)).await;
+            db.remove_visibility_rule(&repo_id, "**/secret/*")
+                .await
+                .expect("remove rule");
+            handle.await.expect("seal task")
+        })
+        .await
+        .expect("wedge guard: the fence abort must not take 30s");
+
+        assert!(
+            sealed.len() < oids.len(),
+            "a reader removal landing mid-batch must abort before every blob is sealed: {}",
+            sealed.len()
+        );
+        assert!(
+            !sealed.is_empty(),
+            "at least the blob already in flight before the removal completes"
+        );
+    }
+
+    /// A hung git must not hold the seal loop past its read budget (R1-P2): the
+    /// git read runs under `spawn_blocking` against `read_object_bounded`, so
+    /// the watchdog reaps a wedged child at the batch deadline and the loop
+    /// keeps its shape instead of blocking a runtime worker indefinitely.
+    #[cfg(unix)]
+    #[sqlx::test]
+    async fn encrypt_and_pin_returns_by_budget_with_a_hung_git(pool: sqlx::PgPool) {
+        use std::os::unix::fs::PermissionsExt;
+        let db = crate::db::Db::for_testing(pool);
+        db.run_migrations().await.expect("migrations");
+        let tmp = tempfile::TempDir::new().unwrap();
+        let repo_path = tmp.path().join("seal-hung.git");
+        let oids: Vec<String> = {
+            crate::git::store::init_bare(&repo_path).expect("init bare repo");
+            (0..2)
+                .map(|i| {
+                    let mut cmd = std::process::Command::new("git");
+                    cmd.args(["hash-object", "-w", "--stdin"])
+                        .current_dir(&repo_path)
+                        .stdin(std::process::Stdio::piped())
+                        .stdout(std::process::Stdio::piped());
+                    let mut child = cmd.spawn().expect("spawn git hash-object");
+                    {
+                        use std::io::Write;
+                        child
+                            .stdin
+                            .as_mut()
+                            .expect("stdin")
+                            .write_all(format!("secret blob {i}\n").as_bytes())
+                            .expect("write stdin");
+                    }
+                    let out = child.wait_with_output().expect("hash-object output");
+                    assert!(out.status.success());
+                    String::from_utf8_lossy(&out.stdout).trim().to_string()
+                })
+                .collect()
+        };
+
+        // A git that wedges forever, ignoring SIGTERM, so only the watchdog's
+        // SIGKILL can reap it.
+        let fake = tmp.path().join("hanging-git");
+        std::fs::write(&fake, "#!/bin/sh\ntrap '' TERM\necho $$ > pid\nsleep 30\n").unwrap();
+        let mut perm = std::fs::metadata(&fake).unwrap().permissions();
+        perm.set_mode(0o755);
+        std::fs::set_permissions(&fake, perm).unwrap();
+
+        let now = chrono::Utc::now();
+        let repo_id = uuid::Uuid::new_v4().to_string();
+        db.create_repo(&crate::db::RepoRecord {
+            id: repo_id.clone(),
+            name: "seal-hung-repo".into(),
+            owner_did: "did:key:zSealHungOwner".into(),
+            description: None,
+            is_public: true,
+            default_branch: "main".into(),
+            created_at: now,
+            updated_at: now,
+            disk_path: repo_path.display().to_string(),
+            forked_from: None,
+            machine_id: None,
+        })
+        .await
+        .expect("create repo");
+        db.set_visibility_rule(
+            &repo_id,
+            "**/secret/*",
+            crate::db::VisibilityMode::B,
+            &[did_key(1)],
+            "did:key:zSealHungOwner",
+        )
+        .await
+        .expect("set rule");
+
+        let recipients: HashMap<String, BTreeSet<String>> = oids
+            .iter()
+            .cloned()
+            .map(|oid| {
+                let mut s = BTreeSet::new();
+                s.insert(did_key(1));
+                (oid, s)
+            })
+            .collect();
+
+        // Unreachable endpoint: even if a read somehow succeeded, the pin would
+        // fail; the read itself is the thing under test.
+        let started = std::time::Instant::now();
+        let sealed = tokio::time::timeout(
+            Duration::from_secs(60),
+            encrypt_and_pin(
+                "http://127.0.0.1:9",
+                &repo_path,
+                &db,
+                &repo_id,
+                &SEED,
+                fake.to_str().unwrap(),
+                Duration::from_secs(2),
+                &recipients,
+                None,
+            ),
+        )
+        .await
+        .expect("a hung git must not hold the seal past the outer wedge guard");
+
+        let elapsed = started.elapsed();
+        assert!(
+            elapsed < Duration::from_secs(10),
+            "a hung git must be watchdog-reaped inside the read budget, not block the loop for ~10s+ (took {elapsed:?})"
+        );
+        assert!(
+            sealed.is_empty(),
+            "with a hung git no blob can be read, so nothing may be reported sealed"
+        );
+    }
+
+    /// Local TCP endpoint that answers `{ "Hash": "QmMock" }` after an optional
+    /// per-request delay, so a seal can be made to straddle a policy mutation.
+    async fn delaying_cid_endpoint(delays: Vec<Duration>) -> String {
+        use tokio::io::{AsyncReadExt, AsyncWriteExt};
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let endpoint = format!("http://{}", listener.local_addr().unwrap());
+        tokio::spawn(async move {
+            let mut seen = 0usize;
+            while let Ok((mut sock, _)) = listener.accept().await {
+                let delay = *delays
+                    .get(seen)
+                    .or_else(|| delays.last())
+                    .unwrap_or(&Duration::ZERO);
+                seen += 1;
+                tokio::spawn(async move {
+                    let mut acc = Vec::new();
+                    let mut buf = [0u8; 4096];
+                    loop {
+                        let n = match sock.read(&mut buf).await {
+                            Ok(0) | Err(_) => break,
+                            Ok(n) => n,
+                        };
+                        acc.extend_from_slice(&buf[..n]);
+                        if let Some(hdr_end) =
+                            acc.windows(4).position(|w| w == b"\r\n\r\n").map(|p| p + 4)
+                        {
+                            let headers = String::from_utf8_lossy(&acc[..hdr_end]).to_lowercase();
+                            let len: usize = headers
+                                .lines()
+                                .find_map(|l| l.strip_prefix("content-length:"))
+                                .and_then(|v| v.trim().parse().ok())
+                                .unwrap_or(0);
+                            if acc.len() >= hdr_end + len {
+                                break;
+                            }
+                        }
+                    }
+                    tokio::time::sleep(delay).await;
+                    let body = br#"{"Hash":"bafkreifjjcie6lypi6ny7amxnfftagclbuxndqonfipmb64f2km2devei4"}"#;
+                    let _ = sock
+                        .write_all(
+                            format!("HTTP/1.1 200 OK\r\nContent-Length: {}\r\n\r\n", body.len())
+                                .as_bytes(),
+                        )
+                        .await;
+                    let _ = sock.write_all(body).await;
+                    let _ = sock.flush().await;
+                });
+            }
+        });
+        endpoint
+    }
 }
@@ -12,6 +12,9 @@ pub enum AppError {
    #[error("repo already exists: {0}")]
    RepoExists(String),

+    #[error("conflict: {0}")]
+    Conflict(String),
+
    #[error("not found: {0}")]
    NotFound(String),

@@ -50,6 +53,16 @@ pub enum AppError {
    #[error("incomplete: {0}")]
    Incomplete(String),

+    /// A bounded search that could not complete. `continuation`, when present, is the
+    /// sealed scan position the caller echoes as `?scan=` to resume where the search
+    /// stopped (#173 round 13, F2). It is AEAD-sealed at the mint site, never plaintext:
+    /// the row it names is by construction one the caller was denied (INV-13).
+    #[error("search incomplete: {message}")]
+    SearchIncomplete {
+        message: String,
+        continuation: Option<String>,
+    },
+
    #[error("git error: {0}")]
    Git(String),

@@ -59,6 +72,15 @@ pub enum AppError {
    #[error("server overloaded: {0}")]
    Overloaded(String),

+    #[error("repository is busy")]
+    RepoBusy,
+
+    #[error("repository is temporarily unavailable")]
+    RepoUnavailable,
+
+    #[error("repository write was fenced by a concurrent publish")]
+    RepoWriteFenced,
+
    #[error("database error: {0}")]
    Db(#[from] sqlx::Error),

@@ -82,7 +104,7 @@ pub const DB_ERROR_MESSAGE: &str = "a database error occurred";

 /// Connection-level sqlx failures that mean the database is unreachable right
 /// now (retryable, 503), as opposed to server-reported query errors.
-fn db_unavailable(e: &sqlx::Error) -> bool {
+pub(crate) fn db_unavailable(e: &sqlx::Error) -> bool {
    matches!(
        e,
        sqlx::Error::PoolTimedOut
@@ -100,7 +122,31 @@ impl From<anyhow::Error> for AppError {
    fn from(err: anyhow::Error) -> Self {
        match err.downcast::<sqlx::Error>() {
            Ok(sql) => AppError::Db(sql),
-            Err(err) => AppError::Internal(err),
+            // Lock contention is transient and ordinary, so it must not land as a
+            // 500. The internal message names the owner slug and repo, so the
+            // variant carries nothing: the detail stays in the log at the raise
+            // site and the client gets a fixed retryable body.
+            Err(err) => match err.downcast::<crate::git::repo_store::LockPoolBusy>() {
+                Ok(_) => AppError::Overloaded("git write locks at capacity, retry shortly".into()),
+                Err(err) => match err.downcast::<crate::git::repo_store::RepoBusy>() {
+                    Ok(_) => AppError::RepoBusy,
+                    // Same reasoning one rung down: a refused under-lock refresh is a
+                    // transient storage condition, and its internal message names the
+                    // owner slug and repo, so the variant carries nothing.
+                    Err(err) => match err.downcast::<crate::git::repo_store::RepoUnavailable>() {
+                        Ok(_) => AppError::RepoUnavailable,
+                        // And one more rung: a publish the store refused twice is
+                        // transient in the same way, and the retry is the client's
+                        // to make. The variant carries nothing for the same reason
+                        // as the two above.
+                        Err(err) => match err.downcast::<crate::git::repo_store::RepoWriteFenced>()
+                        {
+                            Ok(_) => AppError::RepoWriteFenced,
+                            Err(err) => AppError::Internal(err),
+                        },
+                    },
+                },
+            },
        }
    }
 }
@@ -146,6 +192,7 @@ impl IntoResponse for AppError {
                "repo_exists",
                format!("repository '{r}' already exists"),
            ),
+            AppError::Conflict(msg) => (StatusCode::CONFLICT, "conflict", msg.clone()),
            AppError::NotFound(msg) => (StatusCode::NOT_FOUND, "not_found", msg.clone()),
            AppError::Unauthorized(msg) => (StatusCode::UNAUTHORIZED, "not_an_agent", msg.clone()),
            AppError::Forbidden(msg) => (StatusCode::FORBIDDEN, "forbidden", msg.clone()),
@@ -161,10 +208,42 @@ impl IntoResponse for AppError {
            AppError::Incomplete(msg) => {
                (StatusCode::UNPROCESSABLE_ENTITY, "incomplete", msg.clone())
            }
+            // A bounded search that could not complete (the CID resolver hit its
+            // legacy-probe or walk ceiling), distinct from the 404 that asserts a
+            // definitive not-found: absence was NOT proven, so the caller should
+            // retry rather than treat it as gone (#173, F2). 503, retryable.
+            AppError::SearchIncomplete { message, .. } => (
+                StatusCode::SERVICE_UNAVAILABLE,
+                "search_incomplete",
+                message.clone(),
+            ),
            AppError::Git(msg) => (StatusCode::INTERNAL_SERVER_ERROR, "git_error", msg.clone()),
            // 504, distinct from the 500 git_error and from the read-gate's 404 /
            // the auth 401, so the client can tell a deadline from a failure.
            AppError::Timeout(msg) => (StatusCode::GATEWAY_TIMEOUT, "git_timeout", msg.clone()),
+            // 503 with a FIXED body: the caller should retry, and must not be told
+            // which repo is contended or for how long.
+            AppError::RepoBusy => (
+                StatusCode::SERVICE_UNAVAILABLE,
+                "repo_busy",
+                "repository is busy — retry".into(),
+            ),
+            // 503 with a FIXED body for the same reason: the caller should retry, and
+            // must not be told which repo could not be refreshed or why.
+            AppError::RepoUnavailable => (
+                StatusCode::SERVICE_UNAVAILABLE,
+                "repo_unavailable",
+                "repository is temporarily unavailable, retry".into(),
+            ),
+            // 503 with a FIXED body again, and its own code: the caller should
+            // retry, but the condition is not contention, so a client that
+            // distinguishes them should be able to. The body must not say which
+            // repo lost its publish or to whom.
+            AppError::RepoWriteFenced => (
+                StatusCode::SERVICE_UNAVAILABLE,
+                "repo_write_fenced",
+                "repository changed underneath this write, retry".into(),
+            ),
            AppError::Db(e) if db_unavailable(e) => (
                StatusCode::SERVICE_UNAVAILABLE,
                DB_UNAVAILABLE_CODE,
@@ -201,16 +280,36 @@ impl IntoResponse for AppError {
            }
        };

-        let body = Json(json!({
+        let mut body = json!({
            "error": code,
            "message": message,
-        }));
+        });
+        // A truncated CID search may carry the sealed position the caller echoes as
+        // `?scan=` to resume. Rendered as a third body field, present only when the
+        // shed actually left something to resume: a wrapped scan and a throttled
+        // request both omit it, and its ABSENCE is what tells a caller the ladder is
+        // over. It is opaque ciphertext; see `gitlawb_core::scan_token`.
+        if let AppError::SearchIncomplete {
+            continuation: Some(token),
+            ..
+        } = &self
+        {
+            body["continuation"] = json!(token);
+        }

-        let mut resp = (status, body).into_response();
-        // Overloaded advertises when to retry. It rides the shared tail above for
-        // its body/status, so the header is attached here rather than in a bespoke
-        // early return — keeping the variant handled in exactly one place.
-        if matches!(self, AppError::Overloaded(_)) {
+        let mut resp = (status, Json(body)).into_response();
+        // Both retryable 503s advertise when to retry: Overloaded (capacity shed) and
+        // SearchIncomplete (a bounded CID search cut short by a cap — retry may complete
+        // it). They ride the shared tail above for body/status, so the header is attached
+        // here rather than in bespoke early returns, keeping each variant handled once.
+        if matches!(
+            self,
+            AppError::Overloaded(_)
+                | AppError::SearchIncomplete { .. }
+                | AppError::RepoBusy
+                | AppError::RepoUnavailable
+                | AppError::RepoWriteFenced
+        ) {
            resp.headers_mut().insert(
                axum::http::header::RETRY_AFTER,
                axum::http::HeaderValue::from_static("1"),
@@ -226,6 +325,16 @@ pub type Result<T> = std::result::Result<T, AppError>;
 mod tests {
    use super::*;

+    #[test]
+    fn conflict_maps_to_409() {
+        assert_eq!(
+            AppError::Conflict("task not claimable".into())
+                .into_response()
+                .status(),
+            StatusCode::CONFLICT
+        );
+    }
+
    #[test]
    fn timeout_maps_to_504_distinct_from_git_500() {
        assert_eq!(
@@ -327,4 +436,31 @@ mod tests {
        let resp = err.into_response();
        assert_eq!(resp.status(), StatusCode::SERVICE_UNAVAILABLE);
    }
+
+    #[test]
+    fn lock_pool_busy_via_anyhow_from_is_503_overloaded() {
+        let err: AppError = anyhow::Error::new(crate::git::repo_store::LockPoolBusy).into();
+        let resp = err.into_response();
+        assert_eq!(resp.status(), StatusCode::SERVICE_UNAVAILABLE);
+        assert_eq!(
+            resp.headers().get("retry-after").unwrap().to_str().unwrap(),
+            "1"
+        );
+    }
+
+    #[test]
+    fn repo_busy_unavailable_and_fenced_advertise_retry_after() {
+        for err in [
+            AppError::RepoBusy,
+            AppError::RepoUnavailable,
+            AppError::RepoWriteFenced,
+        ] {
+            let resp = err.into_response();
+            assert_eq!(resp.status(), StatusCode::SERVICE_UNAVAILABLE);
+            assert_eq!(
+                resp.headers().get("retry-after").unwrap().to_str().unwrap(),
+                "1"
+            );
+        }
+    }
 }
@@ -61,6 +61,27 @@ pub fn create_issue(repo_path: &Path, issue_id: &str, json: &str) -> Result<()>
    Ok(())
 }

+/// Remove a single issue ref while the write guard still holds the advisory lock,
+/// after a definite publish refusal. Best-effort: a failed delete leaves a
+/// local-only orphan, which is still better than telling the client to retry into a
+/// duplicate id. Not used on `UploadUnknowable`, where the PUT may still land.
+pub fn delete_issue_ref(
+    git_bin: &str,
+    repo_path: &Path,
+    issue_id: &str,
+    deadline: std::time::Instant,
+) -> Result<()> {
+    let ref_name = format!("refs/gitlawb/issues/{issue_id}");
+    crate::git::visibility_pack::run_bounded_git(
+        git_bin,
+        &["update-ref", "-d", &ref_name],
+        repo_path,
+        b"",
+        deadline,
+    )?;
+    Ok(())
+}
+
 /// List all issue refs and return their JSON content.
 pub fn list_issues(repo_path: &Path) -> Result<Vec<String>> {
    // List all refs under refs/gitlawb/issues/
@@ -108,19 +129,22 @@ pub fn list_issues(repo_path: &Path) -> Result<Vec<String>> {
 /// Returns Ok(Some(full_id)) on unique match, Ok(None) if not found,
 /// Err if the prefix is ambiguous (matches more than one issue).
 pub fn resolve_issue_id(repo_path: &Path, id_or_prefix: &str) -> Result<Option<String>> {
-    // Try exact match first — fast path for callers passing the full UUID.
-    let exact_ref = format!("refs/gitlawb/issues/{id_or_prefix}");
-    let check = Command::new("git")
-        .args(["cat-file", "-e", &exact_ref])
-        .current_dir(repo_path)
-        .output()
-        .context("failed to run git cat-file -e")?;
-    if check.status.success() {
-        return Ok(Some(id_or_prefix.to_string()));
+    // The id comes from the request path and is spliced into a git revision and a
+    // for-each-ref pattern below. Issue ids are UUIDs, so accept only ASCII
+    // alphanumerics and '-': anything else (`^{blob}`, `@{0}`, `*`, `/`, `..`) is
+    // revision or glob syntax, never an issue id, and is reported as not found.
+    if id_or_prefix.is_empty()
+        || !id_or_prefix
+            .bytes()
+            .all(|b| b.is_ascii_alphanumeric() || b == b'-')
+    {
+        return Ok(None);
    }

-    // Prefix search: list all refs that start with the given string.
-    let prefix_glob = format!("refs/gitlawb/issues/{id_or_prefix}*");
+    // List all refs that start with the given string; the glob also matches the exact
+    // ref. The id handed back is always git's own refname, never the request input.
+    let exact_ref = format!("refs/gitlawb/issues/{id_or_prefix}");
+    let prefix_glob = format!("{exact_ref}*");
    let list = Command::new("git")
        .args(["for-each-ref", "--format=%(refname)", &prefix_glob])
        .current_dir(repo_path)
@@ -132,25 +156,44 @@ pub fn resolve_issue_id(repo_path: &Path, id_or_prefix: &str) -> Result<Option<S
    }

    let output = String::from_utf8_lossy(&list.stdout);
-    let matches: Vec<&str> = output.lines().filter(|l| !l.trim().is_empty()).collect();
+    let matches: Vec<&str> = output
+        .lines()
+        .map(str::trim)
+        .filter(|l| !l.is_empty())
+        .collect();

-    match matches.len() {
-        0 => Ok(None),
-        1 => {
-            // Strip the "refs/gitlawb/issues/" prefix to get the bare ID.
-            let full_id = matches[0]
-                .trim()
-                .strip_prefix("refs/gitlawb/issues/")
-                .unwrap_or(matches[0].trim())
-                .to_string();
-            Ok(Some(full_id))
+    // An exact id wins over longer ids that share it as a prefix, but only if its
+    // object exists: for-each-ref also lists a ref whose blob is missing, so a
+    // dangling exact ref falls through to the prefix rules below.
+    let exact = match matches.iter().find(|m| **m == exact_ref) {
+        Some(m) => {
+            let check = Command::new("git")
+                .args(["cat-file", "-e", m])
+                .current_dir(repo_path)
+                .output()
+                .context("failed to run git cat-file -e")?;
+            check.status.success().then_some(*m)
        }
-        _ => anyhow::bail!(
-            "ambiguous issue prefix '{}': matches {} issues",
-            id_or_prefix,
-            matches.len()
-        ),
-    }
+        None => None,
+    };
+    let found = match exact {
+        Some(m) => m,
+        None => match matches.as_slice() {
+            [] => return Ok(None),
+            [m] => *m,
+            _ => anyhow::bail!(
+                "ambiguous issue prefix '{}': matches {} issues",
+                id_or_prefix,
+                matches.len()
+            ),
+        },
+    };
+    // Strip the "refs/gitlawb/issues/" prefix to get the bare ID.
+    let full_id = found
+        .strip_prefix("refs/gitlawb/issues/")
+        .unwrap_or(found)
+        .to_string();
+    Ok(Some(full_id))
 }

 /// Close an issue by updating its status to "closed" in the git ref.
@@ -234,6 +277,43 @@ mod tests {
        assert_eq!(resolved, Some(full_id.to_string()));
    }

+    #[test]
+    fn test_resolve_exact_id_wins_over_longer_id() {
+        let dir = TempDir::new().unwrap();
+        init_repo(&dir);
+        create_issue(dir.path(), "abc12345", r#"{"status":"open"}"#).unwrap();
+        create_issue(dir.path(), "abc12345-0000", r#"{"status":"open"}"#).unwrap();
+        let resolved = resolve_issue_id(dir.path(), "abc12345").unwrap();
+        assert_eq!(resolved, Some("abc12345".to_string()));
+    }
+
+    #[test]
+    fn test_resolve_dangling_exact_id_does_not_win_over_longer_id() {
+        let dir = TempDir::new().unwrap();
+        init_repo(&dir);
+        create_issue(dir.path(), "abc12345", r#"{"status":"open","n":1}"#).unwrap();
+        create_issue(dir.path(), "abc12345-0000", r#"{"status":"open","n":2}"#).unwrap();
+        // Dangle the exact ref: the ref stays listed by for-each-ref, but its blob is gone.
+        let out = Command::new("git")
+            .args(["rev-parse", "refs/gitlawb/issues/abc12345"])
+            .current_dir(dir.path())
+            .output()
+            .unwrap();
+        let hash = String::from_utf8(out.stdout).unwrap();
+        let hash = hash.trim();
+        std::fs::remove_file(
+            dir.path()
+                .join(".git/objects")
+                .join(&hash[..2])
+                .join(&hash[2..]),
+        )
+        .unwrap();
+
+        let result = resolve_issue_id(dir.path(), "abc12345");
+        assert!(result.unwrap_err().to_string().contains("ambiguous"));
+        assert!(close_issue(dir.path(), "abc12345").is_err());
+    }
+
    #[test]
    fn test_resolve_prefix_matches_unique() {
        let dir = TempDir::new().unwrap();
@@ -249,6 +329,35 @@ mod tests {
        assert_eq!(resolved, Some(full_id.to_string()));
    }

+    #[test]
+    fn test_resolve_rejects_revision_syntax() {
+        let dir = TempDir::new().unwrap();
+        init_repo(&dir);
+        let full_id = "abc12345-0000-0000-0000-000000000000";
+        create_issue(
+            dir.path(),
+            full_id,
+            r#"{"id":"abc12345-0000-0000-0000-000000000000","status":"open"}"#,
+        )
+        .unwrap();
+        // Each of these resolves to (or globs over) the issue blob in git, but none
+        // is an issue id, so the lookup must not accept it.
+        for crafted in [
+            "abc12345-0000-0000-0000-000000000000^{blob}",
+            "abc12345-0000-0000-0000-000000000000^{}",
+            "abc12345-0000-0000-0000-000000000000~0",
+            "*",
+            "",
+        ] {
+            assert_eq!(
+                resolve_issue_id(dir.path(), crafted).unwrap(),
+                None,
+                "{crafted:?}"
+            );
+            assert_eq!(get_issue(dir.path(), crafted).unwrap(), None, "{crafted:?}");
+        }
+    }
+
    #[test]
    fn test_resolve_prefix_not_found() {
        let dir = TempDir::new().unwrap();
@@ -278,6 +387,27 @@ mod tests {
        assert!(result.unwrap_err().to_string().contains("ambiguous"));
    }

+    #[test]
+    fn delete_issue_ref_removes_a_created_ref() {
+        let dir = TempDir::new().unwrap();
+        init_repo(&dir);
+        let full_id = "eee88888-0000-0000-0000-000000000000";
+        create_issue(
+            dir.path(),
+            full_id,
+            r#"{"id":"eee88888-0000-0000-0000-000000000000","status":"open"}"#,
+        )
+        .unwrap();
+        delete_issue_ref(
+            "git",
+            dir.path(),
+            full_id,
+            std::time::Instant::now() + std::time::Duration::from_secs(30),
+        )
+        .unwrap();
+        assert_eq!(resolve_issue_id(dir.path(), full_id).unwrap(), None);
+    }
+
    #[test]
    fn test_close_issue_via_prefix() {
        let dir = TempDir::new().unwrap();
@@ -1,4 +1,5 @@
 pub mod issues;
+pub mod publish;
 pub mod push_delta;
 pub mod repo_store;
 pub mod smart_http;
@@ -0,0 +1,358 @@
+//! The publication boundary: WHO wrote, and HOW FAR that write got.
+//!
+//! Four lifecycle defects on this branch shared one root cause. A resource was
+//! tracked by a coarse state — "release started", "the path exists", "a row with
+//! this name exists" — when the safety question was about the IDENTITY and the
+//! PUBLICATION STAGE of one specific write attempt. A tail replicated refs from
+//! an attempt that had not dispatched a PUT; a read served a tree whose
+//! generation nobody had confirmed; a fork claimed a concurrent attempt's row and
+//! deleted a successor's object; a response-loss failure was compensated as if it
+//! proved non-publication.
+//!
+//! This module is the vocabulary that makes those questions answerable, and it is
+//! deliberately free of any object-storage type:
+//!
+//! - [`PublishAttemptId`] — minted before the request is built, carried with the
+//!   bytes as user metadata, read back off the store to decide whether what is
+//!   published is THIS attempt's work.
+//! - [`PublishStage`] / [`PublishStageCell`] — how far one attempt got, observable
+//!   from outside the future that is performing it, so a CANCELLED attempt can
+//!   still be classified.
+//! - [`UploadError`] — what the client KNOWS about dispatch and commit, not just
+//!   which HTTP status came back. Destructive compensation is licensed only by
+//!   [`UploadError::proves_not_published`].
+//!
+//! # Porting note (#79)
+//!
+//! PR #79 (`feat/storage-abstraction`) deletes `git/tigris.rs` and replaces it
+//! with a `BlobStore`/`RepoArchive` layer. Everything in this file is
+//! backend-agnostic on purpose and is meant to survive that swap unchanged: a new
+//! backend supplies its own error classifier (the one in `tigris.rs` is a single
+//! private function) and keeps these types as the contract its callers read.
+
+use std::sync::Mutex;
+
+/// A durable identity for ONE publish attempt.
+///
+/// The point is reconciliation. A conditional PUT whose response was lost leaves
+/// the client unable to say whether its bytes committed; an attempt id stored
+/// alongside those bytes turns that into a question the store can answer, because
+/// "is the published object mine?" is decidable where "did my request succeed?"
+/// is not.
+///
+/// Fork creation passes the DB row id it is about to insert, so the object, the
+/// on-disk clone and the database row are all stamped with the same attempt
+/// identity and every cleanup can be made conditional on still owning all three.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct PublishAttemptId(String);
+
+impl PublishAttemptId {
+    /// A fresh identity for an attempt that has nothing else to be named after.
+    pub fn new() -> Self {
+        Self(uuid::Uuid::new_v4().to_string())
+    }
+
+    /// Name the attempt after a caller-owned identity — fork creation uses the
+    /// `record.id` it is about to insert, which is what ties the object back to
+    /// the exact row rather than to the logical owner/name.
+    pub fn from_owned(id: impl Into<String>) -> Self {
+        Self(id.into())
+    }
+
+    pub fn as_str(&self) -> &str {
+        &self.0
+    }
+}
+
+impl Default for PublishAttemptId {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl std::fmt::Display for PublishAttemptId {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(&self.0)
+    }
+}
+
+/// The object user-metadata key an attempt id travels in.
+///
+/// S3 lowercases user-metadata keys, so this must already be lowercase or the
+/// read back would never match what was written.
+pub const ATTEMPT_METADATA_KEY: &str = "gitlawb-attempt";
+
+/// The precondition an upload is fenced on.
+///
+/// Object storage is the only place a fence can hold. Dropping the future of an
+/// in-flight PUT does not cancel the request the server is already processing,
+/// so no amount of local locking stops an abandoned writer's bytes from landing
+/// after a successor has published. A conditional PUT the store itself refuses
+/// is what actually stops it.
+#[derive(Clone, Debug)]
+pub enum UploadPrecondition {
+    /// Publish only if the stored object is still the generation we observed.
+    IfMatch(String),
+    /// Publish only if nothing is stored under the key yet.
+    IfAbsent,
+    /// No fence. Last writer wins.
+    Unconditional,
+}
+
+/// What the store holds under a key right now.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct StoredGeneration {
+    /// The generation a later conditional upload can fence itself on.
+    pub etag: Option<String>,
+    /// The attempt that published it, when the object carries our metadata.
+    /// `None` for anything written by something that does not stamp attempts
+    /// (an operator upload, a pre-attempt-id archive).
+    pub attempt: Option<String>,
+}
+
+impl StoredGeneration {
+    /// Is what the store holds the work of `attempt`?
+    ///
+    /// This is the whole reconciliation primitive. It is deliberately a
+    /// three-way question collapsed to a boolean only at the point of use: an
+    /// object with no attempt metadata answers `false`, because "somebody else's
+    /// bytes" and "bytes nobody stamped" license exactly the same caution.
+    pub fn belongs_to(&self, attempt: &PublishAttemptId) -> bool {
+        self.attempt.as_deref() == Some(attempt.as_str())
+    }
+}
+
+/// The receipt of a publish the store ACKNOWLEDGED.
+///
+/// Only produced on a response the client actually read, so holding one is proof
+/// of publication in a way that "the upload future returned" is not.
+#[derive(Clone, Debug)]
+pub struct UploadReceipt {
+    pub attempt: PublishAttemptId,
+    pub etag: Option<String>,
+}
+
+/// How far ONE publish attempt got.
+///
+/// Observable through [`PublishStageCell`] from outside the future doing the
+/// work, which is the property the whole design turns on: a handler future that
+/// is DROPPED never returns an outcome, so the only way to classify a cancelled
+/// attempt is to read the stage it had reached when it was abandoned.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub enum PublishStage {
+    /// Nothing has been attempted.
+    Idle,
+    /// No object-storage backend is configured, so there is no publication to
+    /// confirm and the local write IS the durable copy. Distinct from `Idle`,
+    /// which means an attempt was possible and had not started: a cancellation
+    /// under `NoBackend` loses nothing, while a cancellation under `Idle` means
+    /// a publish that could have happened never did.
+    NoBackend,
+    /// The archive is being built. The request has not been constructed, let
+    /// alone sent. A cancellation here is a DEFINITE "no publication was
+    /// attempted" — not an ambiguous in-flight PUT.
+    PreparingArchive,
+    /// The request is on the wire. Its fate is unknown until the store answers,
+    /// and a cancellation here does NOT stop the server processing it.
+    PutDispatched { attempt: PublishAttemptId },
+    /// The store acknowledged the write. This is the ONLY stage that licenses
+    /// replication, a 2xx, or treating the local tree as durable.
+    Published {
+        attempt: PublishAttemptId,
+        etag: Option<String>,
+    },
+    /// The store definitively did not publish this attempt (a refused
+    /// precondition, or a failure that proves the bytes never committed).
+    Refused,
+    /// The attempt may or may not have committed and has not been reconciled.
+    /// Destructive compensation is NEVER licensed from here.
+    Ambiguous { attempt: PublishAttemptId },
+}
+
+impl PublishStage {
+    /// Did this attempt reach a state where the store may hold its bytes?
+    ///
+    /// True from dispatch onward. `false` is what makes deleting the attempt's
+    /// object or local tree safe.
+    pub fn may_have_published(&self) -> bool {
+        matches!(
+            self,
+            PublishStage::PutDispatched { .. }
+                | PublishStage::Published { .. }
+                | PublishStage::Ambiguous { .. }
+        )
+    }
+
+    /// The attempt whose fate is unresolved, when there is one. This is what a
+    /// reconciliation HEAD is compared against.
+    pub fn unresolved_attempt(&self) -> Option<&PublishAttemptId> {
+        match self {
+            PublishStage::PutDispatched { attempt } | PublishStage::Ambiguous { attempt } => {
+                Some(attempt)
+            }
+            _ => None,
+        }
+    }
+}
+
+/// A [`PublishStage`] an in-flight upload writes and an outside observer reads.
+///
+/// `std::sync::Mutex` rather than an async lock on purpose: every write is a
+/// field assignment with no await inside, and the reader that matters most runs
+/// inside a `Drop` impl, where an async lock cannot be awaited at all.
+#[derive(Debug)]
+pub struct PublishStageCell(Mutex<PublishStage>);
+
+impl PublishStageCell {
+    pub fn new() -> Self {
+        Self::seeded(PublishStage::Idle)
+    }
+
+    pub fn seeded(stage: PublishStage) -> Self {
+        Self(Mutex::new(stage))
+    }
+
+    pub fn set(&self, stage: PublishStage) {
+        *self.0.lock().expect("publish stage mutex poisoned") = stage;
+    }
+
+    pub fn get(&self) -> PublishStage {
+        self.0.lock().expect("publish stage mutex poisoned").clone()
+    }
+}
+
+impl Default for PublishStageCell {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// What the client knows about whether a request it could not complete ever
+/// reached the wire.
+///
+/// The classifier that produces this is the ONLY backend-aware code in the
+/// publish path; a pluggable store (#79) supplies its own and every caller below
+/// keeps working off [`UploadError`].
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum DispatchKnowledge {
+    /// The request could not be built or sent. The store cannot hold these bytes.
+    NeverSent,
+    /// The request may have been sent, and may have committed.
+    MaybeSent,
+}
+
+/// Why an upload did not publish, split by what it leaves the caller ENTITLED TO
+/// DO — not by which HTTP status came back.
+///
+/// The split exists because the previous shape ("precondition lost" vs "some
+/// other error") made every caller treat a lost response as proof of failure.
+/// Smithy timeout, dispatch and response errors all allow that the request was
+/// sent and committed: a server can accept a complete conditional PUT and lose
+/// or corrupt the response before the client observes success. Compensating that
+/// as a definite failure deletes state that IS published.
+#[derive(Debug, thiserror::Error)]
+pub enum UploadError {
+    /// The store explicitly refused the fence (412, or 409 under create-only).
+    /// DEFINITE: this attempt did not publish, and a successor did.
+    #[error("upload precondition lost (HTTP {status})")]
+    PreconditionLost { status: u16 },
+    /// The attempt provably never committed: it failed before the request could
+    /// be dispatched, or the store answered a definite client-side refusal.
+    /// Destructive compensation is safe here and ONLY here.
+    #[error("upload did not reach the store: {0:#}")]
+    NotPublished(#[source] anyhow::Error),
+    /// The request may have been dispatched and may have committed; the client
+    /// never learned which. The attempt id is carried so the caller can ask the
+    /// store rather than guess.
+    #[error("upload outcome is unknowable (attempt {attempt}): {source:#}")]
+    Ambiguous {
+        attempt: PublishAttemptId,
+        #[source]
+        source: anyhow::Error,
+    },
+}
+
+impl UploadError {
+    /// May this caller destroy state that would be needed if the write HAD
+    /// landed — delete the object, drop the only local clone, invalidate the
+    /// cache?
+    ///
+    /// Only a proven non-publication says yes. Every new variant must default to
+    /// `false`, which is why this is a match on the safe arms rather than a
+    /// negation of the unsafe one.
+    pub fn proves_not_published(&self) -> bool {
+        matches!(
+            self,
+            UploadError::PreconditionLost { .. } | UploadError::NotPublished(_)
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn only_definite_outcomes_license_destructive_compensation() {
+        assert!(UploadError::PreconditionLost { status: 412 }.proves_not_published());
+        assert!(
+            UploadError::NotPublished(anyhow::anyhow!("no route to host")).proves_not_published()
+        );
+        assert!(
+            !UploadError::Ambiguous {
+                attempt: PublishAttemptId::new(),
+                source: anyhow::anyhow!("response body truncated"),
+            }
+            .proves_not_published(),
+            "a response-loss failure must never license deleting state the write may own"
+        );
+    }
+
+    #[test]
+    fn a_stage_may_have_published_only_from_dispatch_onward() {
+        let attempt = PublishAttemptId::new();
+        assert!(!PublishStage::Idle.may_have_published());
+        assert!(!PublishStage::NoBackend.may_have_published());
+        assert!(
+            !PublishStage::PreparingArchive.may_have_published(),
+            "compression has not constructed a request, let alone sent one"
+        );
+        assert!(!PublishStage::Refused.may_have_published());
+        assert!(PublishStage::PutDispatched {
+            attempt: attempt.clone()
+        }
+        .may_have_published());
+        assert!(PublishStage::Ambiguous {
+            attempt: attempt.clone()
+        }
+        .may_have_published());
+        assert!(PublishStage::Published {
+            attempt,
+            etag: None
+        }
+        .may_have_published());
+    }
+
+    #[test]
+    fn an_unstamped_object_never_belongs_to_an_attempt() {
+        let attempt = PublishAttemptId::new();
+        assert!(StoredGeneration {
+            etag: Some("\"e\"".into()),
+            attempt: Some(attempt.as_str().to_string()),
+        }
+        .belongs_to(&attempt));
+        assert!(
+            !StoredGeneration {
+                etag: Some("\"e\"".into()),
+                attempt: None,
+            }
+            .belongs_to(&attempt),
+            "bytes nobody stamped must not be claimed by this attempt"
+        );
+        assert!(!StoredGeneration {
+            etag: Some("\"e\"".into()),
+            attempt: Some("someone-else".into()),
+        }
+        .belongs_to(&attempt));
+    }
+}
@@ -19,7 +19,7 @@
 //! objects the reachable withheld set never classified, so subtracting that set
 //! is not enough — a dangling private blob would slip through (#99). The caller
 //! signals a full scan via [`PinCandidateSet::full_scan`] and must then apply
-//! the fail-closed blob filter (`visibility_pack::replicable_objects_fail_closed`)
+//! the fail-closed blob-and-tree filter (`visibility_pack::replicable_objects_fail_closed`)
 //! instead of the plain reachable-only subtraction.
 //!
 //! Because the pin candidate set needs only the OID *set* (never the per-path
@@ -185,10 +185,13 @@ fn rev_list_delta(
 /// List every object in the repository via
 /// `git cat-file --batch-all-objects --batch-check='%(objectname)'`.
 ///
-/// This is the whole-repo enumeration the push path falls back to and the
-/// reconciliation sweep relies on. It returns *all* objects (including
-/// unreachable/dangling ones), which is what the sweep needs to catch
-/// stragglers — do not swap it for a reachability walk.
+/// This is the whole-repo enumeration the push path falls back to. It
+/// returns *all* objects (including unreachable/dangling ones), which is
+/// what the push path needs to catch stragglers — do not swap it for a
+/// reachability walk. The reconciliation sweep no longer uses it: sweep
+/// candidates come from the windowed walk (window commits plus walked
+/// pairs), which are reachable by construction, so dangling objects are
+/// absent rather than filtered.
 pub fn list_all_objects(repo_path: &Path, git_bin: &str, deadline: Instant) -> Result<Vec<String>> {
    let out = crate::git::visibility_pack::run_bounded_git(
        git_bin,
@@ -211,8 +214,10 @@ pub fn list_all_objects(repo_path: &Path, git_bin: &str, deadline: Instant) -> R

 /// Like [`list_all_objects`] but pairs each OID with its object type, via
 /// `--batch-check='%(objectname) %(objecttype)'`. The pin path's fail-closed
-/// filter needs to tell blobs (content, withholdable) from commits/trees
-/// (structural, never withheld) without typing the candidate list itself.
+/// filter needs to tell blobs (content, withholdable) and trees (content,
+/// withholdable since #135/#173) from commits/tags (structural, never withheld)
+/// without typing the candidate list itself.
+#[allow(dead_code)] // used by tests and all_blob_oids
 pub fn list_all_objects_with_type(
    repo_path: &Path,
    git_bin: &str,
@@ -244,24 +249,49 @@ pub fn list_all_objects_with_type(
 /// Every blob OID in the repo, including unreachable/dangling ones. The
 /// fail-closed pin filter drops any candidate blob absent from the reachable,
 /// visibility-allowed set; a dangling private blob is in this set but not the
-/// allowed set, so it never replicates (#99).
+/// allowed set, so it never replicates (#99). Production uses
+/// [`all_blob_and_tree_oids`] (one typed pass); this wrapper stays for tests.
+#[cfg(test)]
 pub fn all_blob_oids(
    repo_path: &Path,
    git_bin: &str,
    deadline: Instant,
 ) -> Result<HashSet<String>> {
-    Ok(list_all_objects_with_type(repo_path, git_bin, deadline)?
-        .into_iter()
-        .filter(|(_, ty)| ty == "blob")
-        .map(|(oid, _)| oid)
-        .collect())
+    Ok(all_blob_and_tree_oids(repo_path, git_bin, deadline)?.0)
+}
+
+/// Every blob OID and every tree OID in the repo, including unreachable/dangling
+/// ones, from one typed enumeration. The fail-closed pin filter drops any
+/// candidate blob or tree absent from the matching reachable visibility-allowed
+/// set; a dangling or withheld object is in this universe but not the allowed
+/// set, so it never replicates (#99, #172).
+#[allow(dead_code)] // production went to the typed sweep path; tests still use it
+pub fn all_blob_and_tree_oids(
+    repo_path: &Path,
+    git_bin: &str,
+    deadline: Instant,
+) -> Result<(HashSet<String>, HashSet<String>)> {
+    let mut blobs = HashSet::new();
+    let mut trees = HashSet::new();
+    for (oid, ty) in list_all_objects_with_type(repo_path, git_bin, deadline)? {
+        match ty.as_str() {
+            "blob" => {
+                blobs.insert(oid);
+            }
+            "tree" => {
+                trees.insert(oid);
+            }
+            _ => {}
+        }
+    }
+    Ok((blobs, trees))
 }

 /// The pin candidate OIDs for a push plus whether they came from a full-repo
 /// scan. `full_scan` is true when the delta could not be used and the whole
 /// object DB (including dangling objects) was enumerated — the caller must then
-/// apply the fail-closed blob filter, because the reachable-only withheld set
-/// cannot classify dangling blobs (#99).
+/// apply the fail-closed blob-and-tree filter, because the reachable-only withheld set
+/// cannot classify dangling blobs or trees (#99, #172).
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct PinCandidateSet {
    pub candidates: Vec<String>,
@@ -278,10 +308,12 @@ pub struct PinCandidateSet {
 /// Every degraded path is **logged**, not silent: a full-scan fallback, a
 /// failed full scan, and a panicked blocking task each emit a warning. On a
 /// failed full scan or a task panic the candidate set is empty (pin nothing
-/// this push); that is a durability gap the reconciliation sweep backstops, and
-/// it can never leak because the withheld/fail-closed filter still runs on
-/// whatever set is returned. `full_scan` rides on the returned set so the caller
-/// knows when the dangling-inclusive filter is required.
+/// this push); that is a durability gap the reconciliation sweep backstops
+/// when it is enabled and a pin backend is configured (a node running with the
+/// sweep disabled or with no IPFS/Pinata backend has no backstop), and it can
+/// never leak because the withheld/fail-closed filter still runs on whatever
+/// set is returned. `full_scan` rides on the returned set so the caller knows
+/// when the dangling-inclusive filter is required.
 ///
 /// `scan_sem` is the post-receive scan admission pool (`git_encrypt_semaphore`,
 /// #174 F4): both git-spawning stages — the per-tip `cat-file` probe + delta
@@ -14,13 +14,13 @@ use tokio::process::Command;
 /// the work they admitted, so admission is released only when that work is truly
 /// done — not the instant the handler future drops on a client disconnect.
 ///
-/// A move-only wrapper: no methods beyond construction and `Drop`. The handler
-/// MOVEs its permits in and keeps no copy (a retained copy would drop early and
-/// release admission the moment the future is dropped, defeating the guard). It is
-/// threaded into `drive_git_child`, whose [`KillGroupOnDrop`] moves it into the
-/// detached reaper on disconnect, so both permits drop only after the process group
-/// is confirmed reaped (`kill(-pgid,0)==ESRCH`) rather than while the group is still
-/// alive holding PIDs past the concurrency cap (#174 P1-a, plain-spawn residual).
+/// A drop-only wrapper: nothing here inspects what it holds. The handler MOVEs its
+/// permits in and keeps no copy (a retained copy would drop early and release admission
+/// the moment the future is dropped, defeating the guard). It is threaded into
+/// `drive_git_child`, whose [`KillGroupOnDrop`] moves it into the detached reaper on
+/// disconnect, so both permits drop only after the process group is confirmed reaped
+/// (`kill(-pgid,0)==ESRCH`) rather than while the group is still alive holding PIDs past
+/// the concurrency cap (#174 P1-a, plain-spawn residual).
 ///
 /// The `be0cdd6` path-scoped upload-pack walk already applies this discipline by
 /// moving its permits into the `spawn_blocking`; this generalizes it to the plain
@@ -33,6 +33,8 @@ pub struct AdmissionGuard {
    // 'static` so the guard can move into the detached reaper task.
    _global: Option<Box<dyn Send + 'static>>,
    _caller: Option<Box<dyn Send + 'static>>,
+    // Any further work-scoped hold that must outlive the process group; see `with_hold`.
+    _hold: Option<Box<dyn Send + 'static>>,
    // Per-repo write lease (#174 U2/F3), `Some` ONLY on the receive-pack write path
    // (via [`with_lease`](Self::with_lease)); `None` on every read path and every
    // non-receive-pack write path. It rides this guard into `KillGroupOnDrop`'s detached
@@ -54,10 +56,26 @@ impl AdmissionGuard {
        Self {
            _global: Some(Box::new(global)),
            _caller: caller.map(|c| Box::new(c) as Box<dyn Send + 'static>),
+            _hold: None,
            _lease: None,
        }
    }

+    /// Attach a further hold that must not be released until the process group is
+    /// reaped, and ride it through the same seam as the permits.
+    ///
+    /// The push handler uses this for the repo WRITE LOCK (#173 F2). Its
+    /// `guard.release(..)` line is only reached if `receive_pack` returns, so on a client
+    /// disconnect the lock used to be freed by the dropped future while the detached
+    /// reaper was still giving the group its SIGTERM grace, admitting a second
+    /// `receive-pack` on the same repo. Carrying the lock here holds it until the group
+    /// is ESRCH-confirmed gone, which is the same invariant the timeout path already
+    /// keeps ("a caller releasing a write lock can't race them", `reap_group_on_timeout`).
+    pub fn with_hold(mut self, hold: impl Send + 'static) -> Self {
+        self._hold = Some(Box::new(hold));
+        self
+    }
+
    /// Attach the per-repo write lease (#174 U2/F3). Called ONLY on the receive-pack
    /// write path, so the lease rides the disconnect reaper; read paths never call this.
    pub fn with_lease(mut self, lease: crate::state::RepoWriteLease) -> Self {
@@ -88,11 +106,11 @@ pub async fn info_refs(

    let mut command = Command::new(git_bin);
    command
-        .arg(service_to_command(service))
+        .arg(service_to_command(service)?)
        .arg("--stateless-rpc")
        .arg("--advertise-refs")
        .arg(repo_path);
-    // No request body — advertise-refs does not read stdin.
+    // No request body: advertise-refs does not read stdin.
    let (stdout, admission) =
        drive_git_child(command, Bytes::new(), timeout, "advertise-refs", admission).await?;
    // Single-stage op: the advertisement's group is reaped by now; release admission
@@ -202,6 +220,7 @@ pub fn response_served_pack(output: &[u8]) -> bool {
 ///
 /// Accepts a push. The caller MUST verify HTTP Signature auth before
 /// calling this function.
+#[allow(dead_code)] // used by tests; production uses receive_pack_raw
 pub async fn receive_pack(
    git_bin: &str,
    repo_path: &Path,
@@ -226,6 +245,219 @@ pub async fn receive_pack(
        .body(Body::from(output))?)
 }

+/// Run `git-receive-pack` and return the raw stdout bytes together with
+/// the process exit status. Unlike [`receive_pack`], this does NOT bail
+/// on a non-zero exit: the caller needs the stdout (which contains the
+/// report-status with per-ref ok/ng results) even when the process
+/// exits non-zero. A timeout still returns `Err`.
+///
+/// `reflog_action` (typically `gitlawb-request:<request_id>`) is
+/// exported as `GIT_REFLOG_ACTION` so the ref transaction writes
+/// request-bound landing evidence into the reflog. Recovery requires
+/// that evidence before auto-promoting a stranded row; without it a
+/// later request recreating the same old/new tuple is
+/// indistinguishable from the stranded one.
+#[allow(dead_code)]
+pub async fn receive_pack_raw(
+    git_bin: &str,
+    repo_path: &Path,
+    request_body: Bytes,
+    timeout: Duration,
+    admission: Option<AdmissionGuard>,
+) -> Result<(Vec<u8>, bool)> {
+    receive_pack_raw_with_reflog(git_bin, repo_path, request_body, timeout, admission, None).await
+}
+
+/// [`receive_pack_raw`] with explicit reflog binding.
+pub async fn receive_pack_raw_with_reflog(
+    git_bin: &str,
+    repo_path: &Path,
+    request_body: Bytes,
+    timeout: Duration,
+    admission: Option<AdmissionGuard>,
+    reflog_action: Option<&str>,
+) -> Result<(Vec<u8>, bool)> {
+    let mut command = Command::new(git_bin);
+    command
+        .arg("receive-pack")
+        .arg("--stateless-rpc")
+        .arg(repo_path);
+    if let Some(action) = reflog_action {
+        command.env("GIT_REFLOG_ACTION", action);
+    }
+    let (out, err, status, _admission) = drive_git_child_raw(
+        command,
+        request_body,
+        timeout,
+        "git-receive-pack",
+        admission,
+    )
+    .await?;
+    // On timeout, drive_git_child_raw returns Err — we never reach here.
+    // On success/non-zero exit, we have the stdout and exit status.
+    if !status.success() {
+        let stderr = String::from_utf8_lossy(&err);
+        tracing::warn!(stderr = %stderr, "git-receive-pack exited non-zero");
+    }
+    Ok((out, status.success()))
+}
+
+/// Parse the git-receive-pack report-status output to determine per-ref
+/// success/failure. Returns `(unpack_ok, per_ref_results)` where
+/// `per_ref_results` is a list of `(ref_name, is_ok)`.
+///
+/// The report-status format (after the sideband framing) is:
+/// ```text
+/// unpack ok\n          (or: unpack fail\n)
+/// ok <refname>\n       (per successful ref)
+/// ng <refname> <reason>\n  (per rejected ref)
+/// \n                   (empty line terminates)
+/// ```
+///
+/// Returns `None` if the output cannot be parsed (e.g. the client did
+/// not request report-status, or the output is truncated). In that
+/// case the caller should treat all refs as uncertain.
+pub fn parse_report_status(output: &[u8]) -> Option<(bool, Vec<(String, bool)>)> {
+    let text = std::str::from_utf8(output).ok()?;
+    // Strip sideband framing: each line starts with a pkt-line length
+    // prefix and a channel byte (1 = stdout, 2 = stderr). The actual
+    // data starts after the first `0000` flush packet or after we
+    // strip sideband bytes.
+    //
+    // Framing completeness is part of authority, at whichever level the
+    // framing lives. A double-framed capture carries the report stream
+    // (itself pkt-line framed, flush-terminated) inside the sideband
+    // envelope — real git output has been observed with NO outer flush,
+    // so the outer level cannot require one; the inner frame's flush is
+    // authoritative there. A single-framed (or unframed) payload has no
+    // inner frame, so its only completion marker is the outer flush:
+    // output ending after a complete `unpack ok` / `ok <ref>` packet but
+    // before the outer `0000` is a truncated report and yields `None`
+    // (uncertain), not a successful report.
+    let (stripped, outer_flush) = strip_sideband(text, false)?;
+    // The report-status vocabulary (`unpack ...`, `ok ...`, `ng ...`)
+    // never opens with four hex digits, so a hex quartet here is inner
+    // pkt-line framing, not report text.
+    let stripped = if stripped.len() >= 4
+        && stripped.as_bytes()[..4]
+            .iter()
+            .all(|b| b.is_ascii_hexdigit())
+    {
+        strip_sideband(&stripped, true)?.0
+    } else if !outer_flush {
+        return None;
+    } else {
+        stripped
+    };
+    // The report is framed TWICE when the client negotiated side-band-64k,
+    // which `git push` over smart HTTP does — so this is the common case, not
+    // an exotic one. The outer frame is the side-band envelope; band 1 carries
+    // the report-status stream, which is ITSELF pkt-line encoded. Real bytes
+    // from git 2.50.1 rejecting one ref of a two-ref push:
+    //
+    //   0057\x01000eunpack ok\n0028ng refs/heads/main non-fast-forward\n...
+    //
+    // After one pass the first line reads `000eunpack ok`, which fails the
+    // `unpack ok` check below and makes this return None — i.e. "no report",
+    // which the caller treats as inconclusive and keeps every declared ref.
+    // The per-ref gate would then be inert for exactly the pushes it exists to
+    // filter. The second pass above already removed the inner pkt-line
+    // framing where present (and required its flush); it is a no-op
+    // on the single-framed shape, because plain report text does not begin
+    // with four hex digits.
+    let lines: Vec<&str> = stripped.lines().collect();
+    if lines.is_empty() {
+        return None;
+    }
+
+    // First non-empty line is "unpack ok" or "unpack fail".
+    let unpack_line = lines.iter().find(|l| !l.is_empty())?;
+    let unpack_ok = if unpack_line.starts_with("unpack ok") {
+        true
+    } else if unpack_line.starts_with("unpack fail") {
+        false
+    } else {
+        return None;
+    };
+
+    let mut results = Vec::new();
+    for line in &lines[1..] {
+        let line = line.trim();
+        if line.is_empty() {
+            break;
+        }
+        if let Some(rest) = line.strip_prefix("ok ") {
+            results.push((rest.to_string(), true));
+        } else if let Some(rest) = line.strip_prefix("ng ") {
+            // "ng <refname> <reason>" — skip the reason
+            let ref_name = rest.split_whitespace().next()?.to_string();
+            results.push((ref_name, false));
+        }
+    }
+
+    Some((unpack_ok, results))
+}
+
+/// Strip git sideband framing from a pkt-line encoded output.
+/// Sideband-encoded lines start with a 4-hex-digit length, then a
+/// channel byte (0x01=stdout, 0x02=stderr), then payload. Returns the
+/// decoded payload plus whether the stream saw its terminating flush
+/// packet, or `None` if the framing is malformed. With `require_flush`,
+/// a stream that never sees its flush is also `None` — the caller uses
+/// this for the level whose completion marker is authoritative (see
+/// [`parse_report_status`]).
+fn strip_sideband(text: &str, require_flush: bool) -> Option<(String, bool)> {
+    let mut output = String::new();
+    let mut pos = 0;
+    let bytes = text.as_bytes();
+    let mut saw_flush = false;
+
+    loop {
+        if pos == bytes.len() {
+            break;
+        }
+        if pos + 4 > bytes.len() {
+            // Truncated framing mid-header: fail closed.
+            return None;
+        }
+        let len_str = std::str::from_utf8(&bytes[pos..pos + 4]).ok()?;
+        let len = usize::from_str_radix(len_str, 16).ok()?;
+        if len == 0 {
+            // Flush packet — end of sideband stream
+            saw_flush = true;
+            break;
+        }
+        if len < 4 || pos + len > bytes.len() {
+            // Truncated pkt-line extends past EOF: incomplete framing.
+            return None;
+        }
+        let pkt_data = std::str::from_utf8(&bytes[pos + 4..pos + len]).ok()?;
+        pos += len;
+
+        // Sideband: first byte is channel (1=stdout, 2=stderr)
+        if let Some(payload) = pkt_data.strip_prefix('\x01') {
+            output.push_str(payload);
+        } else if pkt_data.starts_with('\x02') {
+            // stderr — skip (git error messages)
+        } else {
+            // Not sideband encoded — pass through
+            output.push_str(pkt_data);
+        }
+    }
+
+    if require_flush && !saw_flush {
+        // Truncated at a packet boundary: every packet parsed, but the
+        // stream never completed. Indistinguishable from a report git was
+        // killed mid-write, so no authority.
+        return None;
+    }
+    if output.is_empty() {
+        None
+    } else {
+        Some((output, saw_flush))
+    }
+}
+
 /// Sends SIGTERM to a child's whole process group on drop, unless disarmed first.
 ///
 /// A served `git upload-pack`/`receive-pack` forks helpers such as `pack-objects`.
@@ -408,7 +640,7 @@ async fn run_git_service(
 ) -> Result<Vec<u8>> {
    let mut command = Command::new(git_bin);
    command
-        .arg(service_to_command(service))
+        .arg(service_to_command(service)?)
        .arg("--stateless-rpc")
        .arg(repo_path);
    let (out, admission) = drive_git_child(command, input, timeout, service, admission).await?;
@@ -560,11 +792,128 @@ async fn drive_git_child(
    Ok((out, admission))
 }

-fn service_to_command(service: &str) -> &str {
+/// Like [`drive_git_child`], but returns stdout and stderr even on a
+/// non-zero exit status. Used by [`receive_pack_raw`] so the caller
+/// can parse the report-status output from a failed `git-receive-pack`.
+async fn drive_git_child_raw(
+    mut command: Command,
+    input: Bytes,
+    timeout: Duration,
+    what: &str,
+    admission: Option<AdmissionGuard>,
+) -> Result<(
+    Vec<u8>,
+    Vec<u8>,
+    std::process::ExitStatus,
+    Option<AdmissionGuard>,
+)> {
+    command
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped());
+
+    #[cfg(unix)]
+    command.process_group(0);
+
+    let mut child = command.spawn()?;
+
+    let mut stdin = child.stdin.take();
+    let mut stdout = child.stdout.take().context("git stdout was not piped")?;
+    let mut stderr = child.stderr.take().context("git stderr was not piped")?;
+
+    #[cfg(unix)]
+    let pgid = child.id().map(|id| id as i32);
+    #[cfg(unix)]
+    let mut group_guard = KillGroupOnDrop {
+        child: Some(child),
+        pgid,
+        admission,
+    };
+
+    let mut out = Vec::new();
+    let mut err = Vec::new();
+
+    let interact = async {
+        let write = async {
+            match stdin.take() {
+                Some(mut s) => s.write_all(&input).await,
+                None => Ok(()),
+            }
+        };
+        #[cfg(unix)]
+        let child_ref = group_guard.child_mut();
+        #[cfg(not(unix))]
+        let child_ref = &mut child;
+        let (write_result, r_out, r_err, status) = tokio::join!(
+            write,
+            stdout.read_to_end(&mut out),
+            stderr.read_to_end(&mut err),
+            child_ref.wait(),
+        );
+        r_out?;
+        r_err?;
+        Ok::<_, anyhow::Error>((write_result, status?))
+    };
+
+    let timed = tokio::time::timeout(timeout, interact).await;
+    let (write_result, status, admission) = match timed {
+        Ok(result) => {
+            #[cfg(unix)]
+            let admission = group_guard.disarm();
+            let (write_result, status) = result?;
+            (write_result, status, admission)
+        }
+        Err(_elapsed) => {
+            #[cfg(unix)]
+            {
+                reap_group_on_timeout(group_guard.child_mut()).await;
+                drop(group_guard.disarm());
+            }
+            #[cfg(not(unix))]
+            {
+                let _ = child.start_kill();
+                let _ = tokio::time::timeout(Duration::from_secs(2), child.wait()).await;
+                drop(admission);
+            }
+            return Err(GitServiceTimeout.into());
+        }
+    };
+
+    // Prefer the child's collected answer over a stdin-write error: an early
+    // rejection (unpack failure, command-level deny) makes git print a complete
+    // report-status and exit WITHOUT draining stdin, so a body larger than the
+    // pipe buffer EPIPEs the write while the answer is already on stdout. The
+    // child is reaped (the join above waited on it), so `out`/`err`/`status`
+    // are complete and authoritative — returning Err here would discard them:
+    // `receive_pack_raw_with_reflog` would propagate the write error and the
+    // handler would mark every prepared row `uncertain` instead of driving
+    // `cancelled` per the report. That durable fate is the effect: on a
+    // non-zero exit the client still gets `Err(AppError::Git)` (the handler
+    // maps the failed run itself, not this write), while on exit 0 an error
+    // here would answer a push git actually landed with a server error. The
+    // non-raw `drive_git_child` orders the same way (exit status and git's own
+    // failure first, stdin-write error last — pinned by
+    // `run_git_service_surfaces_git_stderr_over_a_stdin_epipe`), but it must
+    // fail on non-zero because its callers do not parse output from a failed
+    // child; the raw driver exists precisely to parse that output.
+    if let Err(e) = &write_result {
+        tracing::warn!(
+            err = %e,
+            what = %what,
+            "git stdin write failed after child exit; surfacing collected output"
+        );
+    }
+
+    Ok((out, err, status, admission))
+}
+
+/// Map a validated service name to its git subcommand. Only the two fixed
+/// subcommands can come out, so an unrecognized service can never reach git's argv.
+fn service_to_command(service: &str) -> Result<&'static str> {
    match service {
-        "git-upload-pack" => "upload-pack",
-        "git-receive-pack" => "receive-pack",
-        _ => service,
+        "git-upload-pack" => Ok("upload-pack"),
+        "git-receive-pack" => Ok("receive-pack"),
+        other => bail!("unknown git service: {other}"),
    }
 }

@@ -659,14 +1008,15 @@ pub async fn build_filtered_pack(
    command
        .args(["pack-objects", "--stdout"])
        .current_dir(repo_path);
-    drive_git_child(
+    let (out, admission) = drive_git_child(
        command,
        Bytes::from(data),
        deadline.saturating_duration_since(Instant::now()),
        "pack-objects",
        admission,
    )
-    .await
+    .await?;
+    Ok((out, admission))
 }

 /// Serve a clone/fetch with the withheld blobs removed from the response pack.
@@ -847,6 +1197,104 @@ mod tests {
    use std::process::Command;
    use tempfile::TempDir;

+    /// Byte-for-byte capture from `git receive-pack` 2.50.1 rejecting one ref of
+    /// a two-ref push, with side-band-64k negotiated — what `git push` over
+    /// smart HTTP actually sends. The report is framed twice: the outer
+    /// side-band envelope, then the report-status stream's own pkt-lines.
+    ///
+    /// This fixture is the point of the test. A hand-written single-framed
+    /// string parses fine with only one strip pass, so a fixture invented to
+    /// match the parser hides the exact case the parser is for.
+    const REAL_SIDEBAND_REPORT: &[u8] =
+        b"0057\x01000eunpack ok\n0028ng refs/heads/main non-fast-forward\n0018ok refs/heads/third\n0000";
+
+    #[test]
+    fn a_real_double_framed_sideband_report_is_read_not_treated_as_absent() {
+        let (unpack_ok, refs) =
+            parse_report_status(REAL_SIDEBAND_REPORT).expect("a real git report must parse");
+        assert!(unpack_ok, "unpack line must be read through both frames");
+        assert_eq!(
+            refs,
+            vec![
+                ("refs/heads/main".to_string(), false),
+                ("refs/heads/third".to_string(), true),
+            ],
+            "the rejected ref must be distinguished from the accepted one"
+        );
+    }
+
+    /// The single-framed shape must keep working: the second pass has to be a
+    /// no-op there, not a corruption. Plain report text does not begin with
+    /// four hex digits, which is what makes that safe.
+    #[test]
+    fn a_single_framed_report_still_parses_after_the_second_pass() {
+        // Framed programmatically rather than by a hand-counted hex prefix: a
+        // wrong length makes the parser return None, which would look exactly
+        // like the regression this pair of tests exists to catch.
+        let payload = "\x01unpack ok\nok refs/heads/x\n";
+        let single = format!("{:04x}{payload}0000", payload.len() + 4);
+        let (unpack_ok, refs) =
+            parse_report_status(single.as_bytes()).expect("single-framed must parse");
+        assert!(unpack_ok);
+        assert_eq!(refs, vec![("refs/heads/x".to_string(), true)]);
+    }
+
+    #[test]
+    fn truncated_mid_packet_report_is_indeterminate_not_partial() {
+        // Truncate a valid single-framed report mid-packet (overrun):
+        // must return None so the handler persists the whole request as
+        // indeterminate instead of committing a partial ok set.
+        let payload = "\x01unpack ok\nok refs/heads/x\n";
+        let single = format!("{:04x}{payload}0000", payload.len() + 4);
+        let truncated = &single.as_bytes()[..single.len() - 6];
+        assert!(
+            parse_report_status(truncated).is_none(),
+            "mid-packet truncation must be indeterminate"
+        );
+    }
+
+    /// Truncation exactly at a packet boundary (every packet complete, but
+    /// the outer flush never arrived — e.g. git killed after the last `ok`
+    /// line) must also be indeterminate: the payload is complete but the
+    /// frame is not, and framing completeness is part of authority. The
+    /// mid-packet sibling above cannot catch this; only the flush gate can.
+    /// Removing the `require_flush` enforcement returns `Some` and turns
+    /// this red.
+    #[test]
+    fn truncated_at_packet_boundary_report_is_indeterminate_not_partial() {
+        let payload = "\x01unpack ok\nok refs/heads/x\n";
+        let single = format!("{:04x}{payload}0000", payload.len() + 4);
+        let no_flush = &single.as_bytes()[..single.len() - 4];
+        assert!(
+            parse_report_status(no_flush).is_none(),
+            "a report missing its outer flush must be indeterminate, \
+             even with every packet intact"
+        );
+        // Sanity: the untruncated fixture still parses (flush present).
+        assert!(
+            parse_report_status(single.as_bytes()).is_some(),
+            "the same bytes with the flush must parse"
+        );
+    }
+
+    #[test]
+    fn service_to_command_maps_allowlist_and_rejects_passthrough() {
+        assert_eq!(
+            service_to_command("git-upload-pack").unwrap(),
+            "upload-pack"
+        );
+        assert_eq!(
+            service_to_command("git-receive-pack").unwrap(),
+            "receive-pack"
+        );
+        for bad in ["git-fetch-pack", "--version", "upload-pack", ""] {
+            assert!(
+                service_to_command(bad).is_err(),
+                "must reject passthrough for {bad:?}"
+            );
+        }
+    }
+
    /// List OIDs in a pack by writing it to a temp dir and running verify-pack.
    pub(super) fn pack_object_ids(pack: &[u8]) -> std::collections::HashSet<String> {
        let dir = TempDir::new().unwrap();
@@ -1880,6 +2328,143 @@ mod tests {
        );
    }

+    // #174 U1 (R2, KTD3, RED-before/GREEN-after): the path-scoped filtered-pack serve
+    // must hold read + per-caller admission until its pack-objects process group is
+    // reaped on a client disconnect, exactly as the plain upload_pack path does. Before
+    // the fix build_filtered_pack took no AdmissionGuard and the handler's `_hold`
+    // permits dropped the instant the request future was dropped, so disconnect-spam on
+    // a path-scoped repo could hold PIDs past the concurrency cap while the permits were
+    // already free (#174 P1-a, on the filtered path the plain path had already closed).
+    //
+    // A real AdmissionGuard built from two owned semaphore permits rides
+    // rev-list -> pack-objects. We drive the future until pack-objects has forked its
+    // grandchild (the streaming pack-writer stand-in), assert the permits are still held
+    // mid-serve, then DROP the future (client disconnect) and assert the permits are
+    // released only AFTER the group is ESRCH-confirmed gone — never while it is alive.
+    // Goes RED if the guard is not threaded into the pack-objects stage (it would then
+    // drop after rev-list, freeing the permits mid-serve or on the bare future drop).
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn filtered_pack_holds_admission_until_group_reaped_on_disconnect() {
+        use std::sync::Arc;
+        use tokio::sync::Semaphore;
+
+        let tmp = tempfile::TempDir::new().unwrap();
+        let pidfile = tmp.path().join("pids");
+        // rev-list returns one oid fast; pack-objects forks a grandchild (the streaming
+        // writer stand-in), records leader+grandchild pids, then hangs so the future
+        // parks mid-serve with the guard owned by the pack-objects KillGroupOnDrop. The
+        // grandchild inherits (holds open) the stdout pipe, so drive_git_child's
+        // read_to_end blocks and the future stays pending until we drop it.
+        let body = format!(
+            "#!/bin/sh\ncase \"$1\" in\n  rev-list) echo deadbeefdeadbeefdeadbeefdeadbeefdeadbeef ;;\n  pack-objects) sleep 300 &\nprintf '%s\\n%s\\n' \"$$\" \"$!\" > \"{}\"\nwait ;;\n  *) exit 1 ;;\nesac\n",
+            pidfile.display()
+        );
+        let git_bin = write_fake_git(tmp.path(), &body);
+        let withheld = HashSet::new();
+
+        // Retry the fake-git spawn race like the sibling disconnect test; each attempt
+        // gets a FRESH semaphore so a dropped losing attempt can't skew the winning
+        // attempt's permit accounting. Keep the winning attempt's future PENDING so the
+        // drop below exercises the client-disconnect teardown.
+        let (fut, sem, leader, grandchild) = {
+            let mut attempt = 0u64;
+            loop {
+                attempt += 1;
+                let _ = std::fs::remove_file(&pidfile);
+                // Semaphore(4): two owned permits model the handler's global-read +
+                // per-caller admission, leaving 2 available while the op is in flight.
+                let sem = Arc::new(Semaphore::new(4));
+                let g = sem.clone().try_acquire_owned().unwrap();
+                let c = sem.clone().try_acquire_owned().unwrap();
+                let admission = AdmissionGuard::new(g, Some(c));
+                let mut fut = Box::pin(build_filtered_pack(
+                    git_bin.to_str().unwrap(),
+                    tmp.path(),
+                    &withheld,
+                    Duration::from_secs(60),
+                    Some(admission),
+                ));
+                // Advance the future a slice at a time until the fake records its pids
+                // (i.e. pack-objects is running). `Ok(_)` means the future returned
+                // before the pidfile appeared (spawn error / early exit); stop polling
+                // then, since re-polling a completed future panics.
+                let mut pids = None;
+                for _ in 0..500 {
+                    let finished = tokio::time::timeout(Duration::from_millis(10), &mut fut)
+                        .await
+                        .is_ok();
+                    if let Some(p) = read_two_pids(&pidfile) {
+                        pids = Some(p);
+                        break;
+                    }
+                    if finished {
+                        break;
+                    }
+                }
+                match pids {
+                    Some((l, gch)) => break (fut, sem, l, gch),
+                    None => {
+                        // Transient spawn miss: drop the still-armed future so its guard
+                        // reaps anything that spawned (and returns the permits), then
+                        // back off before retrying.
+                        drop(fut);
+                        assert!(
+                            attempt < FAKE_GIT_RETRY_ATTEMPTS,
+                            "fake git failed to reach the pack-objects stage after \
+                             {FAKE_GIT_RETRY_ATTEMPTS} attempts (persistent failure, \
+                             not a transient parallel-runner miss)"
+                        );
+                        tokio::time::sleep(Duration::from_millis(
+                            FAKE_GIT_BACKOFF_STEP_MS * attempt,
+                        ))
+                        .await;
+                    }
+                }
+            }
+        };
+        let _cleanup = ReapOnPanic(vec![leader, grandchild]);
+        assert!(alive(grandchild), "grandchild must be running mid-serve");
+
+        // Mid-serve: the pack-objects stage owns the guard, so the two permits are still
+        // held. A build that dropped the guard after rev-list (unthreaded pack-objects
+        // stage) would have freed them here.
+        assert_eq!(
+            sem.available_permits(),
+            2,
+            "admission permits must be held while the filtered serve is in flight"
+        );
+
+        // Client disconnect: drop the request future. The pack-objects KillGroupOnDrop
+        // must tear the group down AND hold the permits until the group is
+        // ESRCH-confirmed gone, releasing them only then.
+        drop(fut);
+
+        let mut released_while_alive = false;
+        let mut released_after_reap = false;
+        for _ in 0..500 {
+            let released = sem.available_permits() == 4;
+            let group_alive = alive(grandchild);
+            if released && group_alive {
+                released_while_alive = true;
+            }
+            if released && !group_alive {
+                released_after_reap = true;
+                break;
+            }
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+        assert!(
+            !released_while_alive,
+            "admission permits were released while the git group was still alive — the \
+             path-scoped concurrency-cap bypass (#174 P1-a) is open"
+        );
+        assert!(
+            released_after_reap,
+            "admission permits must be released once the group is reaped on disconnect"
+        );
+    }
+
    // ── F1: filtered-serve admission threaded through both pack stages ──────
    //
    // The handler-level disconnect regression lives in api/repos.rs
@@ -2465,6 +3050,62 @@ mod tests {
        );
    }

+    // The raw driver's counterpart: a rejection-shaped early exit where git
+    // prints a COMPLETE report-status and exits without draining stdin, so a
+    // body larger than the pipe buffer (~64 KiB) EPIPEs the write while the
+    // answer is already on stdout. `receive_pack_raw` must surface that report
+    // (Ok + stdout + exit status), not discard it as a stdin-write error — an
+    // Err here loses parseable output: `receive_pack_raw_with_reflog`
+    // propagates it and the handler marks every prepared row `uncertain`
+    // instead of driving `cancelled` from the report (the durable fate is the
+    // effect; the client-visible status on a failed run is `AppError::Git`
+    // either way). Goes RED if the stdin-write check returns Err after the
+    // join. Both exit codes are pinned: receive-pack rejections and successes
+    // both carry reports the handler must parse.
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn receive_pack_raw_surfaces_report_status_over_a_stdin_epipe() {
+        for exit_code in [0, 1] {
+            let tmp = tempfile::TempDir::new().unwrap();
+            std::fs::write(tmp.path().join("report.bin"), REAL_SIDEBAND_REPORT).unwrap();
+            // Never reads stdin: the pending body blocks the pipe, then EPIPEs
+            // once the fake exits. `cat` delivers the report first.
+            let git_bin = write_fake_git(
+                tmp.path(),
+                &format!("#!/bin/sh\ncat \"$(dirname \"$0\")/report.bin\"\nexit {exit_code}\n"),
+            );
+            let body = Bytes::from(vec![0u8; 256 * 1024]);
+            let (out, ok) = receive_pack_raw(
+                git_bin.to_str().unwrap(),
+                tmp.path(),
+                body,
+                Duration::from_secs(60),
+                None,
+            )
+            .await
+            .expect(
+                "a report-status printed before exit must not be discarded as a \
+                 stdin-write error",
+            );
+            assert_eq!(
+                ok,
+                exit_code == 0,
+                "the exit status rides the Ok return for both report shapes"
+            );
+            let (unpack_ok, results) = parse_report_status(&out)
+                .expect("git's report survives the EPIPE and stays parseable");
+            assert!(unpack_ok, "unpack ok parses");
+            assert_eq!(
+                results,
+                vec![
+                    ("refs/heads/main".to_string(), false),
+                    ("refs/heads/third".to_string(), true),
+                ],
+                "the per-ref report reaches the handler instead of an HTTP error"
+            );
+        }
+    }
+
    // ── #192 F1: response_served_pack detector (real git output) ────────────

    // A completing request (want+flush+done) yields NAK + a real packfile;
@@ -5,17 +5,121 @@

 use std::collections::HashMap;
 use std::path::{Path, PathBuf};
+use std::sync::atomic::AtomicBool;
 use std::sync::{Arc, Mutex, OnceLock};

 use anyhow::{Context, Result};
 use aws_sdk_s3::Client as S3Client;
-use tracing::{debug, info};
+use tracing::{debug, info, warn};
+
+// The publication vocabulary lives in `git::publish`, which carries no
+// object-storage types, and is re-exported here so existing `git::tigris::…`
+// imports keep resolving. #79 deletes this file; `git::publish` is what survives
+// the swap, and the only backend-aware code below is `classify_dispatch`.
+pub use super::publish::{
+    DispatchKnowledge, PublishAttemptId, PublishStage, PublishStageCell, StoredGeneration,
+    UploadError, UploadPrecondition, UploadReceipt, ATTEMPT_METADATA_KEY,
+};
+
+/// What happened to a conditional, attempt-guarded delete.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum AttemptDelete {
+    /// The object was this attempt's and is gone.
+    Deleted,
+    /// Nothing is stored under the key.
+    Absent,
+    /// Something is stored, and it is NOT this attempt's work. Left alone.
+    NotOurs,
+}
+
+/// THE ONLY BACKEND-AWARE CLASSIFIER in the publish path.
+///
+/// It answers one question: does this failure PROVE the request never committed?
+/// Anything short of proof is [`DispatchKnowledge::MaybeSent`], because the
+/// caller's next move on a definite failure is destructive (delete the object,
+/// drop the only local clone, invalidate the cache) and being wrong there
+/// destroys published state.
+///
+/// - An HTTP status means the server answered. A 4xx is a refusal it took
+///   BEFORE storing anything, so it proves non-publication. A 5xx does not: the
+///   store can commit and then fail to report it.
+/// - No status at all means no response was read. Only a construction failure
+///   proves the request never left; Smithy's timeout, dispatch and response
+///   errors all allow that the bytes arrived and committed while the answer was
+///   lost, corrupted, or never parsed.
+/// - `SdkError` is `#[non_exhaustive]`, so the fallback arm must be the CAUTIOUS
+///   one. A future variant defaulting to "definitely failed" would silently
+///   re-open exactly the class this closes.
+///
+/// #79 replaces `tigris.rs` with a `BlobStore` layer; this function is the piece
+/// each backend reimplements, and nothing above it changes.
+fn classify_put_failure<E, R>(
+    err: &aws_sdk_s3::error::SdkError<E, R>,
+    status: Option<u16>,
+) -> DispatchKnowledge {
+    if let Some(status) = status {
+        return if (400..500).contains(&status) {
+            DispatchKnowledge::NeverSent
+        } else {
+            DispatchKnowledge::MaybeSent
+        };
+    }
+    match err {
+        aws_sdk_s3::error::SdkError::ConstructionFailure(_) => DispatchKnowledge::NeverSent,
+        _ => DispatchKnowledge::MaybeSent,
+    }
+}

 /// Wrapper around the S3 client with the configured bucket.
 #[derive(Clone)]
 pub struct TigrisClient {
    s3: S3Client,
    bucket: String,
+    /// Test-only seam: when set, the blocking compression inside `upload_tracked`
+    /// parks on this barrier. That is the ONLY window in which a publish attempt
+    /// can be cancelled with `PublishStage::PreparingArchive` still holding, and
+    /// it is unreachable from outside without a seam because compression of a
+    /// test-sized repo finishes in microseconds.
+    #[cfg(test)]
+    compress_gate: Option<Arc<BlockingGate>>,
+}
+
+/// A gate a BLOCKING thread parks on until a test opens it.
+///
+/// A condvar rather than a held `MutexGuard`: the test's assertions run while the
+/// gate is shut, and holding a `std::sync::MutexGuard` across those awaits is
+/// both a lint violation and a real hazard. Here the test owns no guard at all —
+/// it flips a flag and notifies.
+#[cfg(test)]
+#[derive(Default)]
+pub struct BlockingGate {
+    open: Mutex<bool>,
+    opened: std::sync::Condvar,
+}
+
+#[cfg(test)]
+impl BlockingGate {
+    /// A gate that starts SHUT.
+    pub fn shut() -> Self {
+        Self::default()
+    }
+
+    fn wait(&self) {
+        let mut open = self.open.lock().expect("compression gate poisoned");
+        while !*open {
+            open = self
+                .opened
+                .wait(open)
+                .expect("compression gate poisoned while waiting");
+        }
+    }
+
+    /// Let every parked compression through. Call this at teardown so the
+    /// blocking thread is not stranded for the life of the process.
+    pub fn open(&self) {
+        *self.open.lock().expect("compression gate poisoned") = true;
+        self.opened.notify_all();
+    }
 }

 impl TigrisClient {
@@ -28,28 +132,48 @@ impl TigrisClient {
        Ok(Self {
            s3,
            bucket: bucket.to_string(),
+            #[cfg(test)]
+            compress_gate: None,
        })
    }

-    /// Test-only constructor with an explicit S3 endpoint, region, and static
-    /// credentials — no env-var reads, so parallel tests cannot race each other's
-    /// `AWS_*` environment the way the env-based `new` would. Lets a test point
-    /// the client at a non-routable endpoint to exercise acquire-stall paths.
+    /// Build a client pointed at an arbitrary endpoint, for tests.
+    ///
+    /// The production constructor reads the endpoint and credentials from the
+    /// environment, which a test cannot steer without mutating process-global
+    /// state. This takes both explicitly so a test can aim the client at a
+    /// closed port and get a prompt transport error out of `exists()`.
+    ///
+    /// `RetryConfig::disabled()` is load-bearing, not tidiness: the SDK's default
+    /// policy retries a connection refusal with backoff, which turns each failing
+    /// call into seconds of waiting.
    #[cfg(test)]
-    pub(crate) async fn for_testing_with_endpoint(bucket: &str, endpoint_url: &str) -> Self {
-        let creds = aws_sdk_s3::config::Credentials::new("test", "test", None, None, "test");
-        let config = aws_config::defaults(aws_config::BehaviorVersion::latest())
-            .endpoint_url(endpoint_url)
-            .region(aws_config::Region::new("auto"))
-            .credentials_provider(creds)
-            .load()
-            .await;
+    pub fn for_testing_with_endpoint(bucket: &str, endpoint: &str) -> Self {
+        use aws_sdk_s3::config::{retry::RetryConfig, Credentials, Region};
+
+        let config = aws_sdk_s3::config::Config::builder()
+            .endpoint_url(endpoint)
+            .credentials_provider(Credentials::new("test", "test", None, None, "test"))
+            .region(Region::new("auto"))
+            .retry_config(RetryConfig::disabled())
+            .behavior_version_latest()
+            .build();
        Self {
-            s3: S3Client::new(&config),
+            s3: S3Client::from_conf(config),
            bucket: bucket.to_string(),
+            compress_gate: None,
        }
    }

+    /// Test-only: park this client's blocking compression on `gate` until the
+    /// holder releases it, so a cancellation can be aimed at
+    /// [`PublishStage::PreparingArchive`] rather than at the PUT.
+    #[cfg(test)]
+    pub fn with_compress_gate(mut self, gate: Arc<BlockingGate>) -> Self {
+        self.compress_gate = Some(gate);
+        self
+    }
+
    /// S3 key for a given repo: `repos/v1/{owner_slug}/{repo_name}.tar.zst`
    fn repo_key(owner_slug: &str, repo_name: &str) -> String {
        format!("repos/v1/{owner_slug}/{repo_name}.tar.zst")
@@ -77,34 +201,325 @@ impl TigrisClient {
        }
    }

-    /// Upload a local bare repo directory to Tigris as a tar.zst archive.
-    pub async fn upload(&self, owner_slug: &str, repo_name: &str, local_path: &Path) -> Result<()> {
-        let key = Self::repo_key(owner_slug, repo_name);
-        debug!(key = %key, path = %local_path.display(), "uploading repo to tigris");
+    /// Read the ETag of a repo archive, or `None` when nothing is stored under
+    /// the key. The ETag identifies the generation a later conditional upload
+    /// can fence itself on.
+    ///
+    /// Separate from `exists` rather than folded into it: `exists` has callers
+    /// that only want the boolean, and widening its return type would churn
+    /// every one of them for no benefit.
+    pub async fn head_etag(&self, owner_slug: &str, repo_name: &str) -> Result<Option<String>> {
+        Ok(self
+            .head_generation(owner_slug, repo_name)
+            .await?
+            .map(|g| g.etag)
+            .unwrap_or(None))
+    }

-        // Create tar.zst in memory
-        let archive_bytes = tokio::task::spawn_blocking({
-            let local_path = local_path.to_path_buf();
-            move || compress_repo(&local_path)
-        })
+    /// The full stored generation: the ETag AND the attempt that published it.
+    ///
+    /// The attempt half is what makes a lost response recoverable. An ETag alone
+    /// answers "has the generation moved", which every writer sees the same way;
+    /// the attempt id answers "are the published bytes MINE", which is the
+    /// question a client whose PUT response vanished actually needs answered.
+    pub async fn head_generation(
+        &self,
+        owner_slug: &str,
+        repo_name: &str,
+    ) -> Result<Option<StoredGeneration>> {
+        let key = Self::repo_key(owner_slug, repo_name);
+        match self
+            .s3
+            .head_object()
+            .bucket(&self.bucket)
+            .key(&key)
+            .send()
+            .await
+        {
+            Ok(out) => {
+                let etag = out
+                    .e_tag()
+                    .context(format!("tigris HEAD {key}: hit carried no ETag"))?
+                    .to_string();
+                let attempt = out
+                    .metadata()
+                    .and_then(|m| m.get(ATTEMPT_METADATA_KEY))
+                    .cloned();
+                Ok(Some(StoredGeneration {
+                    etag: Some(etag),
+                    attempt,
+                }))
+            }
+            Err(e) => {
+                if e.as_service_error().is_some_and(|e| e.is_not_found()) {
+                    Ok(None)
+                } else {
+                    Err(anyhow::anyhow!("tigris HEAD {key}: {e}"))
+                }
+            }
+        }
+    }
+
+    /// Did `attempt`'s bytes land? The reconciliation an ambiguous dispatch owes
+    /// before anything downstream may treat it as confirmation.
+    ///
+    /// `Ok(false)` is deliberately NOT "the PUT failed": an abandoned request may
+    /// still be in flight, so a negative answer licenses refusing and retrying,
+    /// never deleting. Only `Ok(true)` upgrades an unresolved attempt to
+    /// published.
+    pub async fn attempt_landed(
+        &self,
+        owner_slug: &str,
+        repo_name: &str,
+        attempt: &PublishAttemptId,
+    ) -> Result<bool> {
+        Ok(self
+            .head_generation(owner_slug, repo_name)
+            .await?
+            .is_some_and(|g| g.belongs_to(attempt)))
+    }
+
+    /// Upload a local bare repo directory to Tigris as a tar.zst archive,
+    /// fenced by `precondition`, under a freshly minted attempt identity.
+    pub async fn upload(
+        &self,
+        owner_slug: &str,
+        repo_name: &str,
+        local_path: &Path,
+        precondition: UploadPrecondition,
+    ) -> std::result::Result<UploadReceipt, UploadError> {
+        self.upload_tracked(
+            owner_slug,
+            repo_name,
+            local_path,
+            precondition,
+            PublishAttemptId::new(),
+            None,
+        )
        .await
-        .context("tar task panicked")?
-        .context("compressing repo")?;
+    }
+
+    /// The full form: a caller-chosen attempt identity, and a stage cell the
+    /// upload reports its progress into.
+    ///
+    /// The stage cell is the answer to "a dropped future never returns an
+    /// outcome". Compression runs inside `spawn_blocking` and the conditional PUT
+    /// is not even constructed until it finishes, so a handler cancelled during
+    /// compression definitely never attempted publication — but nothing could
+    /// observe that, because the only report was the return value of a future
+    /// that no longer exists. Marking [`PublishStage::PreparingArchive`] before
+    /// the blocking call and [`PublishStage::PutDispatched`] immediately before
+    /// `send()` makes that boundary readable from outside.
+    pub async fn upload_tracked(
+        &self,
+        owner_slug: &str,
+        repo_name: &str,
+        local_path: &Path,
+        precondition: UploadPrecondition,
+        attempt: PublishAttemptId,
+        stage: Option<&PublishStageCell>,
+    ) -> std::result::Result<UploadReceipt, UploadError> {
+        let key = Self::repo_key(owner_slug, repo_name);
+        debug!(key = %key, path = %local_path.display(), attempt = %attempt, "uploading repo to tigris");
+
+        if let Some(stage) = stage {
+            stage.set(PublishStage::PreparingArchive);
+        }
+
+        // Create tar.zst in memory. Nothing has been sent at this point and
+        // nothing can be: the request below is not constructed until this
+        // returns. A failure or a cancellation here is a definite
+        // non-publication.
+        let compressed = {
+            let local_path = local_path.to_path_buf();
+            #[cfg(test)]
+            let gate = self.compress_gate.clone();
+            tokio::task::spawn_blocking(move || {
+                // Test-only seam: park INSIDE the blocking compression, which is
+                // the window a handler can be cancelled in while
+                // `PublishStage::PreparingArchive` still holds.
+                #[cfg(test)]
+                if let Some(gate) = gate {
+                    // Blocks until the test opens the gate.
+                    gate.wait();
+                }
+                compress_repo(&local_path)
+            })
+            .await
+        };
+        let archive_bytes = match compressed {
+            Ok(Ok(bytes)) => bytes,
+            Ok(Err(e)) => {
+                if let Some(stage) = stage {
+                    stage.set(PublishStage::Refused);
+                }
+                return Err(UploadError::NotPublished(e.context("compressing repo")));
+            }
+            Err(e) => {
+                if let Some(stage) = stage {
+                    stage.set(PublishStage::Refused);
+                }
+                return Err(UploadError::NotPublished(
+                    anyhow::Error::new(e).context("tar task panicked"),
+                ));
+            }
+        };

        let body = aws_sdk_s3::primitives::ByteStream::from(archive_bytes);

-        self.s3
+        let mut req = self
+            .s3
            .put_object()
            .bucket(&self.bucket)
            .key(&key)
+            // The attempt identity travels WITH the bytes. That is what makes a
+            // lost response recoverable: the client can HEAD the key afterwards
+            // and ask whether the published object is its own work, which is
+            // decidable, instead of asking whether its request succeeded, which
+            // is not.
+            .metadata(ATTEMPT_METADATA_KEY, attempt.as_str())
            .body(body)
-            .content_type("application/zstd")
-            .send()
-            .await
-            .context(format!("tigris PUT {key}"))?;
+            .content_type("application/zstd");
+        match &precondition {
+            UploadPrecondition::IfMatch(etag) => req = req.if_match(etag),
+            UploadPrecondition::IfAbsent => req = req.if_none_match("*"),
+            UploadPrecondition::Unconditional => {}
+        }

-        info!(key = %key, "uploaded repo to tigris");
-        Ok(())
+        // From HERE the bytes may reach the store. Everything downstream that
+        // could destroy state has to treat this stage as "maybe published".
+        if let Some(stage) = stage {
+            stage.set(PublishStage::PutDispatched {
+                attempt: attempt.clone(),
+            });
+        }
+
+        let sent = req.send().await;
+        let out = match sent {
+            Ok(out) => out,
+            Err(e) => {
+                // `PutObjectError` models no PreconditionFailed variant (its arms are
+                // EncryptionTypeMismatch, InvalidRequest, InvalidWriteOffset,
+                // TooManyParts, Unhandled), so a refused precondition arrives as
+                // `Unhandled` and matching the enum would classify it as a generic
+                // failure. The raw HTTP status off the response is the only place the
+                // answer actually lives.
+                //
+                // Read it via `raw_response()`, not a `ServiceError`-only match: the
+                // SDK exposes the raw response for BOTH `ServiceError` and
+                // `ResponseError`, and a refused conditional PUT whose error body the
+                // SDK cannot parse (malformed XML, premature close) surfaces as
+                // `ResponseError`. Matching only `ServiceError` would classify that
+                // unparsable 409/412 as a generic failure, and `RepoWriteGuard::release`
+                // would log-and-succeed instead of taking the supersede retry,
+                // acknowledging a write that was definitively not published.
+                let status = e.raw_response().map(|raw| raw.status().as_u16());
+                // 412 is always a lost precondition. 409 is one only when we asked
+                // for create-only, which is how S3-compatible stores report "the key
+                // already exists". Everything else, 404 included, is a real failure:
+                // archive keys are never deleted by the write path, so a 404 here
+                // means something permanent like a missing bucket or a misrouted
+                // endpoint, and reporting that as a lost precondition would tell a
+                // caller to expect a successor that does not exist.
+                let lost = match status {
+                    Some(412) => true,
+                    Some(409) => matches!(precondition, UploadPrecondition::IfAbsent),
+                    _ => false,
+                };
+                if lost {
+                    if let Some(stage) = stage {
+                        stage.set(PublishStage::Refused);
+                    }
+                    return Err(UploadError::PreconditionLost {
+                        status: status.expect("a lost precondition came from a status"),
+                    });
+                }
+                let knowledge = classify_put_failure(&e, status);
+                let ctx = anyhow::Error::new(e).context(format!("tigris PUT {key}"));
+                return Err(match knowledge {
+                    DispatchKnowledge::NeverSent => {
+                        if let Some(stage) = stage {
+                            stage.set(PublishStage::Refused);
+                        }
+                        UploadError::NotPublished(ctx)
+                    }
+                    DispatchKnowledge::MaybeSent => {
+                        warn!(
+                            key = %key,
+                            attempt = %attempt,
+                            status = ?status,
+                            "tigris PUT failed WITHOUT proving it did not commit — the outcome \
+                             is ambiguous and must be reconciled, not compensated"
+                        );
+                        if let Some(stage) = stage {
+                            stage.set(PublishStage::Ambiguous {
+                                attempt: attempt.clone(),
+                            });
+                        }
+                        UploadError::Ambiguous {
+                            attempt: attempt.clone(),
+                            source: ctx,
+                        }
+                    }
+                });
+            }
+        };
+
+        let etag = out.e_tag().map(str::to_string);
+        if let Some(stage) = stage {
+            stage.set(PublishStage::Published {
+                attempt: attempt.clone(),
+                etag: etag.clone(),
+            });
+        }
+        info!(key = %key, attempt = %attempt, "uploaded repo to tigris");
+        Ok(UploadReceipt { attempt, etag })
+    }
+
+    /// Delete a repo archive ONLY while it is still `attempt`'s work.
+    ///
+    /// The unconditional delete this replaces used the logical owner/name as its
+    /// cleanup authority, so a failed attempt compensating late could erase the
+    /// object a SUCCESSOR had already published under the same name and already
+    /// returned 201 for. A second name lookup before the delete only narrows that
+    /// window; reading the attempt id off the object and fencing the delete on
+    /// the generation it came from closes it.
+    pub async fn delete_if_attempt_matches(
+        &self,
+        owner_slug: &str,
+        repo_name: &str,
+        attempt: &PublishAttemptId,
+    ) -> Result<AttemptDelete> {
+        let Some(generation) = self.head_generation(owner_slug, repo_name).await? else {
+            return Ok(AttemptDelete::Absent);
+        };
+        if !generation.belongs_to(attempt) {
+            return Ok(AttemptDelete::NotOurs);
+        }
+        let key = Self::repo_key(owner_slug, repo_name);
+        let mut req = self.s3.delete_object().bucket(&self.bucket).key(&key);
+        // The If-Match guard is what makes this atomic rather than merely
+        // narrowed: between the HEAD above and this call a successor can publish,
+        // and the store refusing on the moved generation is the only thing that
+        // stops the delete landing on their object.
+        if let Some(etag) = generation.etag.as_deref() {
+            req = req.if_match(etag);
+        }
+        match req.send().await {
+            Ok(_) => Ok(AttemptDelete::Deleted),
+            Err(e) => {
+                // A refused conditional delete means the generation moved: the
+                // object is no longer ours, which is a successful outcome for a
+                // guard whose whole job is not to touch somebody else's bytes.
+                if e.raw_response()
+                    .map(|raw| raw.status().as_u16())
+                    .is_some_and(|s| s == 412 || s == 409)
+                {
+                    return Ok(AttemptDelete::NotOurs);
+                }
+                Err(anyhow::anyhow!("tigris conditional DELETE {key}: {e}"))
+            }
+        }
    }

    /// Download a repo archive from Tigris and extract to local disk.
@@ -112,10 +527,32 @@ impl TigrisClient {
        &self,
        owner_slug: &str,
        repo_name: &str,
-        local_path: &Path,
+        local_path: &super::repo_store::ValidatedRepoDiskPath,
+        swap_authority: Option<Arc<AtomicBool>>,
    ) -> Result<()> {
+        self.download_to(owner_slug, repo_name, local_path, true, swap_authority)
+            .await
+            .map(|_| ())
+    }
+
+    /// Download a repo archive from Tigris and extract it, returning the
+    /// directory that was populated.
+    ///
+    /// `publish` controls whether the extract is swapped into `target` in place
+    /// (the live-path mutation used by writes; returns `target`) or unpacked
+    /// into a fresh temp directory under `target`'s parent (a non-mutating
+    /// snapshot read; returns the temp dir, which the caller owns and cleans
+    /// up). The snapshot form never touches the live repo path.
+    pub async fn download_to(
+        &self,
+        owner_slug: &str,
+        repo_name: &str,
+        target: &super::repo_store::ValidatedRepoDiskPath,
+        publish: bool,
+        swap_authority: Option<Arc<AtomicBool>>,
+    ) -> Result<DownloadExtract> {
        let key = Self::repo_key(owner_slug, repo_name);
-        debug!(key = %key, path = %local_path.display(), "downloading repo from tigris");
+        debug!(key = %key, path = %target.as_path().display(), "downloading repo from tigris");

        let resp = self
            .s3
@@ -133,36 +570,66 @@ impl TigrisClient {
            .context("reading tigris response body")?
            .into_bytes();

-        // Extract tar.zst to local path
-        tokio::task::spawn_blocking({
-            let local_path = local_path.to_path_buf();
-            move || decompress_repo(&data, &local_path)
+        // The snapshot temp dir is decided HERE, in the async layer, before the
+        // extraction runs. Cleanup ownership moves into the blocking task so a
+        // cancelled async future cannot drop it while extraction is still running.
+        let snapshot_tmp = if publish {
+            None
+        } else {
+            let parent = target.parent().context("snapshot path has no parent")?;
+            std::fs::create_dir_all(parent).context("creating parent dir")?;
+            let file_name = target
+                .file_name()
+                .context("snapshot path has no file name")?
+                .to_string_lossy();
+            Some(parent.join(format!(
+                ".{file_name}.tmp-snapshot.{}",
+                uuid::Uuid::new_v4()
+            )))
+        };
+
+        // Extract tar.zst to a directory.
+        let extracted = tokio::task::spawn_blocking({
+            let target = target.clone();
+            let snapshot_tmp = snapshot_tmp.clone();
+            move || -> Result<DownloadExtract> {
+                let result = (|| -> Result<DownloadExtract> {
+                    if publish {
+                        decompress_repo(&data, &target, swap_authority.as_ref())?;
+                        return Ok(DownloadExtract::Published(()));
+                    }
+                    // Non-mutating snapshot: unpack into the temp dir decided above.
+                    // The live repo path is never touched.
+                    let tmp_dir = snapshot_tmp.expect("snapshot path was decided above");
+                    std::fs::create_dir_all(&tmp_dir).context("creating temp extract dir")?;
+                    let unpack = (|| -> Result<()> {
+                        let decoder = zstd::stream::Decoder::new(&data[..])?;
+                        let mut archive = tar::Archive::new(decoder);
+                        archive.unpack(&tmp_dir).context("unpacking tar.zst")?;
+                        Ok(())
+                    })();
+                    if let Err(e) = unpack {
+                        let _ = std::fs::remove_dir_all(&tmp_dir);
+                        return Err(e);
+                    }
+                    Ok(DownloadExtract::Snapshot(TempSnapshotDir { path: tmp_dir }))
+                })();
+                result
+            }
        })
        .await
        .context("extract task panicked")?
        .context("extracting repo")?;

-        info!(key = %key, path = %local_path.display(), "downloaded repo from tigris");
-        Ok(())
-    }
-
-    /// Delete a repo archive from Tigris.
-    #[allow(dead_code)]
-    pub async fn delete(&self, owner_slug: &str, repo_name: &str) -> Result<()> {
-        let key = Self::repo_key(owner_slug, repo_name);
-        self.s3
-            .delete_object()
-            .bucket(&self.bucket)
-            .key(&key)
-            .send()
-            .await
-            .context(format!("tigris DELETE {key}"))?;
-        Ok(())
+        info!(key = %key, path = %target.as_path().display(), "downloaded repo from tigris");
+        Ok(extracted)
    }
 }

 /// Compress a bare repo directory into a tar.zst byte vector.
-fn compress_repo(repo_path: &Path) -> Result<Vec<u8>> {
+/// `pub(crate)` for sweep tests that seed a fake object store with a
+/// byte-identical archive (rather than duplicating the format).
+pub(crate) fn compress_repo(repo_path: &Path) -> Result<Vec<u8>> {
    let buf = Vec::new();
    let encoder = zstd::stream::Encoder::new(buf, 3)?; // level 3 = fast + decent ratio
    let mut tar = tar::Builder::new(encoder);
@@ -180,7 +647,7 @@ fn compress_repo(repo_path: &Path) -> Result<Vec<u8>> {
 /// `decompress_repo`. Concurrent extractions unpack into isolated temp dirs in
 /// parallel, but the final `remove_dir_all` + `rename` must not interleave for
 /// the same `local_path`, or they race to a nondeterministic overwrite/failure.
-fn publish_lock(local_path: &Path) -> Arc<Mutex<()>> {
+pub(crate) fn publish_lock(local_path: &Path) -> Arc<Mutex<()>> {
    // KNOWN LIMITATION: this map is never evicted — one (PathBuf, Arc<Mutex>)
    // entry accrues per distinct repo path for the process lifetime. Bounded by
    // the number of repos a node hosts, so it's negligible for normal use, but
@@ -193,6 +660,35 @@ fn publish_lock(local_path: &Path) -> Arc<Mutex<()>> {
        .clone()
 }

+/// Owns a snapshot temp dir from extraction through handoff to [`RepoSnapshot`].
+/// Dropped when the `spawn_blocking` join result is abandoned, so cancellation
+/// before the outer future resumes still removes the directory.
+pub(crate) struct TempSnapshotDir {
+    path: PathBuf,
+}
+
+impl TempSnapshotDir {
+    pub(crate) fn into_repo_snapshot(self) -> super::repo_store::RepoSnapshot {
+        let path = self.path.clone();
+        std::mem::forget(self);
+        super::repo_store::RepoSnapshot::from_owned_path(path)
+    }
+}
+
+impl Drop for TempSnapshotDir {
+    fn drop(&mut self) {
+        let _ = std::fs::remove_dir_all(&self.path);
+    }
+}
+
+/// What [`TigrisClient::download_to`] produced.
+pub enum DownloadExtract {
+    /// Published into the validated live repo path.
+    Published(()),
+    /// Unpacked into a throwaway temp dir that cleans up on drop until adopted.
+    Snapshot(TempSnapshotDir),
+}
+
 /// Decompress a tar.zst byte vector into a local directory.
 ///
 /// Extraction is atomic with respect to `local_path`: the archive is unpacked
@@ -200,11 +696,16 @@ fn publish_lock(local_path: &Path) -> Arc<Mutex<()>> {
 /// fully succeeds. A corrupt or truncated archive therefore can never clobber a
 /// good existing copy at `local_path` — on failure we discard the temp dir and
 /// leave `local_path` exactly as it was.
-fn decompress_repo(data: &[u8], local_path: &Path) -> Result<()> {
-    let parent = local_path.parent().context("repo path has no parent")?;
+fn decompress_repo(
+    data: &[u8],
+    local_path: &super::repo_store::ValidatedRepoDiskPath,
+    swap_authority: Option<&Arc<AtomicBool>>,
+) -> Result<()> {
+    let live = local_path.as_path();
+    let parent = live.parent().context("repo path has no parent")?;
    std::fs::create_dir_all(parent).context("creating parent dir")?;

-    let file_name = local_path
+    let file_name = live
        .file_name()
        .context("repo path has no file name")?
        .to_string_lossy();
@@ -229,17 +730,218 @@ fn decompress_repo(data: &[u8], local_path: &Path) -> Result<()> {
        return Err(e);
    }

-    // Swap the freshly-extracted repo into place. rename within the same parent
-    // is effectively atomic, but most platforms refuse to rename onto a
-    // non-empty dir, so remove the old copy first. Serialize this per repo path:
-    // concurrent extractions unpack into isolated temp dirs, but their swaps
-    // must not interleave or they race to a nondeterministic overwrite/failure.
-    let lock = publish_lock(local_path);
-    let _publish = lock.lock().expect("publish lock poisoned");
-    if local_path.exists() {
-        std::fs::remove_dir_all(local_path).context("removing stale repo dir")?;
-    }
-    std::fs::rename(&tmp_dir, local_path).context("swapping extracted repo into place")?;
+    // Swap through the validated-path helper so CodeQL sees the barrier before the
+    // remove/rename sink (`rust/path-injection`).
+    super::repo_store::swap_extracted_into_validated_repo(local_path, &tmp_dir, swap_authority)?;

    Ok(())
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use aws_sdk_s3::primitives::ByteStream;
+    use futures::FutureExt;
+
+    /// The envs the probe needs, all of them, or it does not run.
+    ///
+    /// `AWS_ENDPOINT_URL_S3` is included on purpose: without it the SDK resolves
+    /// to real AWS S3, and a probe that passed there would say nothing about
+    /// Tigris.
+    fn probe_env() -> Option<String> {
+        if std::env::var("GITLAWB_TIGRIS_PROBE").ok().as_deref() != Some("1") {
+            return None;
+        }
+        for name in [
+            "AWS_ACCESS_KEY_ID",
+            "AWS_SECRET_ACCESS_KEY",
+            "AWS_ENDPOINT_URL_S3",
+        ] {
+            if std::env::var(name).is_err() {
+                eprintln!("tigris conditional-write probe: {name} is unset, skipping");
+                return None;
+            }
+        }
+        match std::env::var("GITLAWB_TIGRIS_BUCKET") {
+            Ok(b) if !b.is_empty() => Some(b),
+            _ => {
+                eprintln!(
+                    "tigris conditional-write probe: GITLAWB_TIGRIS_BUCKET is unset, skipping"
+                );
+                None
+            }
+        }
+    }
+
+    /// One conditional PUT, reported as the status that REFUSED it, or `None`
+    /// when the store accepted the write.
+    ///
+    /// Accepted is the interesting answer here, not an error: it means the
+    /// endpoint ignored the header we fenced on.
+    async fn conditional_put(
+        s3: &S3Client,
+        bucket: &str,
+        key: &str,
+        body: &'static [u8],
+        if_match: Option<&str>,
+        if_none_match: Option<&str>,
+    ) -> Result<Option<u16>, String> {
+        let mut req = s3
+            .put_object()
+            .bucket(bucket)
+            .key(key)
+            .body(ByteStream::from_static(body));
+        if let Some(v) = if_match {
+            req = req.if_match(v);
+        }
+        if let Some(v) = if_none_match {
+            req = req.if_none_match(v);
+        }
+        match req.send().await {
+            Ok(_) => Ok(None),
+            Err(e) => match e.raw_response() {
+                // Same raw-response rule as `upload`: a refused conditional PUT
+                // whose body the SDK cannot parse surfaces as `ResponseError`, and
+                // the status has to come off the raw response for both variants.
+                // Without it, an unparsable 409/412 would report "no HTTP
+                // response" here and skip the supersede retry.
+                Some(raw) => Ok(Some(raw.status().as_u16())),
+                None => Err(format!("conditional PUT {key}: no HTTP response: {e}")),
+            },
+        }
+    }
+
+    /// The probe body, written to RETURN its failures rather than panic on
+    /// them, so the caller's cleanup is reached on every arm.
+    async fn conditional_write_probe(s3: &S3Client, bucket: &str, key: &str) -> Result<(), String> {
+        // 1. A plain PUT under a throwaway key.
+        let seeded = s3
+            .put_object()
+            .bucket(bucket)
+            .key(key)
+            .body(ByteStream::from_static(b"probe-one"))
+            .send()
+            .await
+            .map_err(|e| format!("seeding PUT {key}: {e}"))?;
+
+        // 2. Its ETag, which is the generation the next arm fences against.
+        let etag = seeded
+            .e_tag()
+            .ok_or_else(|| format!("seeding PUT {key} returned no ETag"))?
+            .to_string();
+
+        // 3. A deliberately wrong If-Match. A store honoring it answers 412.
+        let wrong = format!("\"{}\"", "0".repeat(32));
+        if etag.trim_matches('"') == wrong.trim_matches('"') {
+            return Err(format!(
+                "the seeded ETag {etag} collides with the deliberately wrong one, \
+                 so this arm would prove nothing"
+            ));
+        }
+        match conditional_put(s3, bucket, key, b"probe-two", Some(&wrong), None).await? {
+            Some(412) => {}
+            Some(status) => {
+                return Err(format!(
+                    "a stale If-Match must be refused with 412, the endpoint answered {status}"
+                ))
+            }
+            None => {
+                return Err(
+                    "a stale If-Match was ACCEPTED: this endpoint does not honor If-Match, so \
+                     the release fence cannot hold here"
+                        .to_string(),
+                )
+            }
+        }
+
+        // 4. If-None-Match `*` over the object that now exists. This arm matters
+        // MORE than the one above. An ignored If-Match eventually surfaces as
+        // odd behavior, because a stale writer overwrites and someone notices
+        // the lost tree. An ignored If-None-Match just returns 200, so a publish
+        // that should have been fenced lands with no error anywhere: the silent
+        // no-op the bucket-type caveat on this test describes.
+        match conditional_put(s3, bucket, key, b"probe-three", None, Some("*")).await? {
+            // Either status is a pass, and the asymmetry with the If-Match arm
+            // above mirrors `upload`'s classifier exactly: 412 is always a lost
+            // precondition, and 409 is one too when we asked for create-only.
+            // AWS documents 409 for a create-only conflict racing a delete, so a
+            // store answering it is enforcing the precondition and we already
+            // handle it. Pinning 412 alone here would fail the probe against a
+            // backend that is behaving correctly, which sends whoever runs it
+            // chasing a fault that is not there.
+            Some(412) | Some(409) => {}
+            Some(status) => {
+                return Err(format!(
+                    "create-only over an existing object must be refused with 412 or 409, \
+                     the endpoint answered {status}"
+                ))
+            }
+            None => {
+                return Err(
+                    "If-None-Match * was ACCEPTED over an existing object: this endpoint does \
+                     not honor create-only, so a fenced publish lands silently"
+                        .to_string(),
+                )
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Probe the REAL Tigris endpoint for the conditional-write semantics the
+    /// release fence depends on.
+    ///
+    /// UNTIL THIS IS RUN AGAINST REAL CREDENTIALS, the fence is verified against
+    /// vendor documentation and an in-process mock, not against the backend it
+    /// runs on. The mock implements the semantics we believe Tigris has; it
+    /// cannot tell us whether Tigris actually has them.
+    ///
+    /// The bucket matters, not just the endpoint. Tigris documents conditional
+    /// operations as supported on Single-region and Multi-region buckets only.
+    /// Global and Dual-region buckets are eventually consistent, and a
+    /// conditional PUT evaluated against a stale replica would make the fence a
+    /// silent no-op rather than an error. So point `GITLAWB_TIGRIS_BUCKET` at a
+    /// throwaway bucket of the SAME type production uses.
+    ///
+    /// Ignored by default and additionally gated on `GITLAWB_TIGRIS_PROBE=1`,
+    /// because it writes to a real bucket and costs real requests. Run with:
+    /// `GITLAWB_TIGRIS_PROBE=1 cargo test -p gitlawb-node --bin gitlawb-node
+    /// tigris_honors_conditional_writes -- --ignored --nocapture`
+    #[tokio::test]
+    #[ignore = "writes to a real Tigris bucket; needs GITLAWB_TIGRIS_PROBE=1 plus credentials"]
+    async fn tigris_honors_conditional_writes() {
+        let Some(bucket) = probe_env() else {
+            eprintln!(
+                "tigris conditional-write probe: skipped. Set GITLAWB_TIGRIS_PROBE=1, \
+                 GITLAWB_TIGRIS_BUCKET, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY and \
+                 AWS_ENDPOINT_URL_S3 to run it."
+            );
+            return;
+        };
+
+        let config = aws_config::load_defaults(aws_config::BehaviorVersion::latest()).await;
+        let s3 = S3Client::new(&config);
+        // A fresh key per run, so a probe that somehow orphaned an object on an
+        // earlier run cannot change what this one observes.
+        let key = format!("probe/conditional-write-{}.bin", uuid::Uuid::new_v4());
+
+        // CLEANUP MUST RUN ON EVERY ARM, and a failing assertion is precisely
+        // the case this probe exists to catch, so the delete cannot sit after
+        // the checks. The body returns its failures rather than panicking, and
+        // `catch_unwind` covers the panic an SDK call could still raise; either
+        // way the delete below is reached before the verdict is re-raised.
+        let outcome = std::panic::AssertUnwindSafe(conditional_write_probe(&s3, &bucket, &key))
+            .catch_unwind()
+            .await;
+
+        if let Err(e) = s3.delete_object().bucket(&bucket).key(&key).send().await {
+            eprintln!("tigris conditional-write probe: cleanup of {key} failed: {e}");
+        }
+
+        match outcome {
+            Ok(Ok(())) => {}
+            Ok(Err(msg)) => panic!("tigris conditional-write probe: {msg}"),
+            Err(payload) => std::panic::resume_unwind(payload),
+        }
+    }
+}
@@ -3,7 +3,7 @@ pub mod query;
 pub mod subscription;
 pub mod types;

-use async_graphql::Schema;
+use async_graphql::{Schema, SchemaBuilder};
 use std::sync::Arc;

 use crate::db::Db;
@@ -14,6 +14,22 @@ use subscription::SubscriptionRoot;

 pub type GitlawbSchema = Schema<QueryRoot, MutationRoot, SubscriptionRoot>;

+// Keep ordinary schema discovery and application queries usable while
+// bounding validation work and the number of resolver selections per request.
+const GRAPHQL_MAX_COMPLEXITY: usize = 400;
+// The current public schema is shallow; this leaves headroom for composed
+// clients and the canonical getIntrospectionQuery without allowing
+// recursively nested documents to grow unchecked.
+const GRAPHQL_MAX_DEPTH: usize = 14;
+
+fn apply_query_limits<Query, Mutation, Subscription>(
+    builder: SchemaBuilder<Query, Mutation, Subscription>,
+) -> SchemaBuilder<Query, Mutation, Subscription> {
+    builder
+        .limit_complexity(GRAPHQL_MAX_COMPLEXITY)
+        .limit_depth(GRAPHQL_MAX_DEPTH)
+}
+
 /// Client-facing message for GraphQL resolver failures that wrap a real
 /// `sqlx::Error`. The real error is logged server-side; never put sqlx/Postgres
 /// detail in the GraphQL `errors` array (#250).
@@ -60,6 +76,7 @@ pub(crate) fn graphql_app_err(e: crate::error::AppError) -> async_graphql::Error
        // Curated client-safe variants — `Display` is intentional API text.
        safe @ (crate::error::AppError::RepoNotFound(_)
        | crate::error::AppError::RepoExists(_)
+        | crate::error::AppError::Conflict(_)
        | crate::error::AppError::NotFound(_)
        | crate::error::AppError::Unauthorized(_)
        | crate::error::AppError::Forbidden(_)
@@ -76,21 +93,91 @@ pub(crate) fn graphql_app_err(e: crate::error::AppError) -> async_graphql::Error
    }
 }

+/// Classify a failed `claim_task` for GraphQL exactly as REST's claim handler
+/// classifies it: a lost race is a client-safe conflict, a real sqlx fault
+/// stays opaque. Lives here rather than at the three call sites so the
+/// classification cannot drift between transports (#327 review), and so
+/// `every_graphql_map_err_uses_opaque_helpers` can keep whitelisting by name.
+pub(crate) fn graphql_claim_conflict(e: anyhow::Error) -> async_graphql::Error {
+    graphql_app_err(crate::api::tasks::task_write_conflict(
+        e,
+        "task not claimable: not found or already claimed",
+    ))
+}
+
+/// The `finish_task` half of [`graphql_claim_conflict`], covering both
+/// `completeTask` and `failTask`.
+pub(crate) fn graphql_finish_conflict(e: anyhow::Error) -> async_graphql::Error {
+    graphql_app_err(crate::api::tasks::task_write_conflict(
+        e,
+        "task not found or not in claimed state",
+    ))
+}
+
+pub struct TaskReadBrakeExtension;
+
+impl async_graphql::extensions::ExtensionFactory for TaskReadBrakeExtension {
+    fn create(&self) -> Arc<dyn async_graphql::extensions::Extension> {
+        Arc::new(TaskReadBrakeExtensionImpl)
+    }
+}
+
+use async_graphql::async_trait::async_trait;
+
+struct TaskReadBrakeExtensionImpl;
+
+#[async_trait]
+impl async_graphql::extensions::Extension for TaskReadBrakeExtensionImpl {
+    async fn prepare_request(
+        &self,
+        ctx: &async_graphql::extensions::ExtensionContext<'_>,
+        mut request: async_graphql::Request,
+        next: async_graphql::extensions::NextPrepareRequest<'_>,
+    ) -> async_graphql::ServerResult<async_graphql::Request> {
+        if let Some(session_brake) = ctx
+            .session_data
+            .get(&std::any::TypeId::of::<crate::rate_limit::TaskReadBrake>())
+            .and_then(|d| d.downcast_ref::<crate::rate_limit::TaskReadBrake>())
+        {
+            request = request.data(crate::rate_limit::TaskReadBrake {
+                limiter: session_brake.limiter.clone(),
+                key: session_brake.key.clone(),
+                request_count: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
+            });
+        }
+        if let Some(did) = ctx
+            .session_data
+            .get(&std::any::TypeId::of::<crate::auth::AuthenticatedDid>())
+            .and_then(|d| d.downcast_ref::<crate::auth::AuthenticatedDid>())
+        {
+            request = request.data(did.clone());
+        }
+        next.run(ctx, request).await
+    }
+}
+
 pub fn build_schema(
    db: Arc<Db>,
    ref_update_tx: tokio::sync::broadcast::Sender<RefUpdateBroadcast>,
    task_event_tx: tokio::sync::broadcast::Sender<TaskEventBroadcast>,
+    task_cursor_key: crate::api::task_cursor::TaskCursorKey,
 ) -> GitlawbSchema {
-    Schema::build(QueryRoot, MutationRoot, SubscriptionRoot)
+    apply_query_limits(Schema::build(QueryRoot, MutationRoot, SubscriptionRoot))
        .data(db)
        .data(ref_update_tx)
        .data(task_event_tx)
+        .data(task_cursor_key)
+        .extension(TaskReadBrakeExtension)
        .finish()
 }

 #[cfg(test)]
 mod tests {
    use super::*;
+    use crate::api::events::MAX_VISIBLE_REF_UPDATES;
+    use crate::db::{MAX_VISIBLE_REPO_PAGE_SIZE, MAX_VISIBLE_TASK_PAGE_SIZE};
+    use async_graphql::{EmptyMutation, EmptySubscription, Object, Value};
+    use std::sync::atomic::{AtomicUsize, Ordering};

    #[test]
    fn graphql_db_err_opaques_sqlx_chain() {
@@ -160,9 +247,683 @@ mod tests {
        assert!(!err.message.contains("failed to open"));
    }

+    fn production_test_schema() -> GitlawbSchema {
+        let pool = sqlx::postgres::PgPoolOptions::new()
+            .connect_lazy("postgres://localhost/unused")
+            .unwrap();
+        production_test_schema_with_db(Arc::new(Db::for_testing(pool)))
+    }
+
+    fn production_test_schema_with_db(db: Arc<Db>) -> GitlawbSchema {
+        build_schema(
+            db,
+            tokio::sync::broadcast::channel(1).0,
+            tokio::sync::broadcast::channel(1).0,
+            crate::api::task_cursor::TaskCursorKey::derive(&[42u8; 32]),
+        )
+    }
+
+    #[tokio::test]
+    async fn ref_update_subscription_alias_budget() {
+        assert_subscription_alias_budget("refUpdates { repo }").await;
+    }
+
+    #[tokio::test]
+    async fn task_event_subscription_alias_budget() {
+        assert_subscription_alias_budget("taskEvents { taskId }").await;
+    }
+
+    async fn assert_subscription_alias_budget(field: &str) {
+        use futures::StreamExt;
+        use std::time::Duration;
+        let pool = sqlx::postgres::PgPoolOptions::new()
+            .connect_lazy("postgres://localhost/unused")
+            .unwrap();
+        let (ref_tx, _) = tokio::sync::broadcast::channel(16);
+        let (task_tx, _) = tokio::sync::broadcast::channel(16);
+        let schema = build_schema(
+            Arc::new(Db::for_testing(pool)),
+            ref_tx.clone(),
+            task_tx.clone(),
+            crate::api::task_cursor::TaskCursorKey::derive(&[42u8; 32]),
+        );
+        for count in [7, 8] {
+            let fields = (0..count)
+                .map(|n| format!("r{n}: {field}"))
+                .collect::<Vec<_>>()
+                .join(" ");
+            let query = format!("subscription {{ {fields} }}");
+            let mut responses = Box::pin(schema.execute_stream(query));
+            if count == 8 {
+                let response = tokio::time::timeout(Duration::from_secs(1), responses.next())
+                    .await
+                    .expect("over-budget subscriptions must reject before waiting for events")
+                    .expect("the validation error must be returned");
+                assert_eq!(response.errors.len(), 1);
+                assert_eq!(response.errors[0].message, "Query is too complex.");
+                assert_eq!(ref_tx.receiver_count() + task_tx.receiver_count(), 0);
+                continue;
+            }
+            // Poll the accepted operation far enough to register its receivers.
+            assert!(
+                tokio::time::timeout(Duration::from_millis(30), responses.next())
+                    .await
+                    .is_err()
+            );
+            assert_eq!(ref_tx.receiver_count() + task_tx.receiver_count(), 7);
+            let _ = ref_tx.send(RefUpdateBroadcast {
+                repo: "owner/repo".into(),
+                ref_name: "refs/heads/main".into(),
+                old_sha: "0".repeat(40),
+                new_sha: "1".repeat(40),
+                pusher_did: "did:key:test".into(),
+                node_did: "did:key:node".into(),
+                timestamp: "2026-01-01T00:00:00Z".into(),
+                owner_did: "did:key:test".into(),
+            });
+            let _ = task_tx.send(TaskEventBroadcast {
+                task_id: "task".into(),
+                old_status: "pending".into(),
+                new_status: "claimed".into(),
+                by_did: "did:key:test".into(),
+                at: "2026-01-01T00:00:00Z".into(),
+            });
+            let response = tokio::time::timeout(Duration::from_secs(1), responses.next())
+                .await
+                .unwrap()
+                .unwrap();
+            assert!(response.errors.is_empty(), "{:?}", response.errors);
+            assert_ne!(response.data, Value::Null);
+            drop(responses);
+            assert_eq!(ref_tx.receiver_count() + task_tx.receiver_count(), 0);
+        }
+    }
+
+    #[sqlx::test]
+    async fn production_mutation_budget_accepts_seven_and_rejects_eight(pool: sqlx::PgPool) {
+        let db = Arc::new(Db::for_testing(pool));
+        db.run_migrations().await.unwrap();
+        let caller = "did:key:reviewer";
+        let now = chrono::Utc::now().to_rfc3339();
+        for index in 0..8 {
+            db.create_task(&crate::db::AgentTask {
+                id: format!("task-{index}"),
+                repo_id: None,
+                kind: "test".into(),
+                status: "pending".into(),
+                delegator_did: caller.into(),
+                assignee_did: None,
+                capability: "test".into(),
+                ucan_token: None,
+                payload: None,
+                result: None,
+                created_at: now.clone(),
+                updated_at: now.clone(),
+                deadline: None,
+            })
+            .await
+            .unwrap();
+        }
+        let schema = production_test_schema_with_db(db.clone());
+        for count in [8, 7] {
+            let fields = (0..count)
+                .map(|n| {
+                    format!("r{n}: claimTask(id: \"task-{n}\", assigneeDid: \"{caller}\") {{ id }}")
+                })
+                .collect::<Vec<_>>()
+                .join(" ");
+            let response = schema
+                .execute(
+                    async_graphql::Request::new(format!("mutation {{ {fields} }}"))
+                        .data(crate::auth::AuthenticatedDid(caller.into())),
+                )
+                .await;
+            if count == 8 {
+                assert_eq!(response.errors.len(), 1);
+                assert_eq!(response.errors[0].message, "Query is too complex.");
+                assert_eq!(
+                    db.list_tasks_keyset(Some("pending"), None, 8, None)
+                        .await
+                        .unwrap()
+                        .len(),
+                    8
+                );
+                continue;
+            }
+            assert!(response.errors.is_empty(), "{:?}", response.errors);
+            assert_eq!(
+                response
+                    .data
+                    .into_json()
+                    .unwrap()
+                    .as_object()
+                    .unwrap()
+                    .len(),
+                7
+            );
+            assert_eq!(
+                db.list_tasks_keyset(Some("claimed"), Some(caller), 8, None)
+                    .await
+                    .unwrap()
+                    .len(),
+                7
+            );
+            assert_eq!(
+                db.get_task("task-7").await.unwrap().unwrap().status,
+                "pending"
+            );
+        }
+    }
+
+    #[sqlx::test]
+    async fn production_list_cost_scales_with_the_same_alias_count(pool: sqlx::PgPool) {
+        let db = Arc::new(Db::for_testing(pool));
+        db.run_migrations().await.unwrap();
+        let schema = production_test_schema_with_db(db);
+        // Bounds are per-root named constants so the cases move with the
+        // documented maximums instead of going stale as literals.
+        for (field, selection, max) in [
+            ("refUpdates", "repo", MAX_VISIBLE_REF_UPDATES),
+            ("tasks", "items { id }", MAX_VISIBLE_TASK_PAGE_SIZE),
+            (
+                "reposPage",
+                "nodes { name }",
+                MAX_VISIBLE_REPO_PAGE_SIZE as i64,
+            ),
+        ] {
+            // Two aliases at the maximum always exceed the budget.
+            let fields = (0..2)
+                .map(|n| format!("r{n}: {field}(limit: {max}) {{ {selection} }}"))
+                .collect::<Vec<_>>()
+                .join(" ");
+            let response = schema.execute(format!("{{ {fields} }}")).await;
+            assert_eq!(response.errors.len(), 1);
+            assert_eq!(response.errors[0].message, "Query is too complex.");
+
+            // At limit 1 two aliases fit for the row-priced roots; reposPage
+            // prices the full corpus scan, so a single alias is the most that
+            // fits under the budget.
+            let aliases = if field == "reposPage" { 1 } else { 2 };
+            let fields = (0..aliases)
+                .map(|n| format!("r{n}: {field}(limit: 1) {{ {selection} }}"))
+                .collect::<Vec<_>>()
+                .join(" ");
+            let response = schema.execute(format!("{{ {fields} }}")).await;
+            assert!(response.errors.is_empty(), "{field}: {:?}", response.errors);
+            assert_eq!(
+                response
+                    .data
+                    .into_json()
+                    .unwrap()
+                    .as_object()
+                    .unwrap()
+                    .len(),
+                aliases
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn production_limits_reject_mutation_aliases_and_large_lists() {
+        let schema = production_test_schema();
+        // Limits are derived from the MAX_* constants so the boundaries move
+        // with the documented maximums instead of going stale as literals.
+        for (prefix, count, field) in [
+            ("mutation", 8, "claimTask(id: \"missing\", assigneeDid: \"did:key:test\") { id }".to_owned()),
+            ("mutation", 8, "createTask(delegatorDid: \"did:key:test\", input: { kind: \"test\", capability: \"test\" }) { id }".to_owned()),
+            ("mutation", 8, "completeTask(id: \"missing\", byDid: \"did:key:test\", input: {}) { id }".to_owned()),
+            ("mutation", 8, "failTask(id: \"missing\", byDid: \"did:key:test\", input: {}) { id }".to_owned()),
+            ("query", 8, "task(id: \"missing\") { id }".to_owned()),
+            ("query", 2, format!("refUpdates(limit: {MAX_VISIBLE_REF_UPDATES}) {{ repo }}")),
+            ("query", 2, format!("tasks(limit: {MAX_VISIBLE_TASK_PAGE_SIZE}) {{ id }}")),
+            ("query", 2, format!("reposPage(limit: {MAX_VISIBLE_REPO_PAGE_SIZE}) {{ nodes {{ name }} }}")),
+            ("query", 8, "reposPage(limit: 1) { nodes { name } }".to_owned()),
+        ] {
+            let fields = (0..count)
+                .map(|n| format!("r{n}: {field}"))
+                .collect::<Vec<_>>()
+                .join(" ");
+            let response = schema.execute(format!("{prefix} {{ {fields} }}")).await;
+            assert_eq!(response.errors.len(), 1, "{:?}", response.errors);
+            assert_eq!(response.errors[0].message, "Query is too complex.");
+        }
+    }
+
+    #[sqlx::test]
+    async fn production_repos_page_serves_the_documented_maximum(pool: sqlx::PgPool) {
+        let db = Arc::new(Db::for_testing(pool));
+        db.run_migrations().await.unwrap();
+        let schema = production_test_schema_with_db(db);
+        for selection in [
+            "nodes { name }",
+            "nodes { name ownerDid } hasNextPage endCursor",
+        ] {
+            let response = schema
+                .execute(format!(
+                    "{{ reposPage(limit: {MAX_VISIBLE_REPO_PAGE_SIZE}) {{ {selection} }} }}"
+                ))
+                .await;
+            assert!(response.errors.is_empty(), "{:?}", response.errors);
+            assert_eq!(
+                response.data.into_json().unwrap()["reposPage"]["nodes"],
+                serde_json::json!([])
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn seven_root_aliases_are_accepted() {
+        struct Root(Arc<AtomicUsize>);
+        #[Object]
+        impl Root {
+            #[graphql(complexity = "50 + child_complexity")]
+            async fn repos(&self) -> Vec<Nested> {
+                self.0.fetch_add(1, Ordering::Relaxed);
+                vec![Nested]
+            }
+        }
+        let calls = Arc::new(AtomicUsize::new(0));
+        let schema = apply_query_limits(Schema::build(
+            Root(calls.clone()),
+            EmptyMutation,
+            EmptySubscription,
+        ))
+        .finish();
+        let fields = (0..7)
+            .map(|n| format!("r{n}: repos {{ value }}"))
+            .collect::<Vec<_>>()
+            .join(" ");
+        let response = schema.execute(format!("{{ {fields} }}")).await;
+        assert!(response.errors.is_empty(), "{:?}", response.errors);
+        assert_eq!(calls.load(Ordering::Relaxed), 7);
+    }
+
+    #[tokio::test]
+    async fn expensive_root_aliases_are_rejected_before_database_access() {
+        // Use the production builder with a lazy pool: rejection must precede DB access.
+        let schema = production_test_schema();
+        let fields = (0..8)
+            .map(|n| format!("r{n}: repos {{ name }}"))
+            .collect::<Vec<_>>()
+            .join(" ");
+
+        let response = schema.execute(format!("{{ {fields} }}")).await;
+
+        assert_eq!(response.data, Value::Null);
+        assert_eq!(response.errors.len(), 1);
+        assert_eq!(response.errors[0].message, "Query is too complex.");
+    }
+
+    #[tokio::test]
+    async fn ordinary_schema_introspection_remains_available() {
+        let schema = production_test_schema();
+        let canonical_introspection = r#"
+            query IntrospectionQuery {
+              __schema {
+                queryType { name }
+                mutationType { name }
+                subscriptionType { name }
+                types {
+                  ...FullType
+                }
+                directives {
+                  name
+                  description
+                  locations
+                  args {
+                    ...InputValue
+                  }
+                }
+              }
+            }
+
+            fragment FullType on __Type {
+              kind
+              name
+              description
+              fields(includeDeprecated: true) {
+                name
+                description
+                args {
+                  ...InputValue
+                }
+                type {
+                  ...TypeRef
+                }
+                isDeprecated
+                deprecationReason
+              }
+              inputFields {
+                ...InputValue
+              }
+              interfaces {
+                ...TypeRef
+              }
+              enumValues(includeDeprecated: true) {
+                name
+                description
+                isDeprecated
+                deprecationReason
+              }
+              possibleTypes {
+                ...TypeRef
+              }
+            }
+
+            fragment InputValue on __InputValue {
+              name
+              description
+              type { ...TypeRef }
+              defaultValue
+            }
+
+            fragment TypeRef on __Type {
+              kind
+              name
+              ofType {
+                kind
+                name
+                ofType {
+                  kind
+                  name
+                  ofType {
+                    kind
+                    name
+                    ofType {
+                      kind
+                      name
+                      ofType {
+                        kind
+                        name
+                        ofType {
+                          kind
+                          name
+                          ofType {
+                            kind
+                            name
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+        "#;
+        let response = schema.execute(canonical_introspection).await;
+        assert!(
+            response.errors.is_empty(),
+            "canonical introspection query failed against production schema: {:?}",
+            response.errors
+        );
+        assert_ne!(response.data, Value::Null);
+    }
+
+    #[tokio::test]
+    async fn ref_updates_and_tasks_single_alias_max_limit_complexity_contract() {
+        let pool = sqlx::postgres::PgPoolOptions::new()
+            .connect_lazy("postgres://localhost/unused")
+            .unwrap();
+        pool.close().await;
+        let (ref_tx, _) = tokio::sync::broadcast::channel(16);
+        let (task_tx, _) = tokio::sync::broadcast::channel(16);
+        let schema = build_schema(
+            Arc::new(Db::for_testing(pool)),
+            ref_tx,
+            task_tx,
+            crate::api::task_cursor::TaskCursorKey::derive(&[42u8; 32]),
+        );
+
+        // refUpdates(limit: MAX_VISIBLE_REF_UPDATES):
+        // Single-field selection has child_complexity 1 -> cost 100 + 200 + 1 = 301 <= 400.
+        // It passes validation and reaches the resolver (which yields db error on lazy pool).
+        let single_ref = schema
+            .execute(format!(
+                "{{ refUpdates(limit: {MAX_VISIBLE_REF_UPDATES}) {{ repo }} }}"
+            ))
+            .await;
+        assert!(
+            !single_ref
+                .errors
+                .iter()
+                .any(|e| e.message == "Query is too complex."),
+            "single-field refUpdates at max limit must pass complexity validation: {:?}",
+            single_ref.errors
+        );
+        assert!(
+            single_ref
+                .errors
+                .iter()
+                .any(|e| e.message == GRAPHQL_DB_ERROR_MESSAGE),
+            "single-field refUpdates at max limit must reach the resolver: {:?}",
+            single_ref.errors
+        );
+
+        // Two-field selection has child_complexity 2 -> cost 100 + 200 + 2 = 302 <= 400.
+        // The documented max-200 request stays reachable with multiple fields
+        // (the old multiplicative price rejected it at 450); it passes
+        // validation and reaches the resolver like the single-field case.
+        let multi_ref = schema
+            .execute(format!(
+                "{{ refUpdates(limit: {MAX_VISIBLE_REF_UPDATES}) {{ repo refName }} }}"
+            ))
+            .await;
+        assert!(
+            !multi_ref
+                .errors
+                .iter()
+                .any(|e| e.message == "Query is too complex."),
+            "two-field refUpdates at max limit must pass complexity validation: {:?}",
+            multi_ref.errors
+        );
+        assert!(
+            multi_ref
+                .errors
+                .iter()
+                .any(|e| e.message == GRAPHQL_DB_ERROR_MESSAGE),
+            "two-field refUpdates at max limit must reach the resolver: {:?}",
+            multi_ref.errors
+        );
+
+        // tasks(limit: MAX_VISIBLE_TASK_PAGE_SIZE):
+        // Selection items { id } has child_complexity 2, so the cost is
+        // 50 + MAX_VISIBLE_TASK_PAGE_SIZE + 2 <= 400.
+        // It passes validation and reaches the resolver.
+        let single_task = schema
+            .execute(format!(
+                "{{ tasks(limit: {MAX_VISIBLE_TASK_PAGE_SIZE}) {{ items {{ id }} }} }}"
+            ))
+            .await;
+        assert!(
+            !single_task
+                .errors
+                .iter()
+                .any(|e| e.message == "Query is too complex."),
+            "single-field tasks at the maximum page size must pass complexity validation: {:?}",
+            single_task.errors
+        );
+        assert!(
+            single_task
+                .errors
+                .iter()
+                .any(|e| e.message == GRAPHQL_DB_ERROR_MESSAGE),
+            "single-field tasks at the maximum page size must reach the resolver: {:?}",
+            single_task.errors
+        );
+
+        // Selection items { id status } has child_complexity 3, so the cost is
+        // 50 + MAX_VISIBLE_TASK_PAGE_SIZE + 3 <= 400.
+        // A request at the documented maximum stays reachable with multiple
+        // fields (the old multiplicative price rejected it at 450); it passes
+        // validation and reaches the resolver like the single-field case.
+        let multi_task = schema
+            .execute(format!(
+                "{{ tasks(limit: {MAX_VISIBLE_TASK_PAGE_SIZE}) {{ items {{ id status }} }} }}"
+            ))
+            .await;
+        assert!(
+            !multi_task
+                .errors
+                .iter()
+                .any(|e| e.message == "Query is too complex."),
+            "two-field tasks at the maximum page size must pass complexity validation: {:?}",
+            multi_task.errors
+        );
+        assert!(
+            multi_task
+                .errors
+                .iter()
+                .any(|e| e.message == GRAPHQL_DB_ERROR_MESSAGE),
+            "two-field tasks at the maximum page size must reach the resolver: {:?}",
+            multi_task.errors
+        );
+    }
+
+    #[tokio::test]
+    async fn two_repos_aliases_exceed_the_budget() {
+        // `repos` is priced as the full page it scans
+        // (50 + MAX_VISIBLE_REPO_PAGE_SIZE + child), so two `repos { name }`
+        // aliases cost 2 * 251 = 502 > 400 and the document is rejected before
+        // any resolver runs. If the price reverts to `50 + child_complexity`
+        // the aliases cost 2 * 51 = 102, validation passes, and the failure
+        // mode changes (resolver DB errors on the lazy pool): this test goes
+        // red on revert.
+        let schema = production_test_schema();
+        let response = schema
+            .execute("{ r0: repos { name } r1: repos { name } }")
+            .await;
+        assert_eq!(response.errors.len(), 1, "{:?}", response.errors);
+        assert_eq!(response.errors[0].message, "Query is too complex.");
+    }
+
+    #[tokio::test]
+    async fn seven_repos_page_aliases_exceed_the_budget() {
+        // `reposPage` is priced as the full corpus scan
+        // (50 + MAX_VISIBLE_REPO_PAGE_SIZE + child_complexity), so seven
+        // `reposPage(limit: 1) { nodes { name } }` aliases cost
+        // 7 * (50 + 200 + 2) = 1764 > 400 and the document is rejected
+        // before any resolver runs. Under the previous row-priced formula
+        // the same aliases cost 7 * (50 + 1 + 2) = 371, validation passes,
+        // and the failure mode changes (resolver DB errors on the lazy
+        // pool): this test goes red on revert. Eight aliases are rejected
+        // under both formulas, so seven is the load-bearing boundary.
+        let schema = production_test_schema();
+        let fields = (0..7)
+            .map(|n| format!("r{n}: reposPage(limit: 1) {{ nodes {{ name }} }}"))
+            .collect::<Vec<_>>()
+            .join(" ");
+        let response = schema.execute(format!("{{ {fields} }}")).await;
+        assert_eq!(response.errors.len(), 1, "{:?}", response.errors);
+        assert_eq!(response.errors[0].message, "Query is too complex.");
+    }
+
+    #[tokio::test]
+    async fn ref_updates_fixed_base_is_charged() {
+        // The 100 base in the refUpdates complexity formula charges for the
+        // collector's fixed work (loading the deduped repo set, the
+        // quarantined set, and the visibility rules, then walking up to 16
+        // keyset pages) even at limit: 1. Four aliases at limit: 1 cost
+        // 4 * (100 + 1 + 1) = 408 > 400 and must be rejected. If the base
+        // disappears, the cost drops to 4 * (1 + 1) = 8 and this test fails.
+        let schema = production_test_schema();
+        let fields = (0..4)
+            .map(|n| format!("r{n}: refUpdates(limit: 1) {{ repo }}"))
+            .collect::<Vec<_>>()
+            .join(" ");
+        let response = schema.execute(format!("{{ {fields} }}")).await;
+        assert_eq!(response.errors.len(), 1, "{:?}", response.errors);
+        assert_eq!(response.errors[0].message, "Query is too complex.");
+    }
+
+    #[derive(Clone, Copy)]
+    struct Nested;
+
+    #[Object]
+    impl Nested {
+        async fn child(&self) -> Nested {
+            Nested
+        }
+
+        async fn value(&self) -> i32 {
+            1
+        }
+    }
+
+    struct CountingQuery(Arc<AtomicUsize>);
+
+    #[Object]
+    impl CountingQuery {
+        async fn nested(&self) -> Nested {
+            self.0.fetch_add(1, Ordering::Relaxed);
+            Nested
+        }
+    }
+
+    #[tokio::test]
+    async fn query_depth_limit_accepts_fourteen_and_rejects_fifteen() {
+        let calls = Arc::new(AtomicUsize::new(0));
+        let schema = apply_query_limits(Schema::build(
+            CountingQuery(Arc::clone(&calls)),
+            EmptyMutation,
+            EmptySubscription,
+        ))
+        .finish();
+        let query_at_depth = |depth: usize| {
+            let selection = (2..depth).fold("value".to_string(), |selection, _| {
+                format!("child {{ {selection} }}")
+            });
+            format!("{{ nested {{ {selection} }} }}")
+        };
+
+        let accepted = schema.execute(query_at_depth(GRAPHQL_MAX_DEPTH)).await;
+
+        assert!(
+            accepted.errors.is_empty(),
+            "depth {GRAPHQL_MAX_DEPTH} should be accepted: {:?}",
+            accepted.errors
+        );
+        assert_eq!(calls.swap(0, Ordering::Relaxed), 1);
+
+        let rejected = schema.execute(query_at_depth(GRAPHQL_MAX_DEPTH + 1)).await;
+
+        assert_eq!(rejected.data, Value::Null);
+        assert_eq!(rejected.errors.len(), 1);
+        assert_eq!(rejected.errors[0].message, "Query is nested too deep.");
+        assert_eq!(calls.load(Ordering::Relaxed), 0);
+    }
+
+    #[tokio::test]
+    async fn production_schema_enforces_depth_limit() {
+        let schema = production_test_schema();
+        // Depth 14 is accepted through the production build_schema path.
+        let of_types_14 = (0..9).fold("name".to_string(), |acc, _| format!("ofType {{ {acc} }}"));
+        let query_14 =
+            format!("{{ __schema {{ types {{ fields {{ type {{ {of_types_14} }} }} }} }} }}");
+        let accepted = schema.execute(&query_14).await;
+        assert!(accepted.errors.is_empty(), "{:?}", accepted.errors);
+
+        // Construct an introspection query of depth 15 using __schema.
+        // 1: __schema
+        // 2: types
+        // 3: fields
+        // 4: type
+        // 5..14: ofType (10 times)
+        // 15: name
+        // Total depth = 15 > GRAPHQL_MAX_DEPTH (14).
+        let of_types = (0..10).fold("name".to_string(), |acc, _| format!("ofType {{ {acc} }}"));
+        let query = format!("{{ __schema {{ types {{ fields {{ type {{ {of_types} }} }} }} }} }}");
+
+        let rejected = schema.execute(&query).await;
+        assert_eq!(rejected.data, async_graphql::Value::Null);
+        assert_eq!(rejected.errors.len(), 1);
+        assert_eq!(rejected.errors[0].message, "Query is nested too deep.");
+    }
+
    /// Every `.map_err(` in the GraphQL query/mutation resolvers must route
-    /// through the opaque helpers, or discard the error (`|_|`). Same source-
-    /// scrape pattern as `api::authz_guard` (#255 review).
+    /// through one of the curated helpers in this module, or discard the error
+    /// (`|_|`). Same source-scrape pattern as `api::authz_guard` (#255
+    /// review). The list is deliberately a whitelist of names rather than a
+    /// prefix match, so adding a new mapper is a decision a reviewer sees.
    #[test]
    fn every_graphql_map_err_uses_opaque_helpers() {
        for (file, src) in [
@@ -177,6 +938,8 @@ mod tests {
                let after = code[idx + ".map_err(".len()..].trim_start();
                let ok = after.starts_with("crate::graphql::graphql_db_err")
                    || after.starts_with("crate::graphql::graphql_app_err")
+                    || after.starts_with("crate::graphql::graphql_claim_conflict")
+                    || after.starts_with("crate::graphql::graphql_finish_conflict")
                    || after.starts_with("|_|")
                    || after.starts_with("|_ ");
                assert!(
@@ -22,6 +22,7 @@ pub struct MutationRoot;

 #[Object]
 impl MutationRoot {
+    #[graphql(complexity = "50 + child_complexity")]
    async fn create_task(
        &self,
        ctx: &Context<'_>,
@@ -36,6 +37,33 @@ impl MutationRoot {
        }
        let delegator_did = caller.to_string();
        let db = ctx.data_unchecked::<Arc<Db>>();
+        // #496: same write gate as the REST handler — a repo_id naming a
+        // hosted, non-quarantined repo admits tasks only from its owner, so a
+        // stranger cannot file under a foreign repo id. An id naming no hosted
+        // repo (unknown, or quarantined — hidden as if it did not exist) is
+        // stored verbatim as an opaque label; the gate does not bind an
+        // unresolved id to a repo. Task reads are caller-independent on this
+        // base (the tasks resolver is not repo-scoped), before and after this
+        // change. As in tasks.rs, the success-vs-denial split between a hosted
+        // repo id and an unknown one is itself a one-bit oracle — a stranger
+        // holding a private repo's id learns it is hosted here — but denying
+        // foreign repos requires drawing exactly that line.
+        if let Some(repo_id) = input.repo_id.as_deref() {
+            let record = db
+                .get_repo_by_id(repo_id)
+                .await
+                .map_err(crate::graphql::graphql_db_err)?;
+            if let Some(record) = record {
+                let quarantined = db
+                    .is_repo_quarantined(&record.id)
+                    .await
+                    .map_err(crate::graphql::graphql_db_err)?;
+                if !quarantined {
+                    crate::api::require_repo_owner(&record, caller)
+                        .map_err(crate::graphql::graphql_app_err)?;
+                }
+            }
+        }
        let now = Utc::now().to_rfc3339();
        let task = AgentTask {
            id: Uuid::new_v4().to_string(),
@@ -52,12 +80,17 @@ impl MutationRoot {
            updated_at: now,
            deadline: input.deadline,
        };
-        db.create_task(&task)
+        // Return the stored row, not the pre-insert input: create_task
+        // normalizes the assignee (whitespace-only becomes NULL/open), and a
+        // response that shows the raw input lies about the reservation.
+        let stored = db
+            .create_task(&task)
            .await
            .map_err(crate::graphql::graphql_db_err)?;
-        Ok(AgentTaskType::from(task))
+        Ok(AgentTaskType::from(stored))
    }

+    #[graphql(complexity = "50 + child_complexity")]
    async fn claim_task(
        &self,
        ctx: &Context<'_>,
@@ -73,20 +106,41 @@ impl MutationRoot {
        let assignee_did = caller.to_string();
        let db = ctx.data_unchecked::<Arc<Db>>();
        let tx = ctx.data_unchecked::<tokio::sync::broadcast::Sender<TaskEventBroadcast>>();
+        crate::api::tasks::get_claimable_task(db, &id, caller)
+            .await
+            .map_err(crate::graphql::graphql_app_err)?
+            .ok_or_else(|| async_graphql::Error::new("task not found"))?;
+        // Same classification REST's claim handler applies: a claim race is a
+        // client-safe conflict, a genuine sqlx failure stays opaque. Routing
+        // both transports through `task_write_conflict` is what stops a normal
+        // race from reaching a GraphQL caller as a generic database error
+        // (#327 review).
        let task = db
            .claim_task(&id, &assignee_did)
            .await
-            .map_err(crate::graphql::graphql_db_err)?;
-        let _ = tx.send(TaskEventBroadcast {
-            task_id: id,
-            old_status: "pending".to_string(),
-            new_status: "claimed".to_string(),
-            by_did: assignee_did,
-            at: Utc::now().to_rfc3339(),
-        });
+            .map_err(crate::graphql::graphql_claim_conflict)?;
+        let by_did = task
+            .assignee_did
+            .as_deref()
+            .map(crate::db::trim_assignee_did)
+            .unwrap_or(assignee_did.as_str())
+            .to_string();
+        crate::api::tasks::announce_task_event(
+            db,
+            tx,
+            TaskEventBroadcast {
+                task_id: id,
+                old_status: "pending".to_string(),
+                new_status: "claimed".to_string(),
+                by_did,
+                at: Utc::now().to_rfc3339(),
+            },
+        )
+        .await;
        Ok(AgentTaskType::from(task))
    }

+    #[graphql(complexity = "50 + child_complexity")]
    async fn complete_task(
        &self,
        ctx: &Context<'_>,
@@ -103,32 +157,47 @@ impl MutationRoot {
        let by_did = caller.to_string();
        let db = ctx.data_unchecked::<Arc<Db>>();
        let tx = ctx.data_unchecked::<tokio::sync::broadcast::Sender<TaskEventBroadcast>>();
-        // Authorize the actor: binding by_did to the signer is necessary but not
-        // sufficient — only the task's assignee may finish it.
-        let existing = db
-            .get_task(&id)
+        // Authorize the actor: the task must be visible to the caller (returning
+        // not found for invisible tasks so existence is not leaked), and only
+        // the task's assignee may finish it.
+        let existing = crate::api::tasks::get_visible_task(db, &id, Some(caller))
            .await
-            .map_err(crate::graphql::graphql_db_err)?
+            .map_err(crate::graphql::graphql_app_err)?
            .ok_or_else(|| async_graphql::Error::new("task not found"))?;
-        if !crate::api::did_matches(caller, existing.assignee_did.as_deref().unwrap_or_default()) {
+        if !crate::api::did_matches(
+            caller,
+            crate::db::trim_assignee_did(existing.assignee_did.as_deref().unwrap_or_default()),
+        ) {
            return Err(async_graphql::Error::new(
                "only the task assignee can complete it",
            ));
        }
        let task = db
-            .finish_task(&id, "completed", input.result.as_deref())
+            .finish_task(&id, "completed", input.result.as_deref(), caller)
            .await
-            .map_err(crate::graphql::graphql_db_err)?;
-        let _ = tx.send(TaskEventBroadcast {
-            task_id: id,
-            old_status: "claimed".to_string(),
-            new_status: "completed".to_string(),
-            by_did,
-            at: Utc::now().to_rfc3339(),
-        });
+            .map_err(crate::graphql::graphql_finish_conflict)?;
+        let by_did = task
+            .assignee_did
+            .as_deref()
+            .map(crate::db::trim_assignee_did)
+            .unwrap_or(by_did.as_str())
+            .to_string();
+        crate::api::tasks::announce_task_event(
+            db,
+            tx,
+            TaskEventBroadcast {
+                task_id: id,
+                old_status: "claimed".to_string(),
+                new_status: "completed".to_string(),
+                by_did,
+                at: Utc::now().to_rfc3339(),
+            },
+        )
+        .await;
        Ok(AgentTaskType::from(task))
    }

+    #[graphql(complexity = "50 + child_complexity")]
    async fn fail_task(
        &self,
        ctx: &Context<'_>,
@@ -145,29 +214,44 @@ impl MutationRoot {
        let by_did = caller.to_string();
        let db = ctx.data_unchecked::<Arc<Db>>();
        let tx = ctx.data_unchecked::<tokio::sync::broadcast::Sender<TaskEventBroadcast>>();
-        // Authorize the actor: only the task's assignee may fail it.
-        let existing = db
-            .get_task(&id)
+        // Authorize the actor: the task must be visible to the caller (returning
+        // not found for invisible tasks so existence is not leaked), and only
+        // the task's assignee may fail it.
+        let existing = crate::api::tasks::get_visible_task(db, &id, Some(caller))
            .await
-            .map_err(crate::graphql::graphql_db_err)?
+            .map_err(crate::graphql::graphql_app_err)?
            .ok_or_else(|| async_graphql::Error::new("task not found"))?;
-        if !crate::api::did_matches(caller, existing.assignee_did.as_deref().unwrap_or_default()) {
+        if !crate::api::did_matches(
+            caller,
+            crate::db::trim_assignee_did(existing.assignee_did.as_deref().unwrap_or_default()),
+        ) {
            return Err(async_graphql::Error::new(
                "only the task assignee can fail it",
            ));
        }
        let reason = input.reason.unwrap_or_default();
        let task = db
-            .finish_task(&id, "failed", Some(&reason))
+            .finish_task(&id, "failed", Some(&reason), caller)
            .await
-            .map_err(crate::graphql::graphql_db_err)?;
-        let _ = tx.send(TaskEventBroadcast {
-            task_id: id,
-            old_status: "claimed".to_string(),
-            new_status: "failed".to_string(),
-            by_did,
-            at: Utc::now().to_rfc3339(),
-        });
+            .map_err(crate::graphql::graphql_finish_conflict)?;
+        let by_did = task
+            .assignee_did
+            .as_deref()
+            .map(crate::db::trim_assignee_did)
+            .unwrap_or(by_did.as_str())
+            .to_string();
+        crate::api::tasks::announce_task_event(
+            db,
+            tx,
+            TaskEventBroadcast {
+                task_id: id,
+                old_status: "claimed".to_string(),
+                new_status: "failed".to_string(),
+                by_did,
+                at: Utc::now().to_rfc3339(),
+            },
+        )
+        .await;
        Ok(AgentTaskType::from(task))
    }
 }
@@ -186,57 +270,94 @@ mod tests {
            .join("; ")
    }

-    /// N2: GraphQL mutations require a verified signer and bind the acting DID to
-    /// it. Unsigned is rejected; a signer other than the claimed actor is
+    /// The `id` scalar returned by a successful `createTask` mutation.
+    fn created_id(resp: &Response) -> String {
+        assert!(resp.errors.is_empty(), "graphql errors: {:?}", resp.errors);
+        let async_graphql::Value::Object(obj) = &resp.data else {
+            panic!("data not an object: {:?}", resp.data);
+        };
+        let async_graphql::Value::Object(task) = obj.get("createTask").expect("createTask key")
+        else {
+            panic!("createTask not an object");
+        };
+        let async_graphql::Value::String(id) = task.get("id").expect("id key") else {
+            panic!("id not a string");
+        };
+        id.clone()
+    }
+
+    /// N2: every GraphQL mutation requires a verified signer and binds the acting
+    /// DID to it. Unsigned is rejected; a signer other than the claimed actor is
    /// rejected; the matching signer passes the auth gate.
    #[sqlx::test]
-    async fn mutation_requires_and_binds_signer(pool: PgPool) {
+    async fn each_mutation_requires_and_binds_signer(pool: PgPool) {
        let state = crate::test_support::test_state(pool).await;
        let schema = state.graphql_schema.as_ref();
-        let assignee = "did:key:zASSIGNEEAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
-        let q = format!(
-            r#"mutation {{ claimTask(id: "no-such-task", assigneeDid: "{assignee}") {{ id }} }}"#
-        );
+        let actor = "did:key:zGQLACTORAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
+        let other = "did:key:zGQLOTHERBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB";
+        let cases: Vec<(&str, String)> = vec![
+            (
+                "createTask",
+                format!(
+                    r#"mutation {{ createTask(delegatorDid: "{actor}", input: {{ kind: "build", capability: "repo:write" }}) {{ id }} }}"#
+                ),
+            ),
+            (
+                "claimTask",
+                format!(
+                    r#"mutation {{ claimTask(id: "no-such-task", assigneeDid: "{actor}") {{ id }} }}"#
+                ),
+            ),
+            (
+                "completeTask",
+                format!(
+                    r#"mutation {{ completeTask(id: "no-such-task", byDid: "{actor}", input: {{}}) {{ id }} }}"#
+                ),
+            ),
+            (
+                "failTask",
+                format!(
+                    r#"mutation {{ failTask(id: "no-such-task", byDid: "{actor}", input: {{}}) {{ id }} }}"#
+                ),
+            ),
+        ];

-        // 1. Unsigned → rejected before any DB work.
-        let resp = schema.execute(Request::new(&q)).await;
-        assert!(
-            errors(&resp).contains("authentication required"),
-            "unsigned mutation must be rejected: {}",
-            errors(&resp)
-        );
+        for (name, q) in &cases {
+            let resp = schema.execute(Request::new(q.as_str())).await;
+            assert!(
+                errors(&resp).contains("authentication required"),
+                "{name}: unsigned mutation must be rejected: {}",
+                errors(&resp)
+            );

-        // 2. Signed as someone other than the claimed assignee → rejected.
-        let resp = schema
-            .execute(Request::new(&q).data(AuthenticatedDid(
-                "did:key:zOTHERBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB".into(),
-            )))
-            .await;
-        assert!(
-            errors(&resp).contains("authenticated signer"),
-            "DID mismatch must be rejected: {}",
-            errors(&resp)
-        );
+            let resp = schema
+                .execute(Request::new(q.as_str()).data(AuthenticatedDid(other.into())))
+                .await;
+            assert!(
+                errors(&resp).contains("authenticated signer"),
+                "{name}: DID mismatch must be rejected: {}",
+                errors(&resp)
+            );

-        // 3. Signed as the claimed assignee → passes the auth gate. The missing
-        //    task is a business error from claim_task, not a sqlx fault, so the
-        //    actionable message must survive (not the opaque DB string) (#250).
-        let resp = schema
-            .execute(Request::new(&q).data(AuthenticatedDid(assignee.into())))
-            .await;
-        let errs = errors(&resp);
-        assert!(
-            !errs.contains("authentication required") && !errs.contains("authenticated signer"),
-            "matching signer must pass the auth gate: {errs}"
-        );
-        assert!(
-            errs.contains("task not claimable"),
-            "claim race / missing task must keep its business message: {errs}"
-        );
-        assert!(
-            !errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
-            "business error must not be rewritten as opaque DB error: {errs}"
-        );
+            let resp = schema
+                .execute(Request::new(q.as_str()).data(AuthenticatedDid(actor.into())))
+                .await;
+            let errs = errors(&resp);
+            assert!(
+                !errs.contains("authentication required") && !errs.contains("authenticated signer"),
+                "{name}: matching signer must pass the auth gate: {errs}"
+            );
+            if *name == "claimTask" {
+                assert!(
+                    errs.contains("task not found"),
+                    "claim race / missing task must keep its business message: {errs}"
+                );
+                assert!(
+                    !errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
+                    "business error must not be rewritten as opaque DB error: {errs}"
+                );
+            }
+        }
    }

    /// #250: mutation DB faults must be opaque; create_task hits agent_tasks.
@@ -272,6 +393,210 @@ mod tests {
        );
    }

+    /// #327 review: `claimTask`/`completeTask`/`failTask` called
+    /// `graphql_db_err` directly on db-layer business failures, so an ordinary
+    /// claim race or stale finish reached GraphQL clients as a generic
+    /// database error while REST clients got an actionable conflict. All three
+    /// now route through the shared `task_write_conflict` classifier.
+    ///
+    /// Deleting either `map_err` from a mutation turns this red: the message
+    /// becomes the raw anyhow text instead of the fixed conflict form.
+    #[sqlx::test]
+    async fn task_write_races_surface_as_conflicts_not_db_errors(pool: PgPool) {
+        let state = crate::test_support::test_state(pool).await;
+        let schema = state.graphql_schema.as_ref();
+        let assignee = "did:key:zGQLRACEASSIGNEEAAAAAAAAAAAAAAAAAAAAAAA";
+        let rival = "did:key:zGQLRACERIVALBBBBBBBBBBBBBBBBBBBBBBBBBB";
+        let delegator = "did:key:zGQLRACEDELEGATORCCCCCCCCCCCCCCCCCCCCC";
+        let now = chrono::Utc::now().to_rfc3339();
+
+        state
+            .db
+            .create_repo(&crate::db::RepoRecord {
+                id: "race-repo".into(),
+                name: "race".into(),
+                owner_did: delegator.into(),
+                description: None,
+                is_public: true,
+                default_branch: "main".into(),
+                created_at: chrono::Utc::now(),
+                updated_at: chrono::Utc::now(),
+                disk_path: "/tmp/race-repo".into(),
+                forked_from: None,
+                machine_id: None,
+            })
+            .await
+            .unwrap();
+        state
+            .db
+            .create_task(&crate::db::AgentTask {
+                id: "race-task".into(),
+                repo_id: Some("race-repo".into()),
+                kind: "build".into(),
+                status: "pending".into(),
+                delegator_did: delegator.into(),
+                assignee_did: None,
+                capability: "repo:write".into(),
+                ucan_token: None,
+                payload: None,
+                result: None,
+                created_at: now.clone(),
+                updated_at: now,
+                deadline: None,
+            })
+            .await
+            .unwrap();
+
+        // The assignee wins the claim.
+        let claim = |did: &str| {
+            format!(r#"mutation {{ claimTask(id: "race-task", assigneeDid: "{did}") {{ id }} }}"#)
+        };
+        let resp = schema
+            .execute(Request::new(claim(assignee)).data(AuthenticatedDid(assignee.into())))
+            .await;
+        assert!(resp.errors.is_empty(), "first claim: {}", errors(&resp));
+
+        // The rival loses the race: ineligible caller receives opaque not-found.
+        let resp = schema
+            .execute(Request::new(claim(rival)).data(AuthenticatedDid(rival.into())))
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.contains("task not found"),
+            "ineligible caller after task is claimed must receive opaque not-found: {errs}"
+        );
+        assert!(
+            !errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
+            "ineligible claim is not a database failure: {errs}"
+        );
+
+        // A second claim that reaches the conditional SQL write (e.g. assignee claiming
+        // again on an already-claimed task) exercises the conflict mapper:
+        let resp = schema
+            .execute(Request::new(claim(assignee)).data(AuthenticatedDid(assignee.into())))
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.contains("task not claimable: not found or already claimed"),
+            "a lost claim race must render the fixed conflict message: {errs}"
+        );
+        assert!(
+            !errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
+            "a claim race is not a database failure: {errs}"
+        );
+
+        // A stale finish on an already-completed task is the same class.
+        let complete = format!(
+            r#"mutation {{ completeTask(id: "race-task", byDid: "{assignee}", input: {{ result: "ok" }}) {{ id }} }}"#
+        );
+        let resp = schema
+            .execute(Request::new(&complete).data(AuthenticatedDid(assignee.into())))
+            .await;
+        assert!(resp.errors.is_empty(), "first complete: {}", errors(&resp));
+
+        let resp = schema
+            .execute(Request::new(&complete).data(AuthenticatedDid(assignee.into())))
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.contains("task not found or not in claimed state"),
+            "a stale complete must render the fixed conflict message: {errs}"
+        );
+        assert!(
+            !errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
+            "a stale complete is not a database failure: {errs}"
+        );
+
+        let fail = format!(
+            r#"mutation {{ failTask(id: "race-task", byDid: "{assignee}", input: {{ reason: "nope" }}) {{ id }} }}"#
+        );
+        let resp = schema
+            .execute(Request::new(&fail).data(AuthenticatedDid(assignee.into())))
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.contains("task not found or not in claimed state"),
+            "a stale fail must render the fixed conflict message: {errs}"
+        );
+        assert!(
+            !errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
+            "a stale fail is not a database failure: {errs}"
+        );
+    }
+
+    /// The other half of the same classification: a genuine SQL fault on the
+    /// write itself must stay opaque. Without this, routing conflicts through
+    /// `task_write_conflict` could be "fixed" by making every db failure a
+    /// client-visible conflict message.
+    #[sqlx::test]
+    async fn task_write_sql_faults_stay_opaque(pool: PgPool) {
+        let state = crate::test_support::test_state(pool.clone()).await;
+        let assignee = "did:key:zGQLOPAQUEASSIGNEEAAAAAAAAAAAAAAAAAAAAA";
+        let delegator = "did:key:zGQLOPAQUEDELEGATORBBBBBBBBBBBBBBBBBBB";
+        let now = chrono::Utc::now().to_rfc3339();
+        state
+            .db
+            .create_task(&crate::db::AgentTask {
+                id: "opaque-task".into(),
+                repo_id: None,
+                kind: "build".into(),
+                status: "pending".into(),
+                delegator_did: delegator.into(),
+                assignee_did: Some(assignee.into()),
+                capability: "repo:write".into(),
+                ucan_token: None,
+                payload: None,
+                result: None,
+                created_at: now.clone(),
+                updated_at: now,
+                deadline: None,
+            })
+            .await
+            .unwrap();
+
+        // Fault the write itself, not the schema. Dropping a column the
+        // visibility pre-check also SELECTs would fail in `get_visible_task`
+        // and never reach `graphql_claim_conflict`, so this test would pass
+        // against a claim mapper that leaks (#327 review). A BEFORE UPDATE
+        // trigger keeps every read valid and faults only inside
+        // `Db::claim_task`.
+        sqlx::raw_sql(
+            "CREATE FUNCTION gl_test_fault_task_update() RETURNS trigger
+             LANGUAGE plpgsql AS $fn$
+             BEGIN RAISE EXCEPTION 'gl_test_forced_update_fault'; END;
+             $fn$;
+             CREATE TRIGGER gl_test_fault_task_update
+             BEFORE UPDATE ON agent_tasks
+             FOR EACH ROW EXECUTE FUNCTION gl_test_fault_task_update();",
+        )
+        .execute(&pool)
+        .await
+        .unwrap();
+
+        let resp = state
+            .graphql_schema
+            .execute(
+                Request::new(format!(
+                    r#"mutation {{ claimTask(id: "opaque-task", assigneeDid: "{assignee}") {{ id }} }}"#
+                ))
+                .data(AuthenticatedDid(assignee.into())),
+            )
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
+            "a real sqlx fault on the write must stay opaque: {errs}"
+        );
+        assert!(
+            !errs.contains("task not claimable"),
+            "a write-time sqlx fault must not be reclassified as a claim race: {errs}"
+        );
+        assert!(
+            !errs.contains("gl_test_forced_update_fault") && !errs.contains("trigger"),
+            "database text leaked through the conflict mapper: {errs}"
+        );
+    }
+
    /// Adversarial-review GATE-1 (GraphQL): completing a task requires being its
    /// assignee, not merely signing as the by_did you pass. A signer who is not
    /// the assignee is rejected even though the by_did binding passes; the
@@ -294,7 +619,7 @@ mod tests {
            payload: None,
            result: None,
            created_at: now.clone(),
-            updated_at: now,
+            updated_at: now.clone(),
            deadline: None,
        };
        state.db.create_task(&task).await.expect("seed task");
@@ -311,14 +636,69 @@ mod tests {
            )
        };

-        // Stranger signs as themselves and passes byDid=self (so the signer
-        // binding passes), but is not the assignee → rejected by authorization.
+        // Stranger signs as themselves on a repo-less task they cannot see:
+        // invisible task returns "task not found" so existence is not leaked.
        let resp = schema
            .execute(Request::new(q(stranger)).data(AuthenticatedDid(stranger.into())))
            .await;
+        assert!(
+            errors(&resp).contains("task not found"),
+            "an invisible task must return not found, got: {}",
+            errors(&resp)
+        );
+
+        // Seed a task on a public repo that stranger CAN see, but is not assignee of:
+        let pub_repo = crate::db::RepoRecord {
+            id: "pub-r".into(),
+            name: "pub-r".into(),
+            owner_did: "did:key:zGQLDELEGATORCCCCCCCCCCCCCCCCCCCCCCCCCC".into(),
+            description: None,
+            is_public: true,
+            default_branch: "main".into(),
+            created_at: chrono::Utc::now(),
+            updated_at: chrono::Utc::now(),
+            disk_path: "/tmp/pub-r".into(),
+            forked_from: None,
+            machine_id: None,
+        };
+        state.db.create_repo(&pub_repo).await.expect("create repo");
+        let pub_task = crate::db::AgentTask {
+            id: "task-pub".into(),
+            repo_id: Some("pub-r".into()),
+            kind: "build".into(),
+            status: "pending".into(),
+            delegator_did: "did:key:zGQLDELEGATORCCCCCCCCCCCCCCCCCCCCCCCCCC".into(),
+            assignee_did: None,
+            capability: "repo:write".into(),
+            ucan_token: None,
+            payload: None,
+            result: None,
+            created_at: now.clone(),
+            updated_at: now,
+            deadline: None,
+        };
+        state
+            .db
+            .create_task(&pub_task)
+            .await
+            .expect("seed pub task");
+        state
+            .db
+            .claim_task("task-pub", assignee)
+            .await
+            .expect("claim pub task");
+
+        let q_pub = |actor: &str| {
+            format!(
+                r#"mutation {{ completeTask(id: "task-pub", byDid: "{actor}", input: {{}}) {{ id status }} }}"#
+            )
+        };
+        let resp = schema
+            .execute(Request::new(q_pub(stranger)).data(AuthenticatedDid(stranger.into())))
+            .await;
        assert!(
            errors(&resp).contains("assignee"),
-            "a non-assignee signer must be rejected: {}",
+            "a non-assignee signer on a visible task must be rejected: {}",
            errors(&resp)
        );

@@ -332,4 +712,341 @@ mod tests {
            errors(&resp)
        );
    }
+
+    /// #496 (GraphQL): createTask applies the same repo-ownership gate as the
+    /// REST handler — a stranger naming a hosted repo is rejected without
+    /// persisting anything, while the owner files. The quarantined and
+    /// unknown-id arms are pinned too, including that each accepted write
+    /// stores the supplied id verbatim, so this copy of the gate cannot regress
+    /// silently.
+    #[sqlx::test]
+    async fn create_task_rejects_foreign_repo_id(pool: PgPool) {
+        let state = crate::test_support::test_state(pool).await;
+        let owner = "did:key:zGQLTASKOWNERAAAAAAAAAAAAAAAAAAAAAAAAAA";
+        let stranger = "did:key:zGQLTASKSTRANGERBBBBBBBBBBBBBBBBBBBBBB";
+        let now = chrono::Utc::now();
+        let repo = crate::db::RepoRecord {
+            id: uuid::Uuid::new_v4().to_string(),
+            name: "gql-task-gate-repo".to_string(),
+            owner_did: owner.to_string(),
+            description: None,
+            is_public: true,
+            default_branch: "main".to_string(),
+            created_at: now,
+            updated_at: now,
+            disk_path: "/tmp/gql-task-gate-repo".to_string(),
+            forked_from: None,
+            machine_id: None,
+        };
+        state.db.create_repo(&repo).await.expect("seed repo");
+        let schema = state.graphql_schema.as_ref();
+
+        let q = |actor: &str| {
+            format!(
+                r#"mutation {{ createTask(delegatorDid: "{actor}", input: {{ kind: "build", capability: "repo:write", repoId: "{}" }}) {{ id }} }}"#,
+                repo.id
+            )
+        };
+
+        // Stranger signs as themselves (so the signer binding passes) but
+        // names the victim's repo → rejected by the ownership gate, leaking
+        // nothing about the repo.
+        let resp = schema
+            .execute(Request::new(q(stranger)).data(AuthenticatedDid(stranger.into())))
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.contains(crate::api::REPO_OWNER_ONLY_MESSAGE),
+            "a non-owner must not file under a foreign repo id (shared denial text): {errs}"
+        );
+
+        // The denial must not persist: no task row may carry the victim's
+        // repo id afterwards.
+        let stored = state
+            .db
+            .list_tasks_keyset(None, None, 50, None)
+            .await
+            .expect("list tasks");
+        assert!(
+            stored
+                .iter()
+                .all(|t| t.repo_id.as_deref() != Some(repo.id.as_str())),
+            "a denied write must not leave a task under the foreign repo id"
+        );
+
+        // The owner files under their own repo id.
+        let resp = schema
+            .execute(Request::new(q(owner)).data(AuthenticatedDid(owner.into())))
+            .await;
+        assert!(
+            errors(&resp).is_empty(),
+            "the owner should file against their repo: {}",
+            errors(&resp)
+        );
+
+        // Unknown repo ids stay accepted as opaque labels.
+        let q_unknown = format!(
+            r#"mutation {{ createTask(delegatorDid: "{stranger}", input: {{ kind: "build", capability: "repo:write", repoId: "no-such-repo-id" }}) {{ id }} }}"#
+        );
+        let resp = schema
+            .execute(Request::new(q_unknown).data(AuthenticatedDid(stranger.into())))
+            .await;
+        assert!(
+            errors(&resp).is_empty(),
+            "an id naming no hosted repo stays an opaque label: {}",
+            errors(&resp)
+        );
+        // Read back the row: the unresolved id must be stored verbatim, not
+        // dropped or rewritten to a repo it never named.
+        let stored = state
+            .db
+            .get_task(&created_id(&resp))
+            .await
+            .expect("get task")
+            .expect("task must be stored");
+        assert_eq!(
+            stored.repo_id.as_deref(),
+            Some("no-such-repo-id"),
+            "an unresolved id must be stored verbatim as a label"
+        );
+
+        // A quarantined repo id is treated as unknown (accepted, never a 403
+        // that would confirm the id names a real repo). This pins the
+        // `!quarantined` arm on this copy of the gate.
+        state
+            .db
+            .set_repo_quarantine(&repo.id, true)
+            .await
+            .expect("quarantine");
+        let resp = schema
+            .execute(Request::new(q(stranger)).data(AuthenticatedDid(stranger.into())))
+            .await;
+        assert!(
+            errors(&resp).is_empty(),
+            "a quarantined repo id must not 403 (existence oracle): {}",
+            errors(&resp)
+        );
+        // Read back the row: the quarantined id is stored verbatim as a label
+        // too, so the accepted write is not silently dropped.
+        let stored = state
+            .db
+            .get_task(&created_id(&resp))
+            .await
+            .expect("get task")
+            .expect("task must be stored");
+        assert_eq!(
+            stored.repo_id.as_deref(),
+            Some(repo.id.as_str()),
+            "a quarantined id must be stored verbatim as a label"
+        );
+    }
+
+    /// #496 (GraphQL): a repo-lookup fault in createTask's gate must stay
+    /// opaque. Dropping a column `get_repo_by_id` selects faults that lookup;
+    /// the resolver must return the fixed DB message and leak no schema text
+    /// (#250), as `create_task_db_error_message_is_opaque` pins for the
+    /// agent_tasks fault.
+    #[sqlx::test]
+    async fn create_task_gate_repo_lookup_fault_is_opaque(pool: PgPool) {
+        let state = crate::test_support::test_state(pool.clone()).await;
+        let owner = "did:key:zGQLTASKFAULTOWNERAAAAAAAAAAAAAAAAAAAAAAAA";
+        let now = chrono::Utc::now();
+        let repo = crate::db::RepoRecord {
+            id: uuid::Uuid::new_v4().to_string(),
+            name: "gql-task-gate-fault-repo".to_string(),
+            owner_did: owner.to_string(),
+            description: None,
+            is_public: true,
+            default_branch: "main".to_string(),
+            created_at: now,
+            updated_at: now,
+            disk_path: "/tmp/gql-task-gate-fault-repo".to_string(),
+            forked_from: None,
+            machine_id: None,
+        };
+        state.db.create_repo(&repo).await.expect("seed repo");
+
+        sqlx::query("ALTER TABLE repos DROP COLUMN machine_id")
+            .execute(&pool)
+            .await
+            .expect("drop machine_id");
+
+        let q = format!(
+            r#"mutation {{ createTask(delegatorDid: "{owner}", input: {{ kind: "build", capability: "repo:write", repoId: "{}" }}) {{ id }} }}"#,
+            repo.id
+        );
+        let resp = state
+            .graphql_schema
+            .execute(Request::new(&q).data(AuthenticatedDid(owner.into())))
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.contains(crate::graphql::GRAPHQL_DB_ERROR_MESSAGE),
+            "gate sqlx fault must be opaque: {errs}"
+        );
+        assert!(
+            !errs.contains("machine_id") && !errs.contains("column"),
+            "schema text leaked through the gate: {errs}"
+        );
+    }
+
+    /// Pre-assigned GraphQL claimTask must reject a stranger (no UCAN leak)
+    /// and admit the reserved assignee.
+    #[sqlx::test]
+    async fn claim_task_honors_preassigned_assignee(pool: PgPool) {
+        let state = crate::test_support::test_state(pool).await;
+        let reserved = "did:key:zGQLRESERVEDAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
+        let thief = "did:key:zGQLTHIEFBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB";
+        let now = chrono::Utc::now().to_rfc3339();
+        let task = crate::db::AgentTask {
+            id: "task-gql-reserved".into(),
+            repo_id: None,
+            kind: "build".into(),
+            status: "pending".into(),
+            delegator_did: "did:key:zGQLDELEGATORCCCCCCCCCCCCCCCCCCCCCCCCCC".into(),
+            assignee_did: Some(reserved.into()),
+            capability: "repo:write".into(),
+            ucan_token: Some("gql-ucan-secret".into()),
+            payload: None,
+            result: None,
+            created_at: now.clone(),
+            updated_at: now,
+            deadline: None,
+        };
+        state.db.create_task(&task).await.expect("seed");
+        let schema = state.graphql_schema.as_ref();
+
+        let q = |actor: &str| {
+            format!(
+                r#"mutation {{ claimTask(id: "task-gql-reserved", assigneeDid: "{actor}") {{ id status ucanToken }} }}"#
+            )
+        };
+
+        let resp = schema
+            .execute(Request::new(q(thief)).data(AuthenticatedDid(thief.into())))
+            .await;
+        let errs = errors(&resp);
+        // The claim gate hides existence: an ineligible claimer gets the same
+        // not-found as a missing task, not a reserved deny that would leak it.
+        assert!(
+            errs.contains("task not found"),
+            "stranger claim must read as not-found, got: {errs}"
+        );
+        assert!(
+            !errs.contains("gql-ucan-secret"),
+            "UCAN must not leak in GraphQL errors: {errs}"
+        );
+        if let Ok(data) = resp.data.into_json() {
+            let s = data.to_string();
+            assert!(
+                !s.contains("gql-ucan-secret"),
+                "UCAN must not leak in GraphQL data on failed claim: {s}"
+            );
+        }
+        // Parity with REST: failed steal must leave the row untouched.
+        let still = state
+            .db
+            .get_task("task-gql-reserved")
+            .await
+            .expect("get")
+            .expect("task exists");
+        assert_eq!(
+            still.status, "pending",
+            "status must stay pending after steal"
+        );
+        assert_eq!(
+            still.assignee_did.as_deref(),
+            Some(reserved),
+            "assignee_did must stay reserved after steal"
+        );
+
+        let resp = schema
+            .execute(Request::new(q(reserved)).data(AuthenticatedDid(reserved.into())))
+            .await;
+        assert!(
+            errors(&resp).is_empty(),
+            "reserved assignee must claim: {}",
+            errors(&resp)
+        );
+        let data = resp.data.into_json().expect("json data");
+        assert_eq!(data["claimTask"]["status"], "claimed");
+        assert_eq!(data["claimTask"]["ucanToken"], "gql-ucan-secret");
+    }
+
+    /// create_task normalizes a whitespace-only assignee to NULL (open task),
+    /// so the create response must show the stored row, not echo the input: a
+    /// client trusting the echoed "reservation" would believe a claimable task
+    /// is reserved. Before the fix the response was built from the pre-insert
+    /// struct and returned the padded input verbatim.
+    #[sqlx::test]
+    async fn create_task_reports_stored_assignee_not_input(pool: PgPool) {
+        let state = crate::test_support::test_state(pool).await;
+        let caller = "did:key:zGQLCREATORAAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
+        let schema = state.graphql_schema.as_ref();
+        let q = format!(
+            r#"mutation {{ createTask(delegatorDid: "{caller}", input: {{ kind: "build", capability: "repo:write", assigneeDid: "   " }}) {{ id assigneeDid }} }}"#
+        );
+        let resp = schema
+            .execute(Request::new(q).data(AuthenticatedDid(caller.into())))
+            .await;
+        let data = resp.data.into_json().expect("json data");
+        assert!(
+            data["createTask"]["assigneeDid"].is_null(),
+            "whitespace-only assigneeDid must come back null (stored shape): {data}"
+        );
+    }
+
+    /// A task seeded with a whitespace-padded assignee must still be
+    /// completable by that assignee over GraphQL: REST and Db::finish_task
+    /// trim the stored value before comparing, and the GraphQL gate must do
+    /// the same or padded reservations finish only through REST.
+    #[sqlx::test]
+    async fn graphql_finish_accepts_padded_stored_assignee(pool: PgPool) {
+        let state = crate::test_support::test_state(pool).await;
+        let assignee = "did:key:zGQLPADDEDAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
+        let now = chrono::Utc::now().to_rfc3339();
+        let task = crate::db::AgentTask {
+            id: "task-gql-padded".into(),
+            repo_id: None,
+            kind: "build".into(),
+            status: "pending".into(),
+            delegator_did: "did:key:zGQLDELEGATORDDDDDDDDDDDDDDDDDDDDDDDDDD".into(),
+            assignee_did: Some(format!("\t{assignee}\n")),
+            capability: "repo:write".into(),
+            ucan_token: None,
+            payload: None,
+            result: None,
+            created_at: now.clone(),
+            updated_at: now,
+            deadline: None,
+        };
+        state.db.create_task(&task).await.expect("seed");
+        let schema = state.graphql_schema.as_ref();
+
+        let claim = format!(
+            r#"mutation {{ claimTask(id: "task-gql-padded", assigneeDid: "{assignee}") {{ id status }} }}"#
+        );
+        let resp = schema
+            .execute(Request::new(claim).data(AuthenticatedDid(assignee.into())))
+            .await;
+        assert!(
+            errors(&resp).is_empty(),
+            "reserved assignee must claim a padded reservation: {}",
+            errors(&resp)
+        );
+
+        let finish = format!(
+            r#"mutation {{ completeTask(id: "task-gql-padded", byDid: "{assignee}", input: {{}}) {{ id status }} }}"#
+        );
+        let resp = schema
+            .execute(Request::new(finish).data(AuthenticatedDid(assignee.into())))
+            .await;
+        let errs = errors(&resp);
+        assert!(
+            errs.is_empty(),
+            "assignee must complete a padded reservation: {errs}"
+        );
+        let data = resp.data.into_json().expect("json data");
+        assert_eq!(data["completeTask"]["status"], "completed");
+    }
 }
@@ -11,15 +11,16 @@ pub struct SubscriptionRoot;

 #[Subscription]
 impl SubscriptionRoot {
-    /// Live ref-update stream. `/graphql/ws` is mounted outside the
-    /// `optional_signature` layer, so this resolver has NO caller identity and
-    /// cannot gate per-subscriber — it relays whatever enters the broadcast
+    /// Live ref-update stream. `/graphql/ws` is mounted under the
+    /// `optional_signature` layer and carries optional caller identity, but this
+    /// resolver does not inspect it — it relays whatever enters the broadcast
    /// channel to any anonymous client. Its visibility safety therefore rests
    /// entirely on the WRITE side: the push handler only sends a
    /// `RefUpdateBroadcast` for repos the anonymous public may read (inside its
    /// `if announce` block, `api/repos.rs`). This is a single-point invariant —
    /// any new sender to `ref_update_tx` MUST be `announce`-gated, or private-repo
    /// ref metadata leaks here to unauthenticated subscribers (#112/#114 class).
+    #[graphql(complexity = "50 + child_complexity")]
    async fn ref_updates(
        &self,
        ctx: &Context<'_>,
@@ -46,6 +47,7 @@ impl SubscriptionRoot {
        })
    }

+    #[graphql(complexity = "50 + child_complexity")]
    async fn task_events(
        &self,
        ctx: &Context<'_>,
@@ -11,6 +11,14 @@ pub struct RepoType {
    pub created_at: String,
 }

+#[derive(SimpleObject)]
+pub struct RepoPageType {
+    pub nodes: Vec<RepoType>,
+    pub has_next_page: bool,
+    /// Position after the last returned visible repository; null for an empty page.
+    pub end_cursor: Option<String>,
+}
+
 #[derive(SimpleObject, Clone)]
 pub struct AgentTaskType {
    pub id: String,
@@ -48,6 +56,65 @@ impl From<AgentTask> for AgentTaskType {
    }
 }

+/// Read-only projection of `AgentTask` for the `tasks`/`task` queries, as
+/// opposed to `AgentTaskType`, which the task mutations (`createTask`,
+/// `claimTask`, `completeTask`, `failTask` — all `require_signer`-gated)
+/// return. Identical except for the missing `ucan_token` (#268): a read
+/// surface never needs to echo it back, since the assignee already received it
+/// at delegation/claim time via the mutation response.
+#[derive(SimpleObject, Clone)]
+pub struct AgentTaskReadType {
+    pub id: String,
+    pub repo_id: Option<String>,
+    pub kind: String,
+    pub status: String,
+    pub delegator_did: String,
+    pub assignee_did: Option<String>,
+    pub capability: String,
+    pub payload: Option<String>,
+    pub result: Option<String>,
+    pub created_at: String,
+    pub updated_at: String,
+    pub deadline: Option<String>,
+}
+
+/// Wraps a `tasks` page with the same completion signals the REST list route
+/// exposes, so the two surfaces answer pagination identically.
+///
+/// `hasMore` and `incomplete` are separate because they are separate facts
+/// (#327 review): `hasMore` says more candidates remain, `incomplete` says
+/// this page is short *only* because the authorization scan hit its safety
+/// wall. `nextCursor` is present exactly when `hasMore` is true and is an
+/// opaque MAC'd token — it can name the last examined candidate, denied or
+/// not, without disclosing it, which is what lets a caller page past a denied
+/// window instead of stalling on it.
+#[derive(SimpleObject, Clone)]
+pub struct TaskPageType {
+    pub items: Vec<AgentTaskReadType>,
+    pub has_more: bool,
+    pub incomplete: bool,
+    pub next_cursor: Option<String>,
+}
+
+impl From<AgentTask> for AgentTaskReadType {
+    fn from(t: AgentTask) -> Self {
+        Self {
+            id: t.id,
+            repo_id: t.repo_id,
+            kind: t.kind,
+            status: t.status,
+            delegator_did: t.delegator_did,
+            assignee_did: t.assignee_did,
+            capability: t.capability,
+            payload: t.payload,
+            result: t.result,
+            created_at: t.created_at,
+            updated_at: t.updated_at,
+            deadline: t.deadline,
+        }
+    }
+}
+
 #[derive(SimpleObject, Clone)]
 pub struct RefUpdateType {
    pub repo: String,
@@ -1,4 +1,4 @@
-//! Prometheus metrics for the gitlawb node.
+//! Prometheus metrics for the Twigpine node.
 //!
 //! Exposes a small, deliberately conservative set of counters and histograms
 //! that cover the questions an operator actually asks of a live node:
@@ -11,10 +11,18 @@
 //!     `gitlawb_sync_queue_processed_total{status}`
 //!   * are webhooks reaching their endpoints? —
 //!     `gitlawb_webhook_deliveries_total{result}`
+//!   * is inbound gossip being admitted or shed, and for which reason?
+//!     `gitlawb_gossip_ingest_events_total{outcome}`
+//!   * is the gossip replay seen-set full, and therefore admitting events it
+//!     cannot deduplicate?
+//!     `gitlawb_gossip_replay_guard_saturated_total`
 //!   * how big are the packs we're sending and receiving? —
 //!     `gitlawb_pack_size_bytes`
 //!   * a single `gitlawb_info{version, did}` gauge = 1, for joins/dashboards
 //!   * currently-connected peer count — `gitlawb_peers_connected`
+//!   * reconciliation sweep gaps found and filled —
+//!     `gitlawb_reconciliation_gaps_found_total` /
+//!     `gitlawb_reconciliation_gaps_filled_total`
 //!
 //! All metrics live in a single process-wide registry initialized by
 //! [`init`]. Increment helpers (`record_push`, `record_auth_failure`, ...)
@@ -33,8 +41,8 @@
 use std::sync::OnceLock;

 use prometheus::{
-    Encoder, Histogram, HistogramOpts, IntCounterVec, IntGauge, IntGaugeVec, Opts, Registry,
-    TextEncoder,
+    Encoder, Histogram, HistogramOpts, IntCounter, IntCounterVec, IntGauge, IntGaugeVec, Opts,
+    Registry, TextEncoder,
 };

 /// The single, process-wide metrics registry. Initialized by [`init`].
@@ -49,8 +57,17 @@ static AUTH_SUCCESSES: OnceLock<IntCounterVec> = OnceLock::new();
 static AUTH_FAILURES: OnceLock<IntCounterVec> = OnceLock::new();
 static SYNC_PROCESSED: OnceLock<IntCounterVec> = OnceLock::new();
 static WEBHOOK_DELIVERIES: OnceLock<IntCounterVec> = OnceLock::new();
+static GOSSIP_INGEST: OnceLock<IntCounterVec> = OnceLock::new();
+static GOSSIP_REPLAY_GUARD_SATURATED: OnceLock<IntCounter> = OnceLock::new();
 static PACK_SIZE: OnceLock<Histogram> = OnceLock::new();
 static PEERS_CONNECTED: OnceLock<IntGauge> = OnceLock::new();
+static RECONCILIATION_GAPS_FOUND: OnceLock<IntCounter> = OnceLock::new();
+static RECONCILIATION_GAPS_FILLED: OnceLock<IntCounter> = OnceLock::new();
+/// 1 when the node started without its p2p identity because the key storage
+/// could not be used. The node keeps serving HTTP in that state and /health
+/// stays green, so without this the outage is visible only to whoever reads
+/// the startup log.
+static P2P_KEY_LOAD_FAILED: OnceLock<IntGauge> = OnceLock::new();

 /// One-time initializer. Builds the registry, registers every metric,
 /// and sets the constant `gitlawb_info` gauge. Idempotent — calling
@@ -168,6 +185,33 @@ fn init_inner(version: &str, node_did: &str) {
        .set(webhook_deliveries)
        .expect("set WEBHOOK_DELIVERIES once");

+    let gossip_ingest = IntCounterVec::new(
+        Opts::new(
+            "gitlawb_gossip_ingest_events_total",
+            "Total inbound gossip ref-update events and what the ingest path decided about each",
+        ),
+        &["outcome"],
+    )
+    .expect("gitlawb_gossip_ingest_events_total definition");
+    registry
+        .register(Box::new(gossip_ingest.clone()))
+        .expect("register gitlawb_gossip_ingest_events_total");
+    GOSSIP_INGEST
+        .set(gossip_ingest)
+        .expect("set GOSSIP_INGEST once");
+
+    let gossip_replay_guard_saturated = IntCounter::with_opts(Opts::new(
+        "gitlawb_gossip_replay_guard_saturated_total",
+        "Inbound gossip ref-updates admitted without being recorded because the replay seen-set was at capacity",
+    ))
+    .expect("gitlawb_gossip_replay_guard_saturated_total definition");
+    registry
+        .register(Box::new(gossip_replay_guard_saturated.clone()))
+        .expect("register gitlawb_gossip_replay_guard_saturated_total");
+    GOSSIP_REPLAY_GUARD_SATURATED
+        .set(gossip_replay_guard_saturated)
+        .expect("set GOSSIP_REPLAY_GUARD_SATURATED once");
+
    let pack_size = Histogram::with_opts(
        HistogramOpts::new(
            "gitlawb_pack_size_bytes",
@@ -202,6 +246,41 @@ fn init_inner(version: &str, node_did: &str) {
        .set(peers_connected)
        .expect("set PEERS_CONNECTED once");

+    let gaps_found = IntCounter::with_opts(Opts::new(
+        "gitlawb_reconciliation_gaps_found_total",
+        "Total reconciliation sweep gaps detected (objects that should be pinned but are not)",
+    ))
+    .expect("gitlawb_reconciliation_gaps_found_total definition");
+    registry
+        .register(Box::new(gaps_found.clone()))
+        .expect("register gitlawb_reconciliation_gaps_found_total");
+    RECONCILIATION_GAPS_FOUND
+        .set(gaps_found)
+        .expect("set RECONCILIATION_GAPS_FOUND once");
+
+    let gaps_filled = IntCounter::with_opts(Opts::new(
+        "gitlawb_reconciliation_gaps_filled_total",
+        "Total reconciliation sweep gaps successfully filled (objects pinned by the sweep)",
+    ))
+    .expect("gitlawb_reconciliation_gaps_filled_total definition");
+    registry
+        .register(Box::new(gaps_filled.clone()))
+        .expect("register gitlawb_reconciliation_gaps_filled_total");
+    RECONCILIATION_GAPS_FILLED
+        .set(gaps_filled)
+        .expect("set RECONCILIATION_GAPS_FILLED once");
+    let p2p_key_load_failed = IntGauge::with_opts(Opts::new(
+        "gitlawb_p2p_identity_key_load_failed",
+        "1 if the node is running without p2p because its identity key could not be loaded",
+    ))
+    .expect("gitlawb_p2p_identity_key_load_failed definition");
+    registry
+        .register(Box::new(p2p_key_load_failed.clone()))
+        .expect("register gitlawb_p2p_identity_key_load_failed");
+    P2P_KEY_LOAD_FAILED
+        .set(p2p_key_load_failed)
+        .expect("set P2P_KEY_LOAD_FAILED once");
+
    REGISTRY
        .set(registry)
        .expect("set REGISTRY once (init must be called exactly once)");
@@ -263,13 +342,75 @@ pub fn record_sync_processed(status: &str) {
 }

 /// Record one webhook delivery attempt outcome. `result` ∈
-/// {ok, http_error, network_error, skipped}.
+/// {ok, http_error, network_error, skipped, invalid_url}.
+/// `invalid_url` is a policy skip: the stored hook URL failed the public-URL
+/// check at delivery time and the POST was never sent.
 pub fn record_webhook_delivery(result: &str) {
    if let Some(c) = WEBHOOK_DELIVERIES.get() {
        c.with_label_values(&[result]).inc();
    }
 }

+/// Record what the gossip ingest path decided about one inbound ref-update.
+/// `outcome` ∈ {accepted, unsigned_admitted, write_failed, rejected,
+/// source_rate_limited, author_rate_limited, unsigned_source_rate_limited,
+/// replayed, stale_timestamp}.
+///
+/// `accepted` is reserved for signature-verified events. An unsigned event that
+/// survives the rolling-upgrade window is `unsigned_admitted`, so the
+/// authenticated-admission rate is not silently padded by legacy compatibility
+/// traffic; an operator can tell whether the fleet still relies on the
+/// compatibility allowance.
+///
+/// The three shed reasons stay separate labels rather than one `rate_limited`
+/// because they answer different operator questions: `source_rate_limited` is a
+/// forwarding peer over the pre-parse brake, `unsigned_source_rate_limited` is
+/// that same forwarder relaying unproven traffic, and `author_rate_limited`
+/// names a principal a signature actually proved. Collapsing them would make a
+/// flood of unauthenticated garbage indistinguishable from one registered peer
+/// pushing too hard, which is exactly the distinction an alert needs.
+///
+/// The label is the variant, never the free-form reason string that
+/// `Rejected`/`WriteFailed` carry: those are attacker-influenced and would blow
+/// up the label cardinality of a process-wide registry.
+pub fn record_gossip_ingest(outcome: &str) {
+    if let Some(c) = GOSSIP_INGEST.get() {
+        c.with_label_values(&[outcome]).inc();
+    }
+}
+
+/// Record one inbound gossip ref-update that the replay seen-set could not
+/// record because it was at capacity.
+///
+/// Label-less on purpose: there is one degradation mode and nothing about it
+/// varies per event, so a label would only add cardinality.
+///
+/// This is an alert-worthy signal, not a debug counter, and it is the ONLY way
+/// the degraded state is visible. At capacity the guard fails open, which means
+/// the event is admitted and `gitlawb_gossip_ingest_events_total{outcome=
+/// "accepted"}` counts it exactly as it counts a healthy admission. Fail-open is
+/// the right policy (a saturated set that dropped all fresh gossip would convert
+/// a loud resource attack into quiet mesh-wide censorship), but it is not a free
+/// one: while this counter is moving, a captured signature can again be replayed
+/// against the freshness window's full 10 minutes, spending a `peer_exists`
+/// round trip and a debit from the victim author's budget on every replay.
+pub fn record_gossip_replay_guard_saturated() {
+    if let Some(c) = GOSSIP_REPLAY_GUARD_SATURATED.get() {
+        c.inc();
+    }
+}
+
+/// Test-only: current `gitlawb_gossip_replay_guard_saturated_total` value (0 if
+/// the registry is not initialized). Lets the seen-set tests assert the
+/// saturation signal fires without scraping the encoded text.
+#[cfg(test)]
+pub fn replay_guard_saturated_count_for_test() -> u64 {
+    GOSSIP_REPLAY_GUARD_SATURATED
+        .get()
+        .map(|c| c.get())
+        .unwrap_or(0)
+}
+
 /// Record a pack body size observation (bytes).
 pub fn observe_pack_size(bytes: f64) {
    if let Some(h) = PACK_SIZE.get() {
@@ -284,6 +425,27 @@ pub fn set_peers_connected(count: i64) {
    }
 }

+/// Record reconciliation sweep gaps found (objects that should be pinned but are not).
+pub fn record_reconciliation_gaps_found(count: u64) {
+    if let Some(c) = RECONCILIATION_GAPS_FOUND.get() {
+        c.inc_by(count);
+    }
+}
+
+/// Record reconciliation sweep gaps filled (objects successfully pinned by the sweep).
+pub fn record_reconciliation_gaps_filled(count: u64) {
+    if let Some(c) = RECONCILIATION_GAPS_FILLED.get() {
+        c.inc_by(count);
+    }
+}
+
+/// Record whether the node is running without its p2p identity.
+pub fn set_p2p_key_load_failed(failed: bool) {
+    if let Some(g) = P2P_KEY_LOAD_FAILED.get() {
+        g.set(i64::from(failed));
+    }
+}
+
 /// Encode the registry as the standard Prometheus text exposition format.
 /// Returns an error if `init` was never called.
 pub fn encode() -> Result<String, prometheus::Error> {
@@ -321,6 +483,8 @@ mod tests {
            .expect("PUSHES set after init")
            .with_label_values(&["alice/repo"])
            .inc();
+        record_reconciliation_gaps_found(7);
+        record_reconciliation_gaps_filled(3);

        let body = encode().expect("encode should succeed after init");
        assert!(
@@ -335,6 +499,14 @@ mod tests {
            body.contains("gitlawb_pushes_total{repo=\"alice/repo\"} 1"),
            "expected the incremented counter to be visible in: {body}"
        );
+        assert!(
+            body.contains("gitlawb_reconciliation_gaps_found_total 7"),
+            "expected the reconciliation gaps-found counter to be visible in: {body}"
+        );
+        assert!(
+            body.contains("gitlawb_reconciliation_gaps_filled_total 3"),
+            "expected the reconciliation gaps-filled counter to be visible in: {body}"
+        );
    }

    /// #192 F4: `init` is idempotent and safe to call repeatedly. The panic that
@@ -365,6 +537,12 @@ mod tests {
        record_auth_failure("test/route", "test_reason");
        record_sync_processed("done");
        record_webhook_delivery("ok");
+        // `record_gossip_replay_guard_saturated` is deliberately NOT called
+        // here. It is a label-less process-wide counter and the seen-set test
+        // that asserts it asserts a before/after delta, so an increment from a
+        // second test running concurrently in this binary would be a flake with
+        // no diagnostic value.
+        record_gossip_ingest("accepted");
        observe_pack_size(1024.0);
        set_peers_connected(0);
    }
@@ -70,6 +70,14 @@ impl RateLimiter {
        self.window.min(Duration::from_secs(1))
    }

+    /// The configured window, exposed so a 429 can carry an honest
+    /// `retry-after`: a hardcoded value lies when the operator retunes the
+    /// window, and a client that honors the header would retry into another
+    /// 429 until the real window elapses.
+    pub fn window(&self) -> Duration {
+        self.window
+    }
+
    pub(crate) async fn check(&self, key: &str) -> bool {
        // max_requests == 0 means the limiter is disabled, not "block all".
        if self.max_requests == 0 {
@@ -121,11 +129,43 @@ impl RateLimiter {
        true
    }

-    /// Number of keys currently tracked. Tests use it to observe what a sweep
-    /// reclaimed; there is no production reader.
+    /// Non-consuming check: is this key ALREADY at its limit for the current window?
+    /// Unlike [`check`], it records nothing and never inserts a new key — used to shed
+    /// expensive preparatory work (e.g. the `/ipfs/{cid}` legacy scan's O(repos) DB
+    /// preload) BEFORE it runs, without perturbing the per-unit budget the consuming
+    /// `check` maintains (#173, F3). An unknown key or a disabled limiter is not
+    /// throttled. Prunes the key's expired timestamps as a side effect (keeps state
+    /// tidy) but adds none, so it cannot itself fill or grow the map.
+    pub(crate) async fn is_throttled(&self, key: &str) -> bool {
+        if self.max_requests == 0 {
+            return false;
+        }
+        let now = Instant::now();
+        let mut state = self.state.lock().await;
+        if let Some(window) = state.get_mut(key) {
+            window
+                .timestamps
+                .retain(|t| now.duration_since(*t) < self.window);
+            return window.timestamps.len() >= self.max_requests;
+        }
+        false
+    }
+
+    /// The configured key ceiling. Tests use it to assert that a limiter a
+    /// production factory builds is actually bounded, which flooding the real
+    /// ceiling in distinct keys would be far too slow to show.
    #[cfg(test)]
-    pub async fn tracked_keys(&self) -> usize {
-        self.state.lock().await.len()
+    pub fn max_keys(&self) -> usize {
+        self.max_keys
+    }
+
+    /// The configured per-key request cap. Tests use it to assert that a
+    /// limiter a production factory builds carries the cap it is documented
+    /// with; re-asserting the constant on its own cannot tell whether the
+    /// factory passed that constant or a different one.
+    #[cfg(test)]
+    pub fn max_requests(&self) -> usize {
+        self.max_requests
    }

    pub async fn cleanup(&self) {
@@ -137,6 +177,14 @@ impl RateLimiter {
            !w.timestamps.is_empty()
        });
    }
+
+    /// Number of distinct keys currently tracked. Test-only introspection so a
+    /// cross-module test can assert that a sweep actually evicted expired entries
+    /// and observe what it reclaimed. There is no production reader.
+    #[cfg(test)]
+    pub(crate) async fn tracked_keys(&self) -> usize {
+        self.state.lock().await.len()
+    }
 }

 /// Per-source concurrency cap derived from the write-pool size: one resolved client
@@ -253,12 +301,7 @@ pub async fn rate_limit_by_did(request: Request, next: Next) -> Response {

    if let (Some(limiter), Some(did)) = (limiter, did) {
        if !limiter.check(&did).await {
-            return (
-                StatusCode::TOO_MANY_REQUESTS,
-                [("retry-after", "60")],
-                "rate limit exceeded — try again later",
-            )
-                .into_response();
+            return too_many_requests(limiter.window().as_secs());
        }
    }

@@ -361,18 +404,72 @@ impl<S: Send + Sync> axum::extract::FromRequestParts<S> for PeerAddr {
    }
 }

+/// The message every per-IP brake answers with, shared so the GraphQL surface
+/// (which cannot return a 429 status inside a 200 GraphQL envelope) says the
+/// same thing as the REST routes.
+pub const RATE_LIMIT_MESSAGE: &str = "rate limit exceeded — try again later";
+
 /// The shared 429 response for the per-IP flood brakes. Route-agnostic: this
 /// middleware now serves the push path AND the peer-sync routes, so the message
 /// stays generic (the offending path is recorded in the warn log below).
-pub fn too_many_requests() -> Response {
+pub fn too_many_requests(retry_after_secs: u64) -> Response {
    (
        StatusCode::TOO_MANY_REQUESTS,
-        [("retry-after", "60")],
-        "rate limit exceeded — try again later",
+        [("retry-after", retry_after_secs.to_string())],
+        RATE_LIMIT_MESSAGE,
    )
        .into_response()
 }

+/// The per-IP task-read brake carried as GraphQL request data: the limiter plus
+/// the client key the transport already resolved.
+///
+/// `/graphql` is a single POST endpoint, so this brake cannot sit in middleware
+/// the way `task_read_routes` mounts `rate_limit_by_ip`: that would charge every
+/// unrelated query and every mutation against the task-read bucket. The two task
+/// resolvers debit it instead, which also prices an aliased query honestly — ten
+/// aliased `tasks` fields run the visibility gate ten times and pay ten slots,
+/// where one request-scoped debit would pay one (#327 review).
+/// Default cap on the number of task read fields (across all aliases) permitted
+/// in a single GraphQL request before rejecting further field executions (#327 review).
+pub const MAX_GRAPHQL_TASK_READS_PER_REQUEST: usize = 5;
+
+#[derive(Clone)]
+pub struct TaskReadBrake {
+    pub limiter: RateLimiter,
+    /// `None` when no key could be resolved at all (no trusted header and no
+    /// `ConnectInfo`, e.g. a synthetic test request). Never braked, matching
+    /// [`rate_limit_by_ip`].
+    pub key: Option<String>,
+    pub request_count: Arc<std::sync::atomic::AtomicUsize>,
+}
+
+impl TaskReadBrake {
+    /// Debit one slot for one task-read field. `true` = proceed.
+    pub async fn check(&self, field: &str) -> bool {
+        let Some(key) = self.key.as_deref() else {
+            return true;
+        };
+        let count = self
+            .request_count
+            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+        if count >= MAX_GRAPHQL_TASK_READS_PER_REQUEST {
+            tracing::warn!(
+                key = %key,
+                field = %field,
+                count = count + 1,
+                "per-request GraphQL task field limit exceeded"
+            );
+            return false;
+        }
+        if self.limiter.check(key).await {
+            return true;
+        }
+        tracing::warn!(key = %key, field = %field, "per-IP rate limit exceeded");
+        false
+    }
+}
+
 /// Throttle the git push path by resolved client IP. The socket peer address is
 /// read from `ConnectInfo` (see `into_make_service_with_connect_info` in
 /// `main`). Only skips the limiter when no key can be resolved at all.
@@ -387,7 +484,7 @@ pub async fn rate_limit_by_ip(request: Request, next: Next) -> Response {
        if let Some(key) = client_key(request.headers(), peer, limiter.trust) {
            if !limiter.limiter.check(&key).await {
                tracing::warn!(key = %key, path = %request.uri().path(), "per-IP rate limit exceeded");
-                return too_many_requests();
+                return too_many_requests(limiter.limiter.window().as_secs());
            }
        }
    }
@@ -8,6 +8,19 @@ use crate::git::repo_store::RepoStore;
 use crate::p2p::P2pHandle;
 use crate::rate_limit::RateLimiter;

+/// HKDF salt for [`AppState::derive_scan_token_key`]. A constant, not a secret: HKDF's
+/// salt is a domain qualifier, and the confidentiality of the derived key rests entirely
+/// on the node's private seed being the input keying material.
+const SCAN_TOKEN_KEY_SALT: &[u8] = b"gitlawb/hkdf-salt/ipfs-scan-token";
+
+/// HKDF `info` for [`AppState::derive_scan_token_key`]: the domain separation AND the
+/// rotation handle in one string. Nothing else in the node derives from this label, so
+/// the token key is unrelated to the signing key it shares an input with; bumping the
+/// trailing version rotates every node's token key on its next boot, which invalidates
+/// outstanding continuations (they simply fail to open and the caller restarts at the
+/// front) and needs no migration, no config, and no change to this function.
+const SCAN_TOKEN_KEY_INFO: &[u8] = b"gitlawb/ipfs-scan-token/v1";
+
 #[derive(Clone, Debug)]
 pub struct RefUpdateBroadcast {
    pub repo: String,
@@ -46,6 +59,10 @@ pub struct AppState {
    pub task_event_tx: tokio::sync::broadcast::Sender<TaskEventBroadcast>,
    /// GraphQL schema (queries + mutations + subscriptions)
    pub graphql_schema: Arc<crate::graphql::GitlawbSchema>,
+    /// MAC key for task-list continuation tokens, derived from the node
+    /// keypair seed. Held here (and handed to the GraphQL schema) so REST and
+    /// GraphQL mint and accept the same tokens.
+    pub task_cursor_key: crate::api::task_cursor::TaskCursorKey,
    /// Fly.io machine ID — used for fly-replay routing in multi-machine deployments
    pub machine_id: Option<String>,
    /// Centralized repo storage: local disk cache + optional Tigris backend
@@ -66,6 +83,102 @@ pub struct AppState {
    /// brake a push flood from a DID farm (one throwaway DID per repo), so the
    /// push path throttles on the resolved client IP instead.
    pub push_rate_limiter: RateLimiter,
+    /// Per-client-IP rate limiter for `close_issue`'s pre-lock snapshot path.
+    /// Distinct from `push_rate_limiter` so a flood of close attempts cannot
+    /// drain the receive-pack budget for the same source IP.
+    pub close_issue_rate_limiter: RateLimiter,
+    /// Per-client-IP ROUTE brake for `GET /ipfs/{cid}`: charged ONCE per request by the
+    /// `rate_limit_by_ip` middleware (server.rs), never inside the handler. It bounds
+    /// request RATE (the "requests per hour" contract of `GITLAWB_IPFS_RATE_LIMIT`) on
+    /// the non-farmable source IP, so an anonymous flood of the public route is capped.
+    /// The per-probe/per-walk WORK accounting the resolver does WITHIN a request draws
+    /// from the SEPARATE `ipfs_work_rate_limiter` below — the two cannot share one bucket
+    /// or a single request that spends a route token and then its own probe token off the
+    /// same bucket is admitted at the route and falsely shed mid-request (#173 round-10,
+    /// R6). Keyed by `push_limiter_trust`.
+    pub ipfs_rate_limiter: RateLimiter,
+    /// Per-client-IP WORK-budget limiter for the `GET /ipfs/{cid}` resolver's internal
+    /// fan-out: charged per legacy (NULL-provenance) PROBE (`acquire` + `cat-file`) and
+    /// per provenance-path WALK, and peeked non-consuming before the O(repos) legacy
+    /// preload. A legacy CID from the public pins index otherwise lets one request drive
+    /// O(repos) subprocess spawns and cold Tigris fetches, and repeat requests amplify
+    /// that across requests with zero limiter contact (INV-10, F3). Charging the work to
+    /// the non-farmable source IP bounds it. A bucket DISTINCT from the route brake above:
+    /// one request legitimately spends many work tokens (a full legacy scan is up to
+    /// `ipfs_max_legacy_probes` probes), so it must not double as the once-per-request
+    /// route bucket. Capacity is DERIVED from the route limit (`AppState::ipfs_work_budget`,
+    /// no separate operator knob), floored at the legacy-probe budget so a single default-
+    /// config deep search never self-throttles mid-scan. Keyed by `push_limiter_trust`.
+    pub ipfs_work_rate_limiter: RateLimiter,
+    /// Per-request ceiling on full-history reachability walks the CID resolver
+    /// may spawn (default `api::ipfs::MAX_HISTORY_WALKS_PER_REQUEST`). A field,
+    /// not a bare const, so tests can shrink it to exercise the cap cheaply;
+    /// production keeps the const default.
+    pub ipfs_max_history_walks: u32,
+    /// Per-request ceiling on legacy (NULL-provenance) repo probes in the CID
+    /// resolver's scan fallback (default `api::ipfs::MAX_LEGACY_PROBES_PER_REQUEST`).
+    /// Bounds the anonymous `acquire` + `cat-file` fan-out across the node (#173,
+    /// INV-10); a field for the same test-seam reason as `ipfs_max_history_walks`.
+    pub ipfs_max_legacy_probes: u32,
+    /// How many repo rows the CID resolver's legacy scan pulls per keyset page
+    /// (default `api::ipfs::LEGACY_SCAN_PAGE_ROWS`). Bounds the DATABASE-facing half
+    /// of the same fan-out `ipfs_max_legacy_probes` bounds on the probe side: without
+    /// it the scan materialized every repo row and every matching visibility rule
+    /// before spending a single probe (#173, INV-10). A field for the same test-seam
+    /// reason as the sibling caps.
+    pub ipfs_legacy_scan_page_rows: usize,
+    /// Per-request ceiling on how many repo ROWS the CID resolver's legacy scan may
+    /// fetch (default `api::ipfs::MAX_LEGACY_SCAN_ROWS_PER_REQUEST`, operator-tunable via
+    /// `GITLAWB_IPFS_MAX_LEGACY_SCAN_ROWS`). `ipfs_max_legacy_probes` bounds the PROBE
+    /// fan-out but only starts counting once a probe runs, and quarantine plus a
+    /// root-scope visibility deny both return before a probe or a visit is spent, so an
+    /// all-denying inventory paged the whole repo table at zero probes (#173 round 13, F2).
+    /// Truncating here sheds a retryable 503 carrying a sealed continuation token.
+    pub ipfs_max_legacy_scan_rows: usize,
+    /// Per-request ceiling on the BYTES of visibility rules the CID resolver's legacy
+    /// scan may retain (default `api::ipfs::MAX_LEGACY_SCAN_RULE_BYTES_PER_REQUEST`). The
+    /// row ceiling above bounds the row count but not the memory each row drags in: the
+    /// pager keeps every fetched page's rules for the whole request, and neither the
+    /// number of rules per repo nor the length of a rule's reader list is capped, so a
+    /// rule COUNT would be the wrong unit. Enforced by the rules QUERY
+    /// (`Db::list_visibility_rules_for_repos_bounded`) rather than by a sum taken once the
+    /// page has landed, so the oversized page is never materialized at all. Deliberately
+    /// NOT an operator knob (it is a
+    /// memory guard, not a reach tradeoff); a field only for the same test-seam reason as
+    /// the sibling caps.
+    pub ipfs_max_legacy_scan_rule_bytes: usize,
+    /// Key sealing the legacy scan's continuation tokens (INV-13), derived from the
+    /// node's persistent identity by [`AppState::derive_scan_token_key`].
+    ///
+    /// The token is minted from a FETCHED row on a scan that served nothing, so by
+    /// construction that row is a private or quarantined repo the caller may not read:
+    /// its `created_at` and its `id` (which carries the owner's DID) are withheld
+    /// fields. The token is therefore AEAD-SEALED, never signed plaintext and never
+    /// base64-of-plaintext, since integrity is not confidentiality.
+    ///
+    /// DERIVED, not random per boot. This reverses an earlier revision of this design,
+    /// which argued that derivation "would make old tokens valid across restarts for no
+    /// benefit and tie a throwaway transport secret to a long-lived signing key." Both
+    /// halves were wrong. Surviving a restart IS the benefit: the token is the ONLY way
+    /// a caller resumes a ladder, an unopenable one is treated as absent, and a caller
+    /// treated as absent silently restarts at the front of the scan. A node whose
+    /// inventory needs several ladder steps and which deploys more often than a caller
+    /// can climb therefore keeps that caller from ever reaching a holder, which
+    /// contradicts the reach bound the README states. And the tie to the signing key is
+    /// what the HKDF domain separation removes: the derived key is a one-way function of
+    /// the seed under an `info` string nothing else uses, so it is not the signing key
+    /// and cannot be worked back into one.
+    ///
+    /// Persisting a random key instead would need a schema change for a value the node
+    /// already has on disk, so derivation is also the smaller mechanism. Rotation is a
+    /// bump of the version in [`SCAN_TOKEN_KEY_INFO`].
+    pub ipfs_scan_token_key: Arc<[u8; 32]>,
+    /// Hard ceiling on the byte size of an object `GET /ipfs/{cid}` will buffer and
+    /// serve (default `api::ipfs::MAX_SERVED_OBJECT_BYTES`). The serve reads via a
+    /// blocking `git cat-file` and buffers the whole object; without a bound a large
+    /// public blob could exhaust memory or block a runtime worker (#173, F6, INV-10).
+    /// A field for the same test-seam reason as the sibling caps.
+    pub ipfs_max_served_object_bytes: u64,
    /// Which forwarded header (if any) the edge is trusted to set, for
    /// resolving the push limiter's client-IP key. See `GITLAWB_TRUSTED_PROXY`.
    /// Node-wide; also keys the two peer-sync limiters below.
@@ -81,6 +194,9 @@ pub struct AppState {
    /// sink as trigger and accepts unsigned requests from known peers, so it is
    /// braked too; each peer's distinct IP gets its own bucket.
    pub peer_write_rate_limiter: RateLimiter,
+    /// Per-client-IP federation budget: 12 requests per minute, with at most
+    /// 10,000 tracked source keys. Shared by router instances and swept below.
+    pub federated_rate_limiter: RateLimiter,
    /// Process-wide graceful-shutdown signal. Sending `true` causes every
    /// task that holds a `watch::Receiver` to exit at its next await point.
    /// Used by:
@@ -214,13 +330,16 @@ pub struct AppState {
    /// (`with_default_max_keys`, reject-before-insert) so a source-key farm cannot grow
    /// it (INV-15).
    pub git_ipfs_walk_per_caller: crate::rate_limit::PerCallerConcurrency,
-    /// Per-client-IP rate limiter for `GET /ipfs/{cid}`. The route is publicly
-    /// reachable and each request can drive a full-history git walk, so it carries a
-    /// per-IP flood brake in addition to the concurrency cap above — a rate limit
-    /// bounds request *rate*, the semaphore bounds concurrent slow holds (different
-    /// axes). Keyed on the resolved client IP via `push_limiter_trust`. Layered on the
-    /// `/ipfs` route via `rate_limit_by_ip`.
-    pub ipfs_rate_limiter: RateLimiter,
+    /// Per-client-IP rate limiter for the task read routes. `GET /api/v1/tasks`
+    /// and `GET /api/v1/tasks/{id}` are anonymously reachable and run the #268
+    /// visibility gate (a task lookup plus deduped-repo and visibility-rule
+    /// queries) before returning the opaque 404, so an unauthenticated prober
+    /// costs the node real work per request whether or not anything is visible.
+    /// Keyed on the resolved client IP via `push_limiter_trust`. Layered on
+    /// `task_read_routes` via `rate_limit_by_ip`.
+    /// The same per-IP hourly budget applies to GraphQL `tasks` / `task`
+    /// queries and WebSocket task queries via `TaskReadBrake`.
+    pub task_read_rate_limiter: RateLimiter,
    /// The `git` executable the served-git withheld-blob walk spawns. Production is
    /// `"git"` (resolved via PATH); injectable so a fake `git` can drive the walk's
    /// process-group teardown in handler tests without mutating the process-global
@@ -235,6 +354,24 @@ impl AppState {
        self.shutdown_tx.subscribe()
    }

+    /// Sweep expired entries from every per-IP/DID rate limiter. Driven by the
+    /// periodic cleanup task so a bounded limiter's key map sheds stale entries
+    /// instead of sitting near its cap until an inline capacity sweep reclaims
+    /// them. Every limiter on the state is swept here; adding a new limiter means
+    /// adding it to this list.
+    pub(crate) async fn sweep_rate_limiters(&self) {
+        self.rate_limiter.cleanup().await;
+        self.create_ip_rate_limiter.cleanup().await;
+        self.push_rate_limiter.cleanup().await;
+        self.close_issue_rate_limiter.cleanup().await;
+        self.ipfs_rate_limiter.cleanup().await;
+        self.ipfs_work_rate_limiter.cleanup().await;
+        self.sync_trigger_rate_limiter.cleanup().await;
+        self.peer_write_rate_limiter.cleanup().await;
+        self.task_read_rate_limiter.cleanup().await;
+        self.federated_rate_limiter.cleanup().await;
+    }
+
    /// Trigger graceful shutdown. Idempotent — calling more than once
    /// has no effect. Returns `true` if this call was the one that
    /// flipped the signal.
@@ -255,6 +392,130 @@ impl AppState {
    pub fn is_shutting_down(&self) -> bool {
        *self.shutdown_tx.borrow()
    }
+
+    /// Legacy-probe budget wired from the `GITLAWB_IPFS_MAX_LEGACY_PROBES` operator
+    /// knob. The knob seeds `ipfs_max_legacy_probes` at construction so it controls the
+    /// per-request legacy (NULL-provenance) probe fan-out it advertises. It deliberately
+    /// does NOT feed the history-walk ceiling: that is governed by
+    /// `ipfs_max_repos_walked` under a `MAX_PIN_SOURCES + 1` floor, because a value
+    /// below the floor truncates a provenanced request with a full source set into a
+    /// false 503. The knob is `usize`, the field `u32`; the range cap (1_048_576) keeps
+    /// the cast lossless.
+    pub(crate) fn ipfs_legacy_probe_budget(config: &crate::config::Config) -> u32 {
+        config.ipfs_max_legacy_probes as u32
+    }
+
+    /// Legacy-scan ROW budget wired from the `GITLAWB_IPFS_MAX_LEGACY_SCAN_ROWS` knob,
+    /// the same helper shape as the probe budget above so the knob cannot become a
+    /// silent no-op if a struct literal drifts back to the bare constant.
+    pub(crate) fn ipfs_legacy_scan_row_budget(config: &crate::config::Config) -> usize {
+        config.ipfs_max_legacy_scan_rows
+    }
+
+    /// The key sealing legacy-scan continuation tokens (INV-13), DERIVED from the node's
+    /// persistent identity rather than minted per boot.
+    ///
+    /// HKDF-SHA256 over the node's Ed25519 seed, the same key material
+    /// `load_or_create_keypair` reads back from its PKCS#8 PEM on every boot, so the
+    /// derived key is byte-identical after a restart or a rolling deploy.
+    ///
+    /// Two properties the derivation has to carry, both load-bearing:
+    ///
+    ///   * DOMAIN SEPARATION. The `info` string below is unique to this use, so the
+    ///     derived key is not the signing seed and is not any other secret derived from
+    ///     it. HKDF is one-way, so a leaked token key yields nothing about the signing
+    ///     key and cannot be turned against a signature.
+    ///   * A VERSION component, carried in the same `info` string
+    ///     ([`SCAN_TOKEN_KEY_INFO`]) alongside a fixed [`SCAN_TOKEN_KEY_SALT`]. Bumping
+    ///     the version rotates every token key on the next boot without touching the
+    ///     identity, the token format, or any caller, so rotation is a constant change
+    ///     rather than a fork of this derivation.
+    pub(crate) fn derive_scan_token_key(keypair: &Keypair) -> [u8; 32] {
+        use hmac::{Hmac, Mac};
+        type HmacSha256 = Hmac<sha2::Sha256>;
+
+        let seed = keypair.to_seed();
+        // HKDF-Extract: PRK = HMAC(salt, ikm).
+        let mut extract =
+            HmacSha256::new_from_slice(SCAN_TOKEN_KEY_SALT).expect("HMAC takes a key of any size");
+        extract.update(seed.as_slice());
+        let prk = extract.finalize().into_bytes();
+        // HKDF-Expand, one block: T(1) = HMAC(PRK, info || 0x01). One 32-byte output
+        // needs exactly one block of SHA-256, so there is no counter loop to get wrong.
+        let mut expand =
+            HmacSha256::new_from_slice(prk.as_slice()).expect("HMAC takes a key of any size");
+        expand.update(SCAN_TOKEN_KEY_INFO);
+        expand.update(&[0x01]);
+        expand.finalize().into_bytes().into()
+    }
+
+    /// Work-budget capacity for [`ipfs_work_rate_limiter`](Self#structfield.ipfs_work_rate_limiter)
+    /// (R6, KTD6), DERIVED from the route limit rather than a new operator knob. The route
+    /// limiter (`ipfs_rate_limiter`) charges once per request; this separate bucket absorbs
+    /// the resolver's per-probe/per-walk work charges so both the route "requests per hour"
+    /// contract and the amplification bound hold. Floor: at least one complete COMBINED
+    /// resolution per window, the provenance phase's walks plus a full legacy search
+    /// (probes plus page tolls), so a single default-config resolution cannot
+    /// self-throttle part-way and recreate the admit-then-429 for a legitimate caller.
+    /// `GITLAWB_IPFS_RATE_LIMIT=0`
+    /// disables the route brake and this derived bucket alike (a 0-capacity limiter admits
+    /// everything).
+    ///
+    /// The floor carries a WALK term (#173 round 15, F2). The provenance visibility walk
+    /// in `gate_and_serve` debits this SAME bucket (its `!legacy_scan` charge), once per
+    /// path-scoped source, before the legacy fallback runs at all, and the walk cap is
+    /// charged per phase. A floor counting only the search therefore under-sizes the
+    /// window by up to that cap exactly when the floor binds (a route limit set below
+    /// it): the provenance phase spends from the budget the floor reserved for the
+    /// search, the fallback 429s short of its configured reach, and the retry re-pays the
+    /// same provenance charges. The term is
+    /// `min(MAX_HISTORY_WALKS_PER_REQUEST, ipfs_max_repos_walked)` because that is what
+    /// the resolver's own `walk_cap` in `gate_and_serve` evaluates to. The two
+    /// expressions are separate, not one shared value: this one reads the CONSTANT
+    /// `crate::api::ipfs::MAX_HISTORY_WALKS_PER_REQUEST`, while `walk_cap` reads the
+    /// `AppState` seam `state.ipfs_max_history_walks`. They agree only because every
+    /// construction seeds that seam from that constant: the production one in `main.rs`,
+    /// and the two test ones in `auth`'s test module and `test_support`. Nothing
+    /// mechanically ties the two expressions beyond this comment and its counterpart at
+    /// `walk_cap` (no shared helper, no type, no assertion outside the one fixture that
+    /// pins the seam as a precondition), so the two `min()`s have to be moved together
+    /// by hand, and a construction that seeded the seam from anything else would size
+    /// the floor for a cap the resolver does not enforce. It reads the
+    /// CONSTANT and the CONFIG knob for the same reason the page term below reads the
+    /// constant page size: `ipfs_max_history_walks` is an `AppState` test seam, and
+    /// sizing a production floor from a seam would inflate the budget by whatever a test
+    /// chose.
+    ///
+    /// The PROBE term is the LEGACY-PROBE knob, not `ipfs_max_repos_walked`. Those were
+    /// one field before the walk cap and the probe budget were split apart, and sizing
+    /// the probe term off the walk cap would silently read 64 instead of 256. That is a
+    /// statement about which knob sizes the PROBE term; it is not a claim that the walk
+    /// cap has no place in the floor, since the walk term above is added to this one
+    /// rather than substituted for it.
+    ///
+    /// The floor also carries the scan's PAGE toll (#173 round 13, F2): every page the
+    /// legacy scan buys is charged to this same bucket, so a deep scan spends
+    /// `ceil(ipfs_max_legacy_scan_rows / LEGACY_SCAN_PAGE_ROWS)` tokens on pages on top
+    /// of its probes. Leaving those out would 429 an honest caller part-way down their
+    /// own continuation-token ladder, which is the F6 admit-then-429 shape in a new
+    /// place. The page term uses the CONSTANT page size, not `AppState`'s field: the
+    /// field is a test seam that shrinks pages to make paging observable, and sizing a
+    /// production floor from it would inflate the budget by whatever a test chose.
+    pub(crate) fn ipfs_work_budget(config: &crate::config::Config) -> usize {
+        if config.ipfs_rate_limit == 0 {
+            return 0;
+        }
+        let pages = config
+            .ipfs_max_legacy_scan_rows
+            .div_ceil(crate::api::ipfs::LEGACY_SCAN_PAGE_ROWS);
+        let walks = std::cmp::min(
+            crate::api::ipfs::MAX_HISTORY_WALKS_PER_REQUEST as usize,
+            config.ipfs_max_repos_walked,
+        );
+        config
+            .ipfs_rate_limit
+            .max(config.ipfs_max_legacy_probes + pages + walks)
+    }
 }

 /// Bounds the OUTSTANDING post-push encryption-task set by per-repo coalescing
@@ -1226,3 +1487,80 @@ mod repo_write_lease_tests {
        }
    }
 }
+
+#[cfg(test)]
+mod scan_token_key_tests {
+    use super::AppState;
+    use gitlawb_core::identity::Keypair;
+    use gitlawb_core::scan_token::{open_scan_token, seal_scan_token, ScanPosition};
+
+    const CID: &str = "bafkreiscantokenkeyfixtureaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
+
+    fn pos() -> ScanPosition {
+        ScanPosition {
+            created_at_key: "2020-01-01T12:00:00+00:00".to_string(),
+            id: "did:key:z6MkScanTokenOwner/repo".to_string(),
+            // 40 hex: the production shape, since the node's own repos are sha1.
+            sha256_hex: "a1b2c3d4e5f60718293a4b5c6d7e8f9012345678".to_string(),
+        }
+    }
+
+    /// The RESTART case. A token minted before a restart must still open after it, or a
+    /// caller laddering a deep inventory is silently returned to the front of the scan on
+    /// every rolling deploy and can never reach a holder buried past one window.
+    ///
+    /// The second key is derived from the identity RELOADED THROUGH ITS ON-DISK PEM,
+    /// which is exactly what `load_or_create_keypair` does on boot, so this exercises the
+    /// real restart path rather than a clone of the in-memory keypair.
+    #[test]
+    fn scan_token_key_survives_a_restart_of_the_same_identity() {
+        let kp = Keypair::generate();
+        let pem = kp.to_pem().expect("the identity serializes");
+        let reloaded = Keypair::from_pem(&pem).expect("the identity reloads");
+
+        let before = AppState::derive_scan_token_key(&kp);
+        let after = AppState::derive_scan_token_key(&reloaded);
+
+        let token = seal_scan_token(&before, CID, &pos(), i64::MAX - 1).expect("seal");
+        assert_eq!(
+            open_scan_token(&after, CID, &token, 0),
+            Some(pos()),
+            "a continuation minted before a restart must open after it: the node's \
+             identity is the same, so the derived sealing key must be too"
+        );
+    }
+
+    /// The must-not: a DIFFERENT node identity must derive a DIFFERENT key, so a token is
+    /// no more portable between nodes than it was when the key was random per boot.
+    #[test]
+    fn scan_token_key_does_not_open_under_a_different_identity() {
+        let mine = Keypair::generate();
+        let theirs = Keypair::generate();
+
+        let token = seal_scan_token(
+            &AppState::derive_scan_token_key(&mine),
+            CID,
+            &pos(),
+            i64::MAX - 1,
+        )
+        .expect("seal");
+        assert_eq!(
+            open_scan_token(&AppState::derive_scan_token_key(&theirs), CID, &token, 0),
+            None,
+            "a continuation sealed by one node must not open under another node's identity"
+        );
+    }
+
+    /// Domain separation, executed rather than asserted in prose: the derived token key
+    /// must not be the signing seed itself. Compromising a token key must not hand an
+    /// attacker the material that signs.
+    #[test]
+    fn scan_token_key_is_not_the_signing_seed() {
+        let kp = Keypair::generate();
+        assert_ne!(
+            AppState::derive_scan_token_key(&kp),
+            *kp.to_seed(),
+            "the token key must be a DERIVED secret, never the Ed25519 signing seed"
+        );
+    }
+}
@@ -14,6 +14,7 @@

 use std::collections::HashMap;
 use std::path::Path;
+use std::path::PathBuf;
 use std::sync::Arc;

 use gitlawb_core::identity::Keypair;
@@ -435,9 +436,10 @@ async fn process_batch(
                .await;
            }
            Err(e) => {
-                warn!(repo = %item.repo, origin = %origin_url, err = %e, "repo sync failed");
-                let _ = db.mark_sync_failed(&item.id).await;
-                crate::metrics::record_sync_processed("failed");
+                // Prune vs terminal classification lives in
+                // `handle_sync_item_error` (defer = stays pending for the
+                // next tick; anything else = failed).
+                handle_sync_item_error(db, &item, &origin_url, e).await;
            }
        }
    }
@@ -691,6 +693,284 @@ async fn existing_promisor_state(repo: &str) -> PromisorProbe {
    }
 }

+/// Byte-prefix check for the two node-managed blob namespaces exempt from
+/// the fail-closed commit walk (`visibility_pack.rs`): request markers
+/// (`refs/gitlawb/requests/`) and issue records (`refs/gitlawb/issues/`).
+/// Bytewise so non-UTF-8 refnames (which git accepts) are classified, not
+/// lossy-decoded into a false exempt match. Extracted so the predicate is
+/// unit-testable; [`prune_non_exempt_gitlawb_refs`] enforces it on import.
+fn gitlawb_ref_is_exempt(raw_refname: &[u8]) -> bool {
+    raw_refname.starts_with(b"refs/gitlawb/requests/")
+        || raw_refname.starts_with(b"refs/gitlawb/issues/")
+}
+
+/// Parse one `git show-ref` output line into its refname bytes: `<hex-sha>`
+/// (40-hex SHA-1 or 64-hex SHA-256, which a sha256 mirror inherits from its
+/// origin), one space, then the refname. Empty lines (trailing newline) yield
+/// `Ok(None)`; anything else-shaped is an anomaly the caller fails on rather
+/// than skips — silently passing an unclassifiable line could leave an evil
+/// ref imported. Extracted so malformed input is unit-testable without git.
+fn show_ref_refname(line: &[u8]) -> anyhow::Result<Option<&[u8]>> {
+    if line.is_empty() {
+        return Ok(None);
+    }
+    let Some(sp) = line.iter().position(|b| *b == b' ') else {
+        return Err(anyhow::anyhow!("git show-ref emitted an unparsable line"));
+    };
+    let (sha, raw) = (&line[..sp], &line[sp + 1..]);
+    let sha_ok = (sha.len() == 40 || sha.len() == 64) && sha.iter().all(|b| b.is_ascii_hexdigit());
+    if !sha_ok || raw.is_empty() {
+        return Err(anyhow::anyhow!("git show-ref emitted an unparsable line"));
+    }
+    Ok(Some(raw))
+}
+
+/// Delete imported `refs/gitlawb/*` refs outside the two exempt subtrees.
+///
+/// Mirrors import whatever the origin advertises (`+refs/*:refs/*`), and
+/// `transfer.hideRefs` is a serve-side knob: a repo's own `git fetch` never
+/// consults local hideRefs, and origins hide only `requests/` anyway. So a
+/// planted `refs/gitlawb/<other>` blob ref — creatable on any pre-gate node —
+/// lands on every mirror and then trips `assert_all_refs_are_commits` closed,
+/// wedging serving on each one. Prune at import instead: after clone/fetch,
+/// list `refs/gitlawb` and delete every ref outside the exempt subtrees
+/// (negative refspecs cannot re-include them, so post-import deletion is the
+/// mechanism that works).
+///
+/// Callers map a transient prune error to [`PruneFailed`] so the batch
+/// defers the row (stays `pending`, retried next tick) instead of failing
+/// it: a failed row would leave a possibly-wedged mirror unretried until
+/// the next peer announcement. Deterministic refusals (identity mismatch,
+/// malformed enumeration) map to [`PruneInvalid`], which fails terminally
+/// since no retry can clear them.
+///
+/// Enumeration is `git show-ref` (no pattern), parsed as raw bytes: refnames
+/// cannot contain newlines, so one line per ref is lossless even for
+/// non-UTF-8 names, which a lossy decode would miss. The `<sha> SP <refname>`
+/// structure (not `for-each-ref --format=%(refname)`, which carries no object
+/// id) is what lets the parser width-check 40-hex SHA-1 vs 64-hex SHA-256
+/// and reject anything else-shaped instead of skipping it.
+///
+/// Fail-closed throughout: the exit code is classified (only `1` with empty
+/// output means "no refs"; anything else non-zero, e.g. 128 on a missing
+/// repo, is fatal, since a silent skip would leave an imported evil ref in
+/// place while the sync reports success), and any line that is not
+/// `<hex-sha> SP <refname>` — 40-hex SHA-1 or 64-hex SHA-256, which a sha256
+/// mirror inherits from its origin — fails the sync rather than being
+/// skipped into an unpruned wedge.
+/// Bounded git spawn for the prune below: `kill_on_drop` plus a hard timeout
+/// so a wedged child cannot stall a mirror import or sweep forever. The
+/// refname delete runs on exact bytes, so args arrive as `OsStr`, not `str`.
+/// Non-PruneInvalid failures (including a timeout) wrap as PruneFailed and
+/// defer for retry, matching the transient classification below.
+async fn prune_git(
+    local_path: &Path,
+    args: &[&std::ffi::OsStr],
+    label: &str,
+) -> anyhow::Result<std::process::Output> {
+    let fut = async {
+        // allow-unbounded-git: bounded by the tokio::time::timeout wrapping
+        // this future below, plus kill_on_drop.
+        let mut cmd = tokio::process::Command::new("git");
+        cmd.kill_on_drop(true);
+        cmd.arg("-C").arg(local_path).args(args);
+        cmd.output().await
+    };
+    tokio::time::timeout(std::time::Duration::from_secs(30), fut)
+        .await
+        .map_err(|_| anyhow::anyhow!("{label} timed out"))?
+        .map_err(|e| anyhow::anyhow!("{label} failed to spawn: {e}"))
+}
+
+async fn prune_non_exempt_gitlawb_refs(local_path: &Path) -> anyhow::Result<()> {
+    let local_str = local_path.to_str().unwrap_or(".");
+    // Prove the path is itself a repo before trusting the enumeration:
+    // `git -C` walks up to the nearest ancestor repo, so inside a non-repo
+    // child (a partially-created mirror) `show-ref` would exit 0 with the
+    // ANCESTOR's refs and `update-ref -d` would then run against that repo.
+    // Require `rev-parse --absolute-git-dir` to resolve to the path itself
+    // (mirrors are bare repos, where the git dir IS the path).
+    let git_dir_out = prune_git(
+        local_path,
+        &[
+            std::ffi::OsStr::new("rev-parse"),
+            std::ffi::OsStr::new("--absolute-git-dir"),
+        ],
+        "git rev-parse",
+    )
+    .await?;
+    if !git_dir_out.status.success() {
+        let stderr = String::from_utf8_lossy(&git_dir_out.stderr);
+        return Err(anyhow::anyhow!(PruneInvalid(format!(
+            "mirror path is not a git repo (rev-parse failed): {stderr}"
+        ))));
+    }
+    let actual_dir = PathBuf::from(
+        String::from_utf8_lossy(&git_dir_out.stdout)
+            .trim()
+            .to_string(),
+    );
+    let canonical_path = local_path
+        .canonicalize()
+        .map_err(|e| anyhow::anyhow!(PruneInvalid(format!("mirror path does not resolve: {e}"))))?;
+    let canonical_git_dir = actual_dir
+        .canonicalize()
+        .map_err(|e| anyhow::anyhow!(PruneInvalid(format!("git dir does not resolve: {e}"))))?;
+    if canonical_git_dir != canonical_path {
+        return Err(anyhow::anyhow!(PruneInvalid(format!(
+            "mirror path {local_str} is not itself a repo (git dir resolves elsewhere); refusing to enumerate"
+        ))));
+    }
+    let out = prune_git(
+        local_path,
+        &[std::ffi::OsStr::new("show-ref")],
+        "git show-ref",
+    )
+    .await?;
+    // Classify on the exit code (the `existing_promisor_state` precedent):
+    // 0 parses, 1 with empty output is an empty repo, anything else is a
+    // fatal enumeration failure that must not read as "nothing to prune".
+    match out.status.code() {
+        Some(0) => {}
+        Some(1) if out.stdout.is_empty() => return Ok(()),
+        _ => {
+            let stderr = String::from_utf8_lossy(&out.stderr);
+            return Err(anyhow::anyhow!(
+                "git show-ref failed with code {:?}: {stderr}",
+                out.status.code()
+            ));
+        }
+    }
+    for line in out.stdout.split(|b| *b == b'\n') {
+        // Malformed lines are deterministic (same git, same output every
+        // tick): fail terminally via PruneInvalid, not transiently.
+        let Some(raw) =
+            show_ref_refname(line).map_err(|e| anyhow::anyhow!(PruneInvalid(e.to_string())))?
+        else {
+            continue;
+        };
+        if !(raw == b"refs/gitlawb" || raw.starts_with(b"refs/gitlawb/"))
+            || gitlawb_ref_is_exempt(raw)
+        {
+            continue;
+        }
+        // Delete by exact bytes (not the lossy display string below): a
+        // non-UTF-8 refname would otherwise survive the prune.
+        let refname_os = os_from_raw_refname(raw);
+        let shown = refname_os.to_string_lossy().into_owned();
+        let delete = prune_git(
+            local_path,
+            &[
+                std::ffi::OsStr::new("update-ref"),
+                std::ffi::OsStr::new("-d"),
+                refname_os.as_os_str(),
+                std::ffi::OsStr::new("--no-deref"),
+            ],
+            "git update-ref -d",
+        )
+        .await?;
+        if !delete.status.success() {
+            let stderr = String::from_utf8_lossy(&delete.stderr);
+            return Err(anyhow::anyhow!(
+                "git update-ref -d {shown} failed (mirror still carries a non-exempt refs/gitlawb ref): {stderr}"
+            ));
+        }
+        tracing::info!(
+            gitlawb_ref = %shown,
+            "mirror: pruned non-exempt refs/gitlawb ref on import"
+        );
+    }
+    Ok(())
+}
+
+/// Lossless bytes → OsString for `update-ref -d`: non-UTF-8 refnames must be
+/// deleted by exact bytes. Unix is the serving platform; elsewhere fall back
+/// to a lossy decode (best effort where exact bytes are unrepresentable).
+#[cfg(unix)]
+fn os_from_raw_refname(raw: &[u8]) -> std::ffi::OsString {
+    use std::os::unix::ffi::OsStringExt;
+    std::ffi::OsString::from_vec(raw.to_vec())
+}
+
+#[cfg(not(unix))]
+fn os_from_raw_refname(raw: &[u8]) -> std::ffi::OsString {
+    String::from_utf8_lossy(raw).into_owned().into()
+}
+
+/// Marker for mirror-prune failures: the clone/fetch succeeded, so the
+/// mirror is on disk, but a non-exempt `refs/gitlawb/*` ref may still be
+/// imported. Local-git conditions are transient (a retry can succeed), while
+/// `mark_sync_failed` is terminal until the next peer announcement — and a
+/// failed row would leave a possibly-wedged mirror unretried in the
+/// meantime. The batch error arm downcasts to this and defers the row
+/// (stays `pending`, retried on the next 30s tick, where fetch+prune are
+/// idempotent) instead of failing it.
+#[derive(Debug)]
+struct PruneFailed(String);
+
+impl std::fmt::Display for PruneFailed {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "mirror prune failed: {}", self.0)
+    }
+}
+
+impl std::error::Error for PruneFailed {}
+
+/// Marker for deterministic mirror-prune refusals: the identity check (path
+/// is not itself a repo) and malformed `show-ref` output cannot clear on
+/// retry — re-running fetch+prune every tick would spin forever on a
+/// permanently bad path. Unlike [`PruneFailed`], these fail the row
+/// terminally so the condition surfaces instead of looping silently.
+#[derive(Debug)]
+struct PruneInvalid(String);
+
+/// Preserve a deterministic prune refusal through the clone/fetch call
+/// sites: `prune_non_exempt_gitlawb_refs` already returns `PruneInvalid`
+/// for identity mismatches and malformed enumeration, and re-wrapping
+/// every error as `PruneFailed` would stringify that type away before
+/// `handle_sync_item_error` downcasts it — leaving a permanently invalid
+/// mirror deferred (fetch+prune every tick) instead of failed. Transient
+/// failures keep the `PruneFailed` deferral.
+fn wrap_prune_error(e: anyhow::Error) -> anyhow::Error {
+    match e.downcast::<PruneInvalid>() {
+        Ok(invalid) => anyhow::Error::new(invalid),
+        Err(other) => anyhow::anyhow!(PruneFailed(other.to_string())),
+    }
+}
+
+impl std::fmt::Display for PruneInvalid {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "mirror prune refused: {}", self.0)
+    }
+}
+
+impl std::error::Error for PruneInvalid {}
+
+/// Classify a mirror sync failure for one queue row. Transient prune
+/// failures ([`PruneFailed`]) defer: the mirror is on disk but may still
+/// carry a non-exempt refs/gitlawb ref, a transient local-git condition a
+/// retry can clear — failing the row would leave it unretried until the
+/// next peer announcement, wedging serving in the meantime. Deterministic
+/// refusals ([`PruneInvalid`]) and everything else fail the row terminally:
+/// retrying a permanently bad path every tick would spin forever instead of
+/// surfacing. Extracted so the deferral (not just the prune itself) is
+/// unit-testable.
+async fn handle_sync_item_error(
+    db: &Db,
+    item: &crate::db::SyncQueueItem,
+    origin_url: &str,
+    err: anyhow::Error,
+) {
+    if err.downcast_ref::<PruneFailed>().is_some() {
+        warn!(repo = %item.repo, origin = %origin_url, err = %err, "mirror prune failed; leaving sync row pending for retry");
+        crate::metrics::record_sync_processed("deferred");
+        return;
+    }
+    warn!(repo = %item.repo, origin = %origin_url, err = %err, "repo sync failed");
+    let _ = db.mark_sync_failed(&item.id).await;
+    crate::metrics::record_sync_processed("failed");
+}
+
 /// Mirror-clone a repo from a remote URL into a local bare repo.
 /// `Promisor` mode adds `--filter=blob:limit=10g`, which marks the repo a git
 /// promisor (so a pack with origin-omitted withheld blobs is accepted) while
@@ -701,6 +981,7 @@ async fn clone_repo(remote_url: &str, local_path: &Path, mode: MirrorMode) -> an
    if mode == MirrorMode::Promisor {
        args.push("--filter=blob:limit=10g");
    }
+    args.push("--");
    args.push(remote_url);
    args.push(local_str);

@@ -714,6 +995,13 @@ async fn clone_repo(remote_url: &str, local_path: &Path, mode: MirrorMode) -> an
        let stderr = String::from_utf8_lossy(&out.stderr);
        return Err(anyhow::anyhow!("git clone --mirror failed: {stderr}"));
    }
+    // A fresh mirror imports everything the origin advertises, including any
+    // planted non-exempt `refs/gitlawb/*` ref; prune before serving from it.
+    // Prune outcome flows through `wrap_prune_error`: transient failures
+    // defer (stays pending), deterministic refusals fail terminally.
+    prune_non_exempt_gitlawb_refs(local_path)
+        .await
+        .map_err(wrap_prune_error)?;
    Ok(())
 }

@@ -729,9 +1017,16 @@ async fn clone_repo(remote_url: &str, local_path: &Path, mode: MirrorMode) -> an
 async fn fetch_repo(local_path: &Path, remote_url: &str, mode: MirrorMode) -> anyhow::Result<()> {
    let local_str = local_path.to_str().unwrap_or(".");

-    git_run(&["-C", local_str, "remote", "set-url", "origin", remote_url]).await?;
+    git_run(&[
+        "-C", local_str, "remote", "set-url", "origin", "--", remote_url,
+    ])
+    .await?;

-    match mode {
+    // The fetch imports whatever the origin advertises via the stored
+    // `+refs/*:refs/*` refspec; prune non-exempt `refs/gitlawb/*` refs after
+    // every fetch so one planted origin ref cannot wedge serving on the
+    // mirror (see `prune_non_exempt_gitlawb_refs`).
+    let fetch_result = match mode {
        MirrorMode::Promisor => {
            git_run(&["-C", local_str, "config", "remote.origin.promisor", "true"]).await?;
            git_run(&[
@@ -771,7 +1066,13 @@ async fn fetch_repo(local_path: &Path, remote_url: &str, mode: MirrorMode) -> an
                git_run(&["-C", local_str, "fetch", "--prune", "origin"]).await
            }
        }
-    }
+    };
+    fetch_result?;
+    // Prune after every fetch (see above); outcome flows through
+    // `wrap_prune_error` like the clone path.
+    prune_non_exempt_gitlawb_refs(local_path)
+        .await
+        .map_err(wrap_prune_error)
 }

 #[cfg(test)]
@@ -1006,6 +1307,82 @@ mod tests {
        assert_eq!(probe, PromisorProbe::NotPromisor);
    }

+    /// #374: `clone_repo` interpolates the peer's origin URL into a `git
+    /// clone --mirror` argument list. Without a `--` delimiter git parses a
+    /// remote beginning with `-` as an option instead of a repository.
+    ///
+    /// The delimiter is what forces the same string to be a repository, and
+    /// git says so in its own words: it reports the whole argument as a
+    /// repository it cannot find. Undelimited, git consumes it as an option
+    /// and fails against the *destination* instead (`fatal: repository '<dest>' does not exist`),
+    /// never naming the injected string — which is why this assertion is RED
+    /// before the fix and GREEN after.
+    ///
+    /// The payload is deliberately free of `:` and spaces: a colon would make
+    /// git read the argument as an ssh-style `host:path` URL and report a
+    /// hostname rather than a repository, which is a different code path.
+    #[tokio::test]
+    async fn clone_never_parses_a_remote_as_a_git_option() {
+        let td = TempDir::new().unwrap();
+        let dest = td.path().join("mirror.git");
+        let injected = "--upload-pack=false";
+
+        let err = clone_repo(injected, &dest, MirrorMode::Plain)
+            .await
+            .expect_err("a remote that is not a repository must fail the clone")
+            .to_string();
+
+        assert!(
+            err.contains(injected),
+            "git must report the argument as a repository, not consume it as an option: {err}"
+        );
+    }
+
+    /// #374: `fetch_repo` interpolates the peer's origin URL into
+    /// `git remote set-url`, a second argv construction with the same contract as
+    /// the clone. It needs its own proof: the clone test cannot reach this path.
+    ///
+    /// The assertion is on stored state rather than on the error string, and that
+    /// is deliberate. `git_run` formats its failure as `git {args:?} failed:
+    /// {stderr}`, so the injected value is in the message via the debug-printed
+    /// argv whether or not the delimiter is present — asserting on containment
+    /// would pass with `--` deleted. Delimited, `set-url` succeeds and stores the
+    /// value verbatim; undelimited, git exits 129 with `unknown option` and leaves
+    /// the previous URL in place, so this goes red without the fix.
+    ///
+    /// The fetch that follows the rewrite fails, because the stored URL is not a
+    /// repository. That is expected and ignored: the property under test is what
+    /// `set-url` wrote, not whether the subsequent fetch could succeed.
+    #[tokio::test]
+    async fn fetch_never_parses_a_remote_as_a_git_option() {
+        let td = TempDir::new().unwrap();
+        let local = td.path().join("mirror.git");
+        let local_str = local.to_str().unwrap();
+        let placeholder = "http://example.invalid/x.git";
+
+        assert!(Command::new("git")
+            .args(["init", "-q", "--bare", local_str])
+            .status()
+            .unwrap()
+            .success());
+        assert!(Command::new("git")
+            .args(["-C", local_str, "remote", "add", "origin", placeholder])
+            .status()
+            .unwrap()
+            .success());
+
+        let injected = "--upload-pack=false";
+        let _ = fetch_repo(&local, injected, MirrorMode::Plain).await;
+
+        assert_eq!(
+            git_config_get(local_str, "remote.origin.url")
+                .await
+                .as_deref(),
+            Some(injected),
+            "set-url must store an option-shaped remote verbatim, not reject it as an option"
+        );
+    }
+
    #[tokio::test]
    async fn probe_reports_unknown_on_git_error() {
        // A path git cannot resolve as a repo at all (exit 128) is an indeterminate
@@ -1250,6 +1627,18 @@ mod tests {
        .await;
    }

+    /// Use a rooted path without a Windows drive prefix in the remote fixture.
+    /// The destination still resolves on the temp directory's current drive,
+    /// while the remote's relative path contains no illegal colon component.
+    fn absolute_slug_path(path: &Path) -> String {
+        assert!(path.is_absolute());
+        path.components()
+            .filter(|part| !matches!(part, std::path::Component::Prefix(_)))
+            .collect::<std::path::PathBuf>()
+            .to_string_lossy()
+            .replace('\\', "/")
+    }
+
    #[sqlx::test]
    async fn process_batch_rejects_slug_escaping_repos_dir(pool: PgPool) {
        // The verified escape from #272: `PathBuf::join` discards everything
@@ -1265,12 +1654,13 @@ mod tests {
        // fails on but leaves the parent it created, so assert on both.
        let outside = TempDir::new().unwrap();
        let escape_dir = outside.path().join("nest");
-        let slug = format!("a/{}/escape", escape_dir.display());
+        let escape_path = absolute_slug_path(&escape_dir);
+        let slug = format!("a/{escape_path}/escape");
        let escape_target = escape_dir.join("escape.git");

        // Serve the composed URL for real, so a run without the guard genuinely
        // clones outside the root rather than merely failing at git.
-        let rel = format!("a{}/escape", escape_dir.display());
+        let rel = format!("a{escape_path}/escape");
        let (_remote, peer_url) = rooted_remote(&[&rel]);

        let did = "did:key:z6MkAttacker";
@@ -1390,6 +1780,7 @@ mod tests {
    // ── canonical containment before the git call (issue #272) ───────────────

    /// Every ref in `repo`, as one string, for a before/after comparison.
+    #[cfg(unix)]
    fn refs_of(repo: &Path) -> String {
        let out = Command::new("git")
            .args(["-C", repo.to_str().unwrap(), "for-each-ref"])
@@ -1722,6 +2113,7 @@ mod tests {
    /// starvation tests is fixed rather than dependent on how fast the loop
    /// runs. Two rows enqueued in the same microsecond would otherwise order
    /// arbitrarily.
+    #[cfg(unix)]
    async fn enqueue_at(db: &Db, pool: &PgPool, repo: &str, did: &str, enqueued_at: &str) {
        enqueue(db, repo, did).await;
        sqlx::query("UPDATE sync_queue SET enqueued_at = $1 WHERE repo = $2")
@@ -1915,7 +2307,8 @@ mod tests {

        let outside = TempDir::new().unwrap();
        let escape_dir = outside.path().join("nest");
-        let slug = format!("a/{}/gitlawb-probe", escape_dir.display());
+        let escape_path = absolute_slug_path(&escape_dir);
+        let slug = format!("a/{escape_path}/gitlawb-probe");
        let escape_target = escape_dir.join("gitlawb-probe.git");

        // Two things this fixture must get right or the test is green for the
@@ -1926,7 +2319,7 @@ mod tests {
        // run without the guard genuinely clones outside the root instead of
        // just failing at git. Db::upsert_peer cannot seed this row: it gates on
        // is_public_http_url, which rejects file://.
-        let rel = format!("a{}/gitlawb-probe", escape_dir.display());
+        let rel = format!("a{escape_path}/gitlawb-probe");
        let (_remote, peer_url) = rooted_remote(&[&rel]);
        let did = "did:key:z6MkAttacker";
        seed_local_peer(&pool, did, &peer_url).await;
@@ -2012,4 +2405,463 @@ mod tests {
        assert!(object_count(&mirror) > 0, "mirror has no objects");
        assert_eq!(sync_status(&pool, "z6Mkfoo/hello").await, "done");
    }
+
+    /// The import-exemption predicate is bytewise over both node-managed
+    /// namespaces and nothing else. Non-UTF-8 bytes never match (git accepts
+    /// them, so a lossy decode could false-exempt). Narrowing either prefix
+    /// turns these red.
+    #[test]
+    fn gitlawb_import_exemption_covers_both_namespaces_by_bytes() {
+        for exempt in [
+            &b"refs/gitlawb/requests/550e8400-e29b-41d4-a716-446655440000"[..],
+            &b"refs/gitlawb/issues/abc123"[..],
+        ] {
+            assert!(
+                gitlawb_ref_is_exempt(exempt),
+                "import must keep {}",
+                String::from_utf8_lossy(exempt)
+            );
+        }
+        for pruned in [
+            &b"refs/gitlawb/evil"[..],
+            &b"refs/gitlawb"[..],
+            &b"refs/gitlawbfoo"[..],
+            &b"refs/heads/main"[..],
+            &b""[..],
+            // Raw 0xFF byte: never an exempt prefix, always pruned.
+            &b"refs/gitlawb/\xffvil"[..],
+        ] {
+            assert!(
+                !gitlawb_ref_is_exempt(pruned),
+                "import must prune {}",
+                String::from_utf8_lossy(pruned)
+            );
+        }
+    }
+
+    /// End-to-end prune on a scratch bare repo: a planted non-exempt
+    /// `refs/gitlawb/*` blob ref is deleted, exempt `requests/`/`issues/`
+    /// blob refs and content refs survive. Removing the prune call from
+    /// `clone_repo`/`fetch_repo` leaves the evil ref in place, which is what
+    /// wedges serving on mirrors — this test plants exactly that state and
+    /// proves the prune removes it.
+    #[test]
+    fn prune_removes_non_exempt_gitlawb_refs_and_keeps_exempt() {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .expect("test runtime");
+        rt.block_on(async {
+            let td = TempDir::new().unwrap();
+            let work = td.path().join("work");
+            let bare = td.path().join("m.git");
+            std::fs::create_dir_all(&work).unwrap();
+            // A real commit for the content branch (branch refs must point
+            // at commits; the evil/exempt refs below point at blobs).
+            std::fs::write(work.join("f.txt"), b"x\n").unwrap();
+            let run_work = |args: &[&str]| {
+                assert!(
+                    Command::new("git")
+                        .args(args)
+                        .current_dir(&work)
+                        .status()
+                        .unwrap()
+                        .success(),
+                    "git {args:?} failed"
+                );
+            };
+            run_work(&["init", "-q"]);
+            run_work(&["config", "user.email", "t@t"]);
+            run_work(&["config", "user.name", "t"]);
+            run_work(&["add", "f.txt"]);
+            run_work(&["commit", "-qm", "init"]);
+            assert!(
+                Command::new("git")
+                    .args([
+                        "clone",
+                        "-q",
+                        "--bare",
+                        work.to_str().unwrap(),
+                        bare.to_str().unwrap()
+                    ])
+                    .current_dir(td.path())
+                    .status()
+                    .unwrap()
+                    .success(),
+                "git clone --bare failed"
+            );
+            let run = |args: &[&str]| {
+                assert!(
+                    Command::new("git")
+                        .args(args)
+                        .current_dir(&bare)
+                        .status()
+                        .unwrap()
+                        .success(),
+                    "git {args:?} failed"
+                );
+            };
+            let run_out = |args: &[&str]| {
+                let out = Command::new("git")
+                    .args(args)
+                    .current_dir(&bare)
+                    .output()
+                    .unwrap();
+                assert!(out.status.success(), "git {args:?} failed");
+                out.stdout
+            };
+            run(&["init", "-q", "--bare", "."]);
+            std::fs::write(bare.join("blob-body"), b"marker\n").unwrap();
+            let blob = String::from_utf8_lossy(&run_out(&["hash-object", "-w", "blob-body"]))
+                .trim()
+                .to_string();
+            // The clone carries the content branch; plant the evil ref and
+            // both exempt refs (all blob-pointing, like node metadata).
+            run(&["update-ref", "refs/gitlawb/evil", &blob]);
+            run(&["update-ref", "refs/gitlawb/requests/r1", &blob]);
+            run(&["update-ref", "refs/gitlawb/issues/i1", &blob]);
+
+            prune_non_exempt_gitlawb_refs(&bare)
+                .await
+                .expect("prune succeeds");
+
+            let raw_refs = run_out(&["for-each-ref", "--format=%(refname)"]);
+            let refs = String::from_utf8_lossy(&raw_refs);
+            let set: std::collections::HashSet<&str> = refs
+                .lines()
+                .map(str::trim)
+                .filter(|l| !l.is_empty())
+                .collect();
+            assert!(
+                !set.contains("refs/gitlawb/evil"),
+                "non-exempt blob ref must be pruned, still present: {set:?}"
+            );
+            for keep in ["refs/gitlawb/requests/r1", "refs/gitlawb/issues/i1"] {
+                assert!(set.contains(keep), "{keep} must survive the prune: {set:?}");
+            }
+            assert!(
+                set.iter().any(|r| r.starts_with("refs/heads/")),
+                "content branch ref must survive the prune: {set:?}"
+            );
+        });
+    }
+
+    /// `show-ref` line parser: 40-hex SHA-1 and 64-hex SHA-256 lines yield
+    /// the refname, empty lines yield None, and anything else-shaped errors
+    /// (the caller fails the sync rather than skipping into an unpruned
+    /// wedge). Reverting to a 40-only width check turns the sha256 case red.
+    #[test]
+    fn show_ref_line_parser_accepts_both_hash_widths_and_rejects_junk() {
+        // 40-hex SHA-1 line.
+        let sha1_line = format!("{} refs/gitlawb/evil", "a".repeat(40));
+        assert_eq!(
+            show_ref_refname(sha1_line.as_bytes()).unwrap(),
+            Some(b"refs/gitlawb/evil".as_slice())
+        );
+        // 64-hex SHA-256 line (a sha256 mirror inherits the width).
+        let sha256_line = format!("{} refs/gitlawb/evil", "b".repeat(64));
+        assert_eq!(
+            show_ref_refname(sha256_line.as_bytes()).unwrap(),
+            Some(b"refs/gitlawb/evil".as_slice())
+        );
+        // Non-UTF-8 refname passes through as exact bytes.
+        let mut raw = vec![b'c'; 40];
+        raw.extend_from_slice(b" refs/gitlawb/\xffvil");
+        assert_eq!(
+            show_ref_refname(&raw).unwrap(),
+            Some(b"refs/gitlawb/\xffvil".as_slice())
+        );
+        // Trailing newline residue.
+        assert_eq!(show_ref_refname(b"").unwrap(), None);
+        // No space, short sha, non-hex sha, empty refname: all errors.
+        let bads: Vec<Vec<u8>> = vec![
+            b"nospaceshere".to_vec(),
+            format!("{}refs/x", "a".repeat(40)).into_bytes(),
+            [b'z'; 40]
+                .iter()
+                .chain(b" refs/x".iter())
+                .copied()
+                .collect(),
+            format!("{} ", "a".repeat(40)).into_bytes(),
+            format!("{} refs/x", "a".repeat(41)).into_bytes(),
+        ];
+        for bad in &bads {
+            assert!(
+                show_ref_refname(bad).is_err(),
+                "malformed show-ref line must error, not skip"
+            );
+        }
+    }
+
+    /// Pruning a path that is not a repo errors. Two shapes: a missing path
+    /// fails `rev-parse` outright (exit 128), while a plain subdirectory of
+    /// a repo would otherwise enumerate the ancestor's refs — covered by the
+    /// identity test below. Either way the prune refuses instead of pruning
+    /// the wrong repo or silently skipping an evil ref.
+    #[test]
+    fn prune_on_non_repo_path_errors() {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .expect("test runtime");
+        rt.block_on(async {
+            let err =
+                prune_non_exempt_gitlawb_refs(Path::new("/tmp/gitlawb-definitely-not-a-repo-xyz"))
+                    .await
+                    .expect_err("non-repo path must fail the prune, not skip it");
+            assert!(
+                err.to_string().contains("not a git repo")
+                    || err.to_string().contains("does not resolve"),
+                "unexpected error shape: {err:#}"
+            );
+        });
+    }
+
+    /// Identity check: pruning a plain subdirectory of a repo refuses instead
+    /// of enumerating the ancestor. `git -C` on the child exits 0 from
+    /// `show-ref` with the ANCESTOR's refs, so without the
+    /// `rev-parse --absolute-git-dir` identity check the delete loop would
+    /// enumerate and delete the wrong repository's refs. Deleting the check
+    /// makes the prune succeed and removes the ancestor's evil ref — both
+    /// assertions go red.
+    #[test]
+    fn prune_on_non_repo_child_of_repo_refuses_and_touches_nothing() {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .expect("test runtime");
+        rt.block_on(async {
+            let td = TempDir::new().unwrap();
+            let ancestor = td.path().join("ancestor.git");
+            std::fs::create_dir_all(&ancestor).unwrap();
+            let run_ancestor = |args: &[&str]| {
+                assert!(
+                    Command::new("git")
+                        .args(args)
+                        .current_dir(&ancestor)
+                        .status()
+                        .unwrap()
+                        .success(),
+                    "git {args:?} failed"
+                );
+            };
+            run_ancestor(&["init", "-q", "--bare", "."]);
+            // Blob target for the planted ref (mirrors the node-metadata shape).
+            std::fs::write(ancestor.join("blob-body"), b"marker\n").unwrap();
+            let out = Command::new("git")
+                .args(["hash-object", "-w", "blob-body"])
+                .current_dir(&ancestor)
+                .output()
+                .unwrap();
+            assert!(out.status.success(), "git hash-object failed");
+            let blob = String::from_utf8_lossy(&out.stdout).trim().to_string();
+            run_ancestor(&["update-ref", "refs/gitlawb/evil", &blob]);
+
+            // A plain child directory: not itself a repo.
+            let child = ancestor.join("not-a-repo");
+            std::fs::create_dir_all(&child).unwrap();
+
+            let err = prune_non_exempt_gitlawb_refs(&child)
+                .await
+                .expect_err("pruning a non-repo child must refuse, not enumerate the ancestor");
+            assert!(
+                err.to_string().contains("not itself a repo"),
+                "refusal must name the identity mismatch, got: {err:#}"
+            );
+
+            // The ancestor's evil ref is untouched.
+            let refs_out = Command::new("git")
+                .args(["for-each-ref", "--format=%(refname)"])
+                .current_dir(&ancestor)
+                .output()
+                .unwrap();
+            let refs = String::from_utf8_lossy(&refs_out.stdout);
+            assert!(
+                refs.lines().any(|l| l.trim() == "refs/gitlawb/evil"),
+                "ancestor evil ref must survive the refused prune"
+            );
+        });
+    }
+
+    /// sha256 mirror: a 64-hex `show-ref` line parses and the planted ref is
+    /// pruned. A 40-only parser skips every line and leaves the evil ref.
+    #[test]
+    fn prune_parses_sha256_show_ref_and_prunes() {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .expect("test runtime");
+        rt.block_on(async {
+            let td = TempDir::new().unwrap();
+            let bare = td.path().join("m256.git");
+            std::fs::create_dir_all(&bare).unwrap();
+            let run = |args: &[&str]| {
+                assert!(
+                    Command::new("git")
+                        .args(args)
+                        .current_dir(&bare)
+                        .status()
+                        .unwrap()
+                        .success(),
+                    "git {args:?} failed"
+                );
+            };
+            let run_out = |args: &[&str]| {
+                let out = Command::new("git")
+                    .args(args)
+                    .current_dir(&bare)
+                    .output()
+                    .unwrap();
+                assert!(out.status.success(), "git {args:?} failed");
+                out.stdout
+            };
+            run(&["init", "-q", "--bare", "--object-format=sha256", "."]);
+            std::fs::write(bare.join("blob-body"), b"marker\n").unwrap();
+            let blob = String::from_utf8_lossy(&run_out(&["hash-object", "-w", "blob-body"]))
+                .trim()
+                .to_string();
+            assert_eq!(
+                blob.len(),
+                64,
+                "sanity: sha256 repo must mint 64-hex objects"
+            );
+            run(&["update-ref", "refs/gitlawb/evil", &blob]);
+            run(&["update-ref", "refs/gitlawb/issues/i1", &blob]);
+
+            prune_non_exempt_gitlawb_refs(&bare)
+                .await
+                .expect("prune succeeds on a sha256 repo");
+
+            let raw_refs = run_out(&["for-each-ref", "--format=%(refname)"]);
+            let refs = String::from_utf8_lossy(&raw_refs);
+            let set: std::collections::HashSet<&str> = refs
+                .lines()
+                .map(str::trim)
+                .filter(|l| !l.is_empty())
+                .collect();
+            assert!(
+                !set.contains("refs/gitlawb/evil"),
+                "evil ref must be pruned from sha256 mirror: {set:?}"
+            );
+            assert!(
+                set.contains("refs/gitlawb/issues/i1"),
+                "exempt ref must survive: {set:?}"
+            );
+        });
+    }
+
+    /// Prune failures defer the queue row (stays `pending` for the next
+    /// tick); deterministic prune refusals and any other sync error fail it
+    /// terminally. Deleting the `downcast_ref::<PruneFailed>` arm restores
+    /// terminal failure and turns the first assertion red.
+    /// Prune failures defer the queue row (stays `pending` for the next
+    /// tick); deterministic prune refusals and any other sync error fail it
+    /// terminally. Deleting the `downcast_ref::<PruneFailed>` arm restores
+    /// terminal failure and turns the first assertion red.
+    ///
+    /// The wrapper below is what the clone/fetch call sites run through:
+    /// a `PruneInvalid` from the helper must still downcast as
+    /// `PruneInvalid` afterwards (not stringified into `PruneFailed`),
+    /// and anything else must arrive as `PruneFailed`.
+    #[test]
+    fn wrap_prune_error_preserves_deterministic_type() {
+        let invalid = wrap_prune_error(anyhow::anyhow!(PruneInvalid("x".to_string())));
+        assert!(
+            invalid.downcast_ref::<PruneInvalid>().is_some(),
+            "deterministic refusal must survive the call-site wrap"
+        );
+        assert!(
+            invalid.downcast_ref::<PruneFailed>().is_none(),
+            "deterministic refusal must not also read as transient"
+        );
+        let transient = wrap_prune_error(anyhow::anyhow!("boom"));
+        assert!(
+            transient.downcast_ref::<PruneFailed>().is_some(),
+            "generic failure must arrive as transient"
+        );
+        assert!(
+            transient.downcast_ref::<PruneInvalid>().is_none(),
+            "generic failure must not read as deterministic"
+        );
+    }
+
+    #[sqlx::test]
+    async fn prune_failure_defers_row_other_failures_fail_it(pool: PgPool) {
+        let state = crate::test_support::test_state(pool).await;
+        let item = |id: &str| crate::db::SyncQueueItem {
+            id: id.to_string(),
+            repo: "z6Mkfoo/hello".to_string(),
+            node_did: "did:key:z6MkOrigin".to_string(),
+            ref_name: "refs/heads/main".to_string(),
+            new_sha: "0".repeat(40),
+            cid: None,
+            status: "pending".to_string(),
+            enqueued_at: chrono::Utc::now().to_rfc3339(),
+        };
+        async fn status_of(pool: &sqlx::PgPool, id: &str) -> String {
+            let row: (String,) = sqlx::query_as("SELECT status FROM sync_queue WHERE id = $1")
+                .bind(id)
+                .fetch_one(pool)
+                .await
+                .unwrap();
+            row.0
+        }
+        // Seed three pending rows directly (enqueue_sync mints random ids).
+        for id in ["sync-prune-defer", "sync-prune-invalid", "sync-other-fail"] {
+            sqlx::query(
+                "INSERT INTO sync_queue (id, repo, node_did, ref_name, new_sha, cid, status, enqueued_at)
+                 VALUES ($1, $2, $3, $4, $5, NULL, 'pending', $6)",
+            )
+            .bind(id)
+            .bind("z6Mkfoo/hello")
+            .bind("did:key:z6MkOrigin")
+            .bind("refs/heads/main")
+            .bind("0".repeat(40))
+            .bind(chrono::Utc::now().to_rfc3339())
+            .execute(state.db.pool())
+            .await
+            .unwrap();
+        }
+
+        handle_sync_item_error(
+            &state.db,
+            &item("sync-prune-defer"),
+            "http://origin.example",
+            anyhow::anyhow!(PruneFailed("boom".to_string())),
+        )
+        .await;
+        assert_eq!(
+            status_of(state.db.pool(), "sync-prune-defer").await,
+            "pending",
+            "prune failure must leave the row pending for retry"
+        );
+
+        handle_sync_item_error(
+            &state.db,
+            &item("sync-other-fail"),
+            "http://origin.example",
+            anyhow::anyhow!("boom"),
+        )
+        .await;
+        assert_eq!(
+            status_of(state.db.pool(), "sync-other-fail").await,
+            "failed",
+            "non-prune failure must fail the row"
+        );
+
+        // Deterministic refusal: no retry can clear it, so it fails
+        // terminally instead of spinning every tick forever.
+        handle_sync_item_error(
+            &state.db,
+            &item("sync-prune-invalid"),
+            "http://origin.example",
+            anyhow::anyhow!(PruneInvalid("not itself a repo".to_string())),
+        )
+        .await;
+        assert_eq!(
+            status_of(state.db.pool(), "sync-prune-invalid").await,
+            "failed",
+            "deterministic prune refusal must fail the row, not defer forever"
+        );
+    }
 }
@@ -0,0 +1,75 @@
+//! Native Git fixtures shared by timing tests on Unix and Windows.
+
+use std::ops::Deref;
+use std::path::{Path, PathBuf};
+use std::sync::OnceLock;
+
+/// The fixture delays or hangs in its own process, so the production watchdog
+/// can terminate it without depending on a shell or descendant process support.
+pub(super) enum Behavior<'a> {
+    Delay(u64),
+    Hang,
+    HangOid(&'a str),
+    HangRepo(&'a str),
+    LogTypes(&'a Path),
+}
+
+pub(super) struct GitShim {
+    _directory: tempfile::TempDir,
+    executable: PathBuf,
+}
+
+impl Deref for GitShim {
+    type Target = Path;
+
+    fn deref(&self) -> &Path {
+        &self.executable
+    }
+}
+
+pub(super) fn create(name: &str, behavior: Behavior<'_>) -> GitShim {
+    static COMPILED: OnceLock<tempfile::TempDir> = OnceLock::new();
+    let compiled = COMPILED.get_or_init(|| {
+        let directory = tempfile::tempdir().expect("native Git fixture directory");
+        let source = directory.path().join("git_shim.rs");
+        std::fs::write(&source, include_str!("../tests/fixtures/git_shim.rs"))
+            .expect("write native Git fixture source");
+        let executable = directory
+            .path()
+            .join(format!("git-shim{}", std::env::consts::EXE_SUFFIX));
+        let output =
+            std::process::Command::new(std::env::var_os("RUSTC").unwrap_or_else(|| "rustc".into()))
+                .arg("--edition=2021")
+                .arg("-Dwarnings")
+                .arg(&source)
+                .arg("-o")
+                .arg(executable)
+                .output()
+                .expect("compile native Git fixture with the installed Rust toolchain");
+        assert!(
+            output.status.success(),
+            "{}",
+            String::from_utf8_lossy(&output.stderr)
+        );
+        directory
+    });
+    let directory = tempfile::Builder::new()
+        .prefix(name)
+        .tempdir()
+        .expect("Git fixture instance");
+    let file_name = format!("git-shim{}", std::env::consts::EXE_SUFFIX);
+    let executable = directory.path().join(&file_name);
+    std::fs::copy(compiled.path().join(file_name), &executable).expect("copy native Git fixture");
+    let config = match behavior {
+        Behavior::Delay(milliseconds) => format!("delay\n{milliseconds}"),
+        Behavior::Hang => "hang\n".to_owned(),
+        Behavior::HangOid(oid) => format!("hang-oid\n{oid}"),
+        Behavior::HangRepo(repo) => format!("hang-repo\n{repo}"),
+        Behavior::LogTypes(path) => format!("log-types\n{}", path.display()),
+    };
+    std::fs::write(directory.path().join("config"), config).expect("configure native Git fixture");
+    GitShim {
+        _directory: directory,
+        executable,
+    }
+}
@@ -0,0 +1,968 @@
+//! Scheduled external-forge fetches for canonical INBOUND mirrors.
+//!
+//! This worker is deliberately separate from `sync`, which mirrors repositories
+//! between Gitlawb peers. External upstream URLs are owner-controlled data that
+//! eventually reach `git fetch`, so each cycle resolves DNS itself, validates
+//! every address, and pins the approved answers into Git/libcurl with
+//! `http.curloptResolve`. Redirects, proxies, credential helpers, submodule
+//! recursion, and non-HTTPS protocols are disabled at the command boundary.
+
+use anyhow::{Context, Result};
+use std::collections::BTreeSet;
+use std::net::IpAddr;
+use std::path::Path;
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+use tracing::{info, warn};
+
+use crate::config::Config;
+use crate::db::{Db, InboundMirrorTarget, MirrorStatus, MirrorTransitionPhase};
+use crate::git::repo_store::RepoStore;
+use crate::state::{repo_identity_key, RepoWriteLeases};
+
+const DNS_TIMEOUT: Duration = Duration::from_secs(10);
+const MAX_PINNED_ADDRESSES: usize = 32;
+const MIN_CURL_OPT_RESOLVE_GIT: (u64, u64) = (2, 37);
+
+#[derive(Debug, Clone, Copy)]
+struct EgressPolicy;
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct PinnedUpstream {
+    url: String,
+    curlopt_resolve: Option<String>,
+}
+
+impl EgressPolicy {
+    async fn resolve(&self, url: reqwest::Url) -> Result<PinnedUpstream> {
+        let raw_host = url.host_str().context("mirror upstream URL has no host")?;
+        let connect_host = raw_host
+            .trim_start_matches('[')
+            .trim_end_matches(']')
+            .to_ascii_lowercase();
+        let policy_host = normalize_policy_host(&connect_host);
+        reject_always_local_hostname(&policy_host)?;
+        let port = url
+            .port_or_known_default()
+            .context("mirror upstream URL has no known port")?;
+
+        if let Ok(ip) = connect_host.parse::<IpAddr>() {
+            self.validate_addresses(&policy_host, &[ip])?;
+            return Ok(PinnedUpstream {
+                url: url.to_string(),
+                // An IP literal performs no DNS lookup, so there is no rebinding
+                // window to pin. TLS still verifies the literal from the URL.
+                curlopt_resolve: None,
+            });
+        }
+
+        let resolved = tokio::time::timeout(
+            DNS_TIMEOUT,
+            tokio::net::lookup_host((connect_host.as_str(), port)),
+        )
+        .await
+        .context("mirror upstream DNS lookup timed out")?
+        .context("resolving mirror upstream host")?;
+        let addresses: BTreeSet<IpAddr> = resolved.map(|addr| addr.ip()).collect();
+        if addresses.is_empty() {
+            anyhow::bail!("mirror upstream host resolved to no addresses");
+        }
+        let addresses: Vec<IpAddr> = addresses.into_iter().collect();
+        self.validate_addresses(&policy_host, &addresses)?;
+
+        let pinned = addresses
+            .iter()
+            .map(|ip| match ip {
+                IpAddr::V4(v4) => v4.to_string(),
+                IpAddr::V6(v6) => format!("[{v6}]"),
+            })
+            .collect::<Vec<_>>()
+            .join(",");
+        Ok(PinnedUpstream {
+            url: url.to_string(),
+            curlopt_resolve: Some(format!("+{connect_host}:{port}:{pinned}")),
+        })
+    }
+
+    fn validate_addresses(&self, host: &str, addresses: &[IpAddr]) -> Result<()> {
+        if addresses.is_empty() || addresses.len() > MAX_PINNED_ADDRESSES {
+            anyhow::bail!(
+                "mirror upstream host must resolve to 1..={MAX_PINNED_ADDRESSES} addresses"
+            );
+        }
+        for &ip in addresses {
+            if !address_is_permitted(ip) {
+                anyhow::bail!("mirror upstream host {host:?} resolved to disallowed address {ip}");
+            }
+        }
+        Ok(())
+    }
+}
+
+fn normalize_policy_host(host: &str) -> String {
+    host.trim_end_matches('.').to_ascii_lowercase()
+}
+
+fn reject_always_local_hostname(host: &str) -> Result<()> {
+    if host.is_empty()
+        || host == "localhost"
+        || host.ends_with(".localhost")
+        || host.ends_with(".local")
+        || host.ends_with(".internal")
+    {
+        anyhow::bail!("mirror upstream host is local-only and cannot be fetched");
+    }
+    Ok(())
+}
+
+fn address_is_permitted(ip: IpAddr) -> bool {
+    crate::api::peers::mirror_upstream_ip_permitted(ip)
+}
+
+fn parse_git_version(output: &str) -> Option<(u64, u64)> {
+    let version = output.split_whitespace().find(|part| {
+        part.as_bytes()
+            .first()
+            .is_some_and(|byte| byte.is_ascii_digit())
+    })?;
+    let mut parts = version.split('.');
+    Some((parts.next()?.parse().ok()?, parts.next()?.parse().ok()?))
+}
+
+fn ensure_safe_git_runtime(git_bin: &str) -> Result<()> {
+    // Bounded runner like every other git spawn on this path: --version does
+    // no I/O of its own, but the child still takes a permit and the deadline.
+    let (status, stdout, stderr) = crate::git::visibility_pack::run_bounded_git_raw_isolated_https(
+        git_bin,
+        &["--version"],
+        &std::env::temp_dir(),
+        b"",
+        Instant::now() + Duration::from_secs(10),
+    )?;
+    if !status.success() {
+        anyhow::bail!(
+            "{git_bin} --version failed: {}",
+            String::from_utf8_lossy(&stderr)
+                .chars()
+                .take(256)
+                .collect::<String>()
+        );
+    }
+    let stdout = String::from_utf8_lossy(&stdout);
+    let version = parse_git_version(&stdout).context("parsing git version")?;
+    if version < MIN_CURL_OPT_RESOLVE_GIT {
+        anyhow::bail!(
+            "upstream mirroring requires Git 2.37+ for DNS pinning; found {}.{}",
+            version.0,
+            version.1
+        );
+    }
+    Ok(())
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct FetchCommand {
+    config: Vec<(String, String)>,
+    args: Vec<String>,
+}
+
+fn fetch_command(upstream: &PinnedUpstream) -> FetchCommand {
+    let http_scope = |name: &str| format!("http.{}.{name}", upstream.url);
+    let credential_scope = |name: &str| format!("credential.{}.{name}", upstream.url);
+    let mut config = vec![
+        ("http.proxy".to_string(), "".to_string()),
+        ("http.followRedirects".to_string(), "false".to_string()),
+        // Reset any inherited multi-valued resolver entries before adding the
+        // address set validated for this exact fetch.
+        ("http.curloptResolve".to_string(), "".to_string()),
+        // Never forward operator or repository-scoped HTTP credentials to an
+        // owner-selected forge. Empty values reset Git's multi-valued headers
+        // and disable cookie state for this invocation.
+        ("http.extraHeader".to_string(), "".to_string()),
+        ("http.cookieFile".to_string(), "".to_string()),
+        ("http.saveCookies".to_string(), "false".to_string()),
+        ("http.sslVerify".to_string(), "true".to_string()),
+        // URL-specific configuration outranks generic `http.*` values even
+        // when the generic value is command-scoped. Repeat every security
+        // control at the exact validated URL so a repository-local scoped
+        // setting cannot re-enable redirects/proxies/headers/cookies or disable
+        // TLS. Key/value environment entries avoid `git -c` splitting a valid
+        // URL path that contains `=`.
+        (http_scope("proxy"), "".to_string()),
+        (http_scope("followRedirects"), "false".to_string()),
+        (http_scope("curloptResolve"), "".to_string()),
+        (http_scope("extraHeader"), "".to_string()),
+        (http_scope("cookieFile"), "".to_string()),
+        (http_scope("saveCookies"), "false".to_string()),
+        (http_scope("sslVerify"), "true".to_string()),
+    ];
+    if let Some(resolve) = &upstream.curlopt_resolve {
+        config.extend([
+            ("http.curloptResolve".to_string(), resolve.clone()),
+            (http_scope("curloptResolve"), resolve.clone()),
+        ]);
+    }
+    config.extend([
+        ("credential.helper".to_string(), "".to_string()),
+        (credential_scope("helper"), "".to_string()),
+        ("credential.interactive".to_string(), "false".to_string()),
+        ("core.askPass".to_string(), "false".to_string()),
+        ("fetch.recurseSubmodules".to_string(), "false".to_string()),
+        // Do not let a smart server or repository-local config introduce a
+        // second, unvalidated pack/bundle URL outside the pinned upstream.
+        ("fetch.uriprotocols".to_string(), "".to_string()),
+        ("fetch.bundleURI".to_string(), "".to_string()),
+        ("protocol.allow".to_string(), "never".to_string()),
+        ("protocol.https.allow".to_string(), "always".to_string()),
+    ]);
+    let args = vec![
+        "fetch".to_string(),
+        "--atomic".to_string(),
+        "--force".to_string(),
+        "--prune".to_string(),
+        "--prune-tags".to_string(),
+        "--no-auto-maintenance".to_string(),
+        "--no-recurse-submodules".to_string(),
+        "--no-write-fetch-head".to_string(),
+        upstream.url.clone(),
+        "+refs/heads/*:refs/heads/*".to_string(),
+        "+refs/tags/*:refs/tags/*".to_string(),
+    ];
+    FetchCommand { config, args }
+}
+
+fn reject_local_url_rewrite(
+    git_bin: &str,
+    repo_path: &Path,
+    upstream_url: &str,
+    deadline: Instant,
+) -> Result<()> {
+    let args = ["ls-remote", "--get-url", upstream_url];
+    let (status, stdout, stderr) = crate::git::visibility_pack::run_bounded_git_raw_isolated_https(
+        git_bin, &args, repo_path, b"", deadline,
+    )?;
+    if !status.success() {
+        let stderr = String::from_utf8_lossy(&stderr);
+        anyhow::bail!(
+            "checking mirror upstream URL rewrite failed: {}",
+            stderr.chars().take(4_096).collect::<String>()
+        );
+    }
+    let expanded = std::str::from_utf8(&stdout)
+        .context("Git returned a non-UTF-8 expanded mirror upstream URL")?
+        .trim_end_matches(['\r', '\n']);
+    if expanded != upstream_url {
+        anyhow::bail!("repository-local Git config attempted to rewrite the mirror upstream URL");
+    }
+    Ok(())
+}
+
+fn reject_sensitive_local_config(git_bin: &str, repo_path: &Path, deadline: Instant) -> Result<()> {
+    let args = [
+        "config",
+        "--local",
+        "--includes",
+        "--get-regexp",
+        r"^(http|credential|url)\.",
+    ];
+    let (status, stdout, stderr) = crate::git::visibility_pack::run_bounded_git_raw_isolated_https(
+        git_bin, &args, repo_path, b"", deadline,
+    )?;
+    match status.code() {
+        Some(1) if stdout.is_empty() => Ok(()),
+        Some(0) => anyhow::bail!(
+            "repository-local network or credential Git config is not allowed for upstream mirrors"
+        ),
+        _ => {
+            let stderr = String::from_utf8_lossy(&stderr);
+            anyhow::bail!(
+                "checking repository-local Git config failed: {}",
+                stderr.chars().take(4_096).collect::<String>()
+            )
+        }
+    }
+}
+
+/// Budget for each preflight probe inside [`run_fetch`] (the URL-rewrite
+/// check and the sensitive-local-config scan). Both run local git children
+/// against the on-disk repo, so a short bound is generous; keeping them on
+/// their own deadlines means `upstream_mirror_fetch_timeout_secs` covers
+/// only the external pack transfer it documents.
+const PREFLIGHT_PROBE_TIMEOUT: Duration = Duration::from_secs(30);
+
+fn run_fetch(
+    git_bin: &str,
+    repo_path: &Path,
+    upstream: &PinnedUpstream,
+    timeout: Duration,
+) -> Result<()> {
+    let command = fetch_command(upstream);
+    let args: Vec<&str> = command.args.iter().map(String::as_str).collect();
+    let config: Vec<(&str, &str)> = command
+        .config
+        .iter()
+        .map(|(key, value)| (key.as_str(), value.as_str()))
+        .collect();
+    let rewrite_deadline = Instant::now()
+        .checked_add(PREFLIGHT_PROBE_TIMEOUT)
+        .context("mirror URL-rewrite probe timeout is not representable")?;
+    reject_local_url_rewrite(git_bin, repo_path, &upstream.url, rewrite_deadline)?;
+    let config_deadline = Instant::now()
+        .checked_add(PREFLIGHT_PROBE_TIMEOUT)
+        .context("mirror local-config probe timeout is not representable")?;
+    reject_sensitive_local_config(git_bin, repo_path, config_deadline)?;
+    let deadline = Instant::now()
+        .checked_add(timeout)
+        .context("upstream mirror fetch timeout is not representable")?;
+    let (status, _stdout, stderr) =
+        crate::git::visibility_pack::run_bounded_git_raw_isolated_https_with_config(
+            git_bin, &args, repo_path, b"", deadline, &config,
+        )?;
+    if !status.success() {
+        let stderr = String::from_utf8_lossy(&stderr);
+        let stderr = stderr.chars().take(4_096).collect::<String>();
+        anyhow::bail!("git upstream fetch failed: {stderr}");
+    }
+    Ok(())
+}
+
+#[derive(Clone)]
+struct Worker {
+    db: Arc<Db>,
+    config: Arc<Config>,
+    repo_store: RepoStore,
+    repo_write_leases: RepoWriteLeases,
+    git_write_semaphore: Arc<tokio::sync::Semaphore>,
+    policy: EgressPolicy,
+    git_bin: String,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum FetchOutcome {
+    Updated,
+    SkippedStale,
+}
+
+impl Worker {
+    async fn fetch_target(&self, target: InboundMirrorTarget) -> Result<FetchOutcome> {
+        let url = crate::db::validate_mirror_upstream_url(&target.upstream_url)?;
+        let upstream = self.policy.resolve(url).await?;
+
+        let repo_identity = repo_identity_key(&target.owner_did, &target.name);
+        let steal_after = Duration::from_secs(
+            self.config
+                .upstream_mirror_fetch_timeout_secs
+                .saturating_mul(2)
+                .saturating_add(60),
+        );
+        let _lease = self
+            .repo_write_leases
+            .acquire(&repo_identity, steal_after)
+            .await
+            .context("upstream mirror write-lease waiter cap reached")?;
+
+        let _write_permit = Arc::clone(&self.git_write_semaphore)
+            .acquire_owned()
+            .await
+            .context("upstream mirror write semaphore closed")?;
+        let acquire_timeout = Duration::from_secs(self.config.git_acquire_timeout_secs);
+        let guard = tokio::time::timeout(
+            acquire_timeout,
+            self.repo_store
+                .acquire_write(&target.owner_did, &target.name),
+        )
+        .await
+        .context("upstream mirror write-lock acquisition timed out")??;
+        let repo_path = guard.path().to_path_buf();
+
+        // A row can leave INBOUND state after selection. Recheck only after
+        // both the process-local lease and the cluster-wide advisory write lock
+        // are held. Every future transition path must take these same locks
+        // before changing authority, so no other node can commit a transition
+        // between this check and the fetch.
+        let current = match self.db.get_repo_mirror_state(&target.repo_id).await {
+            Ok(Some(current)) => current,
+            Ok(None) => {
+                let _ = guard.release(false).await;
+                anyhow::bail!("inbound mirror state disappeared before fetch");
+            }
+            Err(error) => {
+                let _ = guard.release(false).await;
+                return Err(error);
+            }
+        };
+        // Both stable shapes the CHECK admits: never transitioned, or a
+        // completed transition whose metadata is retained as audit. Only an
+        // in-flight transition makes the target stale.
+        let stable = (current.transition_id.is_none() && current.transition_phase.is_none())
+            || (current.transition_id.is_some()
+                && current.transition_phase == Some(MirrorTransitionPhase::Completed));
+        if current.status != MirrorStatus::Inbound
+            || !stable
+            || current.upstream_url != target.upstream_url
+        {
+            let _ = guard.release(false).await;
+            info!(repo_id = %target.repo_id, "upstream mirror state changed before fetch; skipping stale target");
+            return Ok(FetchOutcome::SkippedStale);
+        }
+
+        let git_bin = self.git_bin.clone();
+        let fetch_timeout = Duration::from_secs(self.config.upstream_mirror_fetch_timeout_secs);
+        let fetch = tokio::task::spawn_blocking(move || {
+            run_fetch(&git_bin, &repo_path, &upstream, fetch_timeout)
+        })
+        .await
+        .context("upstream mirror fetch task panicked")?;
+        let success = fetch.is_ok();
+        let outcome = guard.release(success).await;
+        // A refused publish (fenced or unknowable upload) must not report
+        // Updated; surface it as the failure it is before mapping the fetch.
+        outcome.into_result()?;
+        fetch.map(|()| FetchOutcome::Updated)
+    }
+
+    async fn scan_once(&self) {
+        let mut cursor: Option<String> = None;
+        loop {
+            let targets = match self
+                .db
+                .list_inbound_mirror_targets(
+                    cursor.as_deref(),
+                    self.config.upstream_mirror_page_size as i64,
+                )
+                .await
+            {
+                Ok(targets) => targets,
+                Err(error) => {
+                    warn!(err = %error, "failed to list inbound mirror targets");
+                    return;
+                }
+            };
+            if targets.is_empty() {
+                return;
+            }
+            let page_len = targets.len();
+            cursor = targets.last().map(|target| target.repo_id.clone());
+            for target in targets {
+                let repo_id = target.repo_id.clone();
+                let upstream_url = target.upstream_url.clone();
+                match self.fetch_target(target).await {
+                    Ok(FetchOutcome::Updated) => {
+                        info!(repo_id, upstream = %upstream_url, "upstream mirror fetch completed")
+                    }
+                    Ok(FetchOutcome::SkippedStale) => {}
+                    Err(error) => {
+                        warn!(repo_id, upstream = %upstream_url, err = %error, "upstream mirror fetch failed")
+                    }
+                }
+            }
+            if page_len < self.config.upstream_mirror_page_size {
+                return;
+            }
+        }
+    }
+}
+
+/// Validate the opt-in runtime and spawn the scheduled worker. The worker runs
+/// one scan immediately, sleeps only after a complete scan, and exits between
+/// scans when graceful shutdown is signalled. An in-flight Git child is bounded
+/// and reaped before the worker observes shutdown.
+pub fn start(
+    db: Arc<Db>,
+    config: Arc<Config>,
+    repo_store: RepoStore,
+    repo_write_leases: RepoWriteLeases,
+    git_write_semaphore: Arc<tokio::sync::Semaphore>,
+    mut shutdown_rx: tokio::sync::watch::Receiver<bool>,
+) -> Result<()> {
+    ensure_safe_git_runtime("git")?;
+    let policy = EgressPolicy;
+    let worker = Worker {
+        db,
+        config: Arc::clone(&config),
+        repo_store,
+        repo_write_leases,
+        git_write_semaphore,
+        policy,
+        git_bin: "git".to_string(),
+    };
+    tokio::spawn(async move {
+        info!(
+            interval_secs = config.upstream_mirror_interval_secs,
+            "upstream mirror worker started"
+        );
+        loop {
+            worker.scan_once().await;
+            tokio::select! {
+                _ = tokio::time::sleep(Duration::from_secs(config.upstream_mirror_interval_secs)) => {}
+                changed = shutdown_rx.changed() => {
+                    if changed.is_err() || *shutdown_rx.borrow() {
+                        info!("upstream mirror worker stopped");
+                        return;
+                    }
+                }
+            }
+        }
+    });
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use clap::Parser;
+    use std::path::PathBuf;
+
+    async fn test_worker(
+        pool: sqlx::PgPool,
+        git_bin: PathBuf,
+        page_size: usize,
+    ) -> (Worker, Arc<Db>, tempfile::TempDir) {
+        let db = Arc::new(Db::for_testing(pool.clone()));
+        db.run_migrations().await.unwrap();
+        let repos = tempfile::tempdir().unwrap();
+        let mut config = Config::parse_from([
+            "gitlawb-node".to_string(),
+            "--upstream-mirror-enabled".to_string(),
+            "--enforce-owner-push".to_string(),
+            "--upstream-mirror-fetch-timeout-secs".to_string(),
+            "5".to_string(),
+            "--upstream-mirror-page-size".to_string(),
+            page_size.to_string(),
+        ]);
+        config.validate().unwrap();
+        let config = Arc::new(config);
+        let worker = Worker {
+            db: Arc::clone(&db),
+            config,
+            repo_store: RepoStore::for_testing(repos.path().to_path_buf(), pool),
+            repo_write_leases: RepoWriteLeases::new(8),
+            git_write_semaphore: Arc::new(tokio::sync::Semaphore::new(4)),
+            policy: EgressPolicy,
+            git_bin: git_bin.to_string_lossy().into_owned(),
+        };
+        (worker, db, repos)
+    }
+
+    async fn add_inbound_mirror(
+        worker: &Worker,
+        db: &Db,
+        id: &str,
+        name: &str,
+        upstream_url: &str,
+    ) -> InboundMirrorTarget {
+        let now = chrono::Utc::now();
+        let record = crate::db::RepoRecord {
+            id: id.to_string(),
+            name: name.to_string(),
+            owner_did: "did:key:z6MkMirrorWorker".to_string(),
+            description: None,
+            is_public: true,
+            default_branch: "main".to_string(),
+            created_at: now,
+            updated_at: now,
+            disk_path: format!("/test/{id}.git"),
+            forked_from: None,
+            machine_id: None,
+        };
+        db.create_repo(&record).await.unwrap();
+        worker
+            .repo_store
+            .init(&record.owner_did, &record.name)
+            .await
+            .unwrap();
+        let state = db
+            .configure_inbound_mirror(&record.id, upstream_url)
+            .await
+            .unwrap();
+        InboundMirrorTarget {
+            repo_id: record.id,
+            owner_did: record.owner_did,
+            name: record.name,
+            upstream_url: state.upstream_url,
+        }
+    }
+
+    #[cfg(unix)]
+    fn fake_git(repos: &tempfile::TempDir) -> (PathBuf, PathBuf) {
+        use std::os::unix::fs::PermissionsExt;
+
+        let git = repos.path().join("fake-git");
+        let calls = repos.path().join("fetch-calls");
+        let script = format!(
+            r#"#!/bin/sh
+if [ "$1" = "ls-remote" ]; then
+  printf '%s\n' "$3"
+  exit 0
+fi
+if [ "$1" = "config" ]; then
+  # Match real Git when no sensitive repository-local keys are configured.
+  exit 1
+fi
+case "$*" in
+  *fail.git*) printf '%s\n' fail >> '{}'; exit 7 ;;
+  *) printf '%s\n' ok >> '{}'; exit 0 ;;
+esac
+"#,
+            calls.display(),
+            calls.display()
+        );
+        std::fs::write(&git, script).unwrap();
+        let mut permissions = std::fs::metadata(&git).unwrap().permissions();
+        permissions.set_mode(0o700);
+        std::fs::set_permissions(&git, permissions).unwrap();
+        (git, calls)
+    }
+
+    #[test]
+    fn git_version_gate_requires_curlopt_resolve_support() {
+        assert_eq!(parse_git_version("git version 2.37.0"), Some((2, 37)));
+        assert_eq!(
+            parse_git_version("git version 2.39.5 (Apple Git-154)"),
+            Some((2, 39))
+        );
+        assert_eq!(parse_git_version("not git"), None);
+        assert!((2, 36) < MIN_CURL_OPT_RESOLVE_GIT);
+    }
+
+    #[test]
+    fn every_resolved_address_must_be_public() {
+        let policy = EgressPolicy;
+        let public_v4: IpAddr = "8.8.8.8".parse().unwrap();
+        let private_v4: IpAddr = "10.2.3.4".parse().unwrap();
+        let private_v6: IpAddr = "fd00::20".parse().unwrap();
+        assert!(policy
+            .validate_addresses("public.example", &[public_v4])
+            .is_ok());
+        assert!(policy
+            .validate_addresses("private.example", &[private_v4, private_v6])
+            .is_err());
+        assert!(policy
+            .validate_addresses("mixed.example", &[public_v4, private_v4])
+            .is_err());
+        let too_many = (1..=MAX_PINNED_ADDRESSES + 1)
+            .map(|last| IpAddr::V4(std::net::Ipv4Addr::new(8, 8, 0, last as u8)))
+            .collect::<Vec<_>>();
+        assert!(policy
+            .validate_addresses("oversized.example", &too_many)
+            .is_err());
+
+        for never in [
+            "127.0.0.1",
+            "169.254.169.254",
+            "192.0.2.1",
+            "198.18.0.1",
+            "198.51.100.1",
+            "203.0.113.10",
+            "::1",
+            "::ffff:198.18.0.1",
+            "2002:c612:1::",
+            "64:ff9b::c612:1",
+            "100::1",
+            "100:0:0:1::1",
+            "2001::1",
+            "2001:2::1",
+            "2001:db8::1",
+            "fec0::1",
+            "fe80::1",
+            "3fff::1",
+            "5f00::1",
+        ] {
+            let ip = never.parse().unwrap();
+            assert!(!address_is_permitted(ip), "{never} must stay denied");
+        }
+    }
+
+    #[test]
+    fn local_only_hostnames_are_always_rejected() {
+        for invalid in [
+            "",
+            "localhost",
+            "forge.localhost",
+            "forge.local",
+            "ghe.internal",
+        ] {
+            assert!(
+                reject_always_local_hostname(invalid).is_err(),
+                "{invalid:?}"
+            );
+        }
+        assert!(reject_always_local_hostname("github.example").is_ok());
+    }
+
+    #[test]
+    fn fetch_command_pins_dns_and_only_updates_branches_and_tags() {
+        let upstream = PinnedUpstream {
+            url: "https://github.example/org/repo.git".to_string(),
+            curlopt_resolve: Some("+github.example:443:8.8.8.8,[2606:4700:4700::1111]".to_string()),
+        };
+        let command = fetch_command(&upstream);
+        let has_config = |key: &str, value: &str| {
+            command
+                .config
+                .contains(&(key.to_string(), value.to_string()))
+        };
+        assert!(has_config("http.followRedirects", "false"));
+        assert!(has_config("http.proxy", ""));
+        assert!(has_config("http.extraHeader", ""));
+        assert!(has_config("http.cookieFile", ""));
+        assert!(has_config("http.saveCookies", "false"));
+        assert!(has_config("http.sslVerify", "true"));
+        assert!(has_config(
+            "http.curloptResolve",
+            "+github.example:443:8.8.8.8,[2606:4700:4700::1111]"
+        ));
+        assert!(has_config(
+            "http.https://github.example/org/repo.git.followRedirects",
+            "false"
+        ));
+        assert!(has_config(
+            "http.https://github.example/org/repo.git.proxy",
+            ""
+        ));
+        assert!(has_config(
+            "http.https://github.example/org/repo.git.extraHeader",
+            ""
+        ));
+        assert!(has_config(
+            "http.https://github.example/org/repo.git.curloptResolve",
+            "+github.example:443:8.8.8.8,[2606:4700:4700::1111]"
+        ));
+        assert!(has_config(
+            "credential.https://github.example/org/repo.git.helper",
+            ""
+        ));
+        assert!(has_config("protocol.allow", "never"));
+        assert!(has_config("protocol.https.allow", "always"));
+        assert!(has_config("fetch.uriprotocols", ""));
+        assert!(has_config("fetch.bundleURI", ""));
+        assert!(!command
+            .config
+            .iter()
+            .any(|(key, _)| key.ends_with("sslCert") || key.ends_with("sslKey")));
+        assert!(command.args.contains(&"--atomic".to_string()));
+        assert!(command
+            .args
+            .contains(&"+refs/heads/*:refs/heads/*".to_string()));
+        assert!(command
+            .args
+            .contains(&"+refs/tags/*:refs/tags/*".to_string()));
+        assert!(!command.args.iter().any(|arg| arg == "+refs/*:refs/*"));
+    }
+
+    /// Spawn `git` inside a fixture repo with inherited repository-location
+    /// and config environment scrubbed, so a test run inside an outer Git
+    /// checkout or custom config can never leak into the fixture.
+    fn fixture_git(dir: &std::path::Path) -> std::process::Command {
+        let mut cmd = std::process::Command::new("git");
+        cmd.current_dir(dir);
+        for var in [
+            "GIT_DIR",
+            "GIT_WORK_TREE",
+            "GIT_CONFIG",
+            "GIT_CONFIG_GLOBAL",
+            "GIT_CONFIG_SYSTEM",
+            "GIT_CONFIG_NOSYSTEM",
+            "GIT_CONFIG_COUNT",
+            "GIT_EXEC_PATH",
+            "GIT_SSH",
+            "GIT_SSH_COMMAND",
+            "GIT_ASKPASS",
+            "SSH_ASKPASS",
+        ] {
+            cmd.env_remove(var);
+        }
+        cmd
+    }
+
+    #[test]
+    fn repository_local_instead_of_rewrite_is_rejected() {
+        let repo = tempfile::tempdir().unwrap();
+        let init = fixture_git(repo.path())
+            .args(["init", "--bare"])
+            .output()
+            .unwrap();
+        assert!(init.status.success());
+        let config = fixture_git(repo.path())
+            .args([
+                "config",
+                "url.https://evil.example/.insteadOf",
+                "https://github.example/",
+            ])
+            .output()
+            .unwrap();
+        assert!(config.status.success());
+
+        let error = reject_local_url_rewrite(
+            "git",
+            repo.path(),
+            "https://github.example/org/repo.git",
+            Instant::now() + Duration::from_secs(5),
+        )
+        .unwrap_err();
+        assert!(error.to_string().contains("attempted to rewrite"));
+    }
+
+    #[test]
+    fn repository_local_network_and_credential_config_is_rejected() {
+        let repo = tempfile::tempdir().unwrap();
+        let init = fixture_git(repo.path())
+            .args(["init", "--bare"])
+            .output()
+            .unwrap();
+        assert!(init.status.success());
+
+        reject_sensitive_local_config("git", repo.path(), Instant::now() + Duration::from_secs(5))
+            .unwrap();
+
+        let config = fixture_git(repo.path())
+            .args(["config", "http.sslCert", "/tmp/untrusted-client.pem"])
+            .output()
+            .unwrap();
+        assert!(config.status.success());
+
+        let error = reject_sensitive_local_config(
+            "git",
+            repo.path(),
+            Instant::now() + Duration::from_secs(5),
+        )
+        .unwrap_err();
+        assert!(error.to_string().contains("network or credential"));
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn hardened_real_git_reaches_https_transport_for_equals_path() {
+        let repo = tempfile::tempdir().unwrap();
+        let init = fixture_git(repo.path())
+            .args(["init", "--bare"])
+            .output()
+            .unwrap();
+        assert!(init.status.success());
+
+        // This is intentionally a loopback test seam around `run_fetch`, below
+        // the egress-policy layer. Reaching the listener proves the generated
+        // Git config neither dies on an empty client certificate nor splits the
+        // valid `=` path at the command-config boundary. The listener closes
+        // immediately, so the expected final result is a TLS transport error.
+        let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap();
+        let port = listener.local_addr().unwrap().port();
+        listener.set_nonblocking(true).unwrap();
+        let accepted = std::thread::spawn(move || {
+            let deadline = Instant::now() + Duration::from_secs(3);
+            while Instant::now() < deadline {
+                match listener.accept() {
+                    Ok((_stream, _address)) => return true,
+                    Err(error) if error.kind() == std::io::ErrorKind::WouldBlock => {
+                        std::thread::sleep(Duration::from_millis(10));
+                    }
+                    Err(error) => panic!("accepting real-Git transport probe: {error}"),
+                }
+            }
+            false
+        });
+        let upstream = PinnedUpstream {
+            url: format!("https://127.0.0.1:{port}/org/repo=mirror.git"),
+            curlopt_resolve: None,
+        };
+        let error = run_fetch("git", repo.path(), &upstream, Duration::from_secs(5)).unwrap_err();
+        assert!(
+            accepted.join().unwrap(),
+            "real Git must reach the HTTPS transport; got {error:#}"
+        );
+        let error = error.to_string();
+        assert!(!error.contains("credential missing host field"));
+        assert!(!error.contains("invalid key"));
+    }
+
+    #[tokio::test]
+    async fn literal_addresses_are_checked_without_dns() {
+        let policy = EgressPolicy;
+        let allowed = policy
+            .resolve(reqwest::Url::parse("https://8.8.8.8/org/repo.git").unwrap())
+            .await
+            .unwrap();
+        assert_eq!(allowed.curlopt_resolve, None);
+
+        for denied in [
+            "127.0.0.1",
+            "10.2.3.4",
+            "169.254.169.254",
+            "192.0.2.1",
+            "198.18.0.1",
+            "203.0.113.10",
+            "[::1]",
+            "[100::1]",
+            "[100:0:0:1::1]",
+            "[2001::1]",
+            "[2001:db8::1]",
+            "[fec0::1]",
+            "[fd00::20]",
+            "[3fff::1]",
+            "[5f00::1]",
+        ] {
+            let url = reqwest::Url::parse(&format!("https://{denied}/org/repo.git")).unwrap();
+            assert!(
+                policy.resolve(url).await.is_err(),
+                "{denied} must stay denied"
+            );
+        }
+    }
+
+    #[sqlx::test]
+    async fn stale_target_is_rechecked_under_the_write_locks(pool: sqlx::PgPool) {
+        let missing_git = PathBuf::from("/definitely/missing/git");
+        let (worker, db, _repos) = test_worker(pool, missing_git, 1).await;
+        let target = add_inbound_mirror(
+            &worker,
+            &db,
+            "00000000-0000-0000-0000-000000000001",
+            "stale",
+            "https://8.8.8.8/org/stale.git",
+        )
+        .await;
+        sqlx::query(
+            "UPDATE repos
+             SET mirror_status = 'outbound'
+             WHERE id = $1",
+        )
+        .bind(&target.repo_id)
+        .execute(db.pool())
+        .await
+        .unwrap();
+
+        let outcome = worker.fetch_target(target).await.unwrap();
+        assert_eq!(outcome, FetchOutcome::SkippedStale);
+        assert!(worker.repo_write_leases.is_empty());
+    }
+
+    #[cfg(unix)]
+    #[sqlx::test]
+    async fn scan_pages_past_one_failure_and_fetches_the_next_repo(pool: sqlx::PgPool) {
+        let bootstrap = tempfile::tempdir().unwrap();
+        let (git, calls) = fake_git(&bootstrap);
+        let (worker, db, _repos) = test_worker(pool, git, 1).await;
+        add_inbound_mirror(
+            &worker,
+            &db,
+            "00000000-0000-0000-0000-000000000001",
+            "fails",
+            "https://8.8.8.8/org/fail.git",
+        )
+        .await;
+        add_inbound_mirror(
+            &worker,
+            &db,
+            "00000000-0000-0000-0000-000000000002",
+            "succeeds",
+            "https://8.8.8.8/org/ok.git",
+        )
+        .await;
+
+        worker.scan_once().await;
+
+        let calls = std::fs::read_to_string(calls).unwrap();
+        assert_eq!(calls.lines().collect::<Vec<_>>(), ["fail", "ok"]);
+        assert!(worker.repo_write_leases.is_empty());
+    }
+}
@@ -437,6 +437,31 @@ mod tests {
        );
    }

+    // #135 T1: a Mode-B rule on `/secret/**` must DENY the withheld directory's
+    // OWN path `/secret` (the `path == prefix` arm), not just strict descendants —
+    // otherwise get_by_cid's tree gate would serve the /secret tree object and leak
+    // its children. Pins parity with get_tree, which denies the /secret path.
+    #[test]
+    fn subtree_rule_denies_the_withheld_directory_itself() {
+        let reader = "did:key:z6MkReader";
+        let rules = [rule("/secret/**", VisibilityMode::B, &[reader])];
+        assert_eq!(
+            visibility_check(&rules, true, OWNER, None, "/secret"),
+            Decision::Deny,
+            "anon denied at the withheld directory's OWN path /secret"
+        );
+        assert_eq!(
+            visibility_check(&rules, true, OWNER, None, "/public"),
+            Decision::Allow,
+            "anon allowed at a sibling path outside the withheld subtree"
+        );
+        assert_eq!(
+            visibility_check(&rules, true, OWNER, Some(reader), "/secret"),
+            Decision::Allow,
+            "listed reader allowed at the withheld directory (caller-aware)"
+        );
+    }
+
    // #153 regression: cross-method DID must still be denied even when the
    // trailing segment collides with a bare owner key.
    #[test]
@@ -38,26 +38,151 @@ pub fn fire_event(
    event: &str,
    payload: serde_json::Value,
 ) {
-    let repo_id = repo_id.to_string();
-    let event = event.to_string();
-    tokio::spawn(async move {
-        fire_event_async(db, http_client, &repo_id, &event, payload).await;
-    });
+    fire_event_occurrence(db, http_client, repo_id, event, payload, None, None);
 }

-async fn fire_event_async(
+/// Occurrence-keyed variant: at most one delivery per
+/// `(request_id, ref_name, hook)` is spawned; concurrent executors
+/// collapse via the `webhook_deliveries` ledger. Best-effort on crash
+/// (ledger is claimed before spawn), which the PR scope allows.
+///
+/// For durable outbox usage, call `claim_webhook_delivery_before_spawn` first
+/// to claim deliveries synchronously, then call
+/// `fire_event_occurrence_with_claimed_hooks` with the claimed hook IDs to
+/// spawn HTTP POST only for claimed hooks.
+pub fn fire_event_occurrence(
    db: Arc<Db>,
    http_client: Arc<reqwest::Client>,
    repo_id: &str,
    event: &str,
    payload: serde_json::Value,
+    request_id: Option<&str>,
+    ref_name: Option<&str>,
 ) {
-    let hooks = match db.list_webhooks_for_event(repo_id, event).await {
-        Ok(h) => h,
-        Err(e) => {
-            tracing::warn!(err = %e, "failed to list webhooks for event {event}");
-            return;
+    let repo_id = repo_id.to_string();
+    let event = event.to_string();
+    let request_id = request_id.map(|s| s.to_string());
+    let ref_name = ref_name.map(|s| s.to_string());
+    tokio::spawn(async move {
+        fire_event_async_occurrence(
+            db,
+            http_client,
+            &repo_id,
+            &event,
+            payload,
+            request_id.as_deref(),
+            ref_name.as_deref(),
+            None,
+        )
+        .await;
+    });
+}
+
+/// Variant for durable outbox: only spawn HTTP POST for previously claimed hooks.
+#[allow(clippy::too_many_arguments)]
+pub fn fire_event_occurrence_with_claimed_hooks(
+    db: Arc<Db>,
+    http_client: Arc<reqwest::Client>,
+    repo_id: &str,
+    event: &str,
+    payload: serde_json::Value,
+    request_id: Option<&str>,
+    ref_name: Option<&str>,
+    claimed_hooks: Vec<crate::db::Webhook>,
+) {
+    let repo_id = repo_id.to_string();
+    let event = event.to_string();
+    let request_id = request_id.map(|s| s.to_string());
+    let ref_name = ref_name.map(|s| s.to_string());
+    tokio::spawn(async move {
+        fire_event_async_occurrence(
+            db,
+            http_client,
+            &repo_id,
+            &event,
+            payload,
+            request_id.as_deref(),
+            ref_name.as_deref(),
+            Some(claimed_hooks),
+        )
+        .await;
+    });
+}
+
+/// Claim webhook delivery before spawning HTTP POST.
+/// Returns `Some(hook ids)` successfully claimed, or `None` when the hook
+/// list itself could not be read (transient DB error, unknown membership).
+/// A `None` caller must fall back to best-effort send rather than dropping
+/// the delivery: children are still deleted downstream, so an empty claim
+/// set on DB error would lose the webhook permanently.
+/// This synchronous claim phase must be awaited before child deletion to prevent
+/// permanent webhook loss on crash between deletion and claim.
+pub async fn claim_webhook_delivery_before_spawn(
+    db: Arc<Db>,
+    repo_id: &str,
+    event: &str,
+    request_id: Option<&str>,
+    ref_name: Option<&str>,
+) -> Option<Vec<crate::db::Webhook>> {
+    let mut claimed_hooks = Vec::new();
+
+    // Only claim when we have occurrence context (request_id + ref_name)
+    if let (Some(req), Some(r)) = (request_id, ref_name) {
+        let hooks = match db.list_webhooks_for_event(repo_id, event).await {
+            Ok(h) => h,
+            Err(e) => {
+                tracing::warn!(err = %e, "failed to list webhooks for event {event}");
+                return None;
+            }
+        };
+
+        for hook in hooks {
+            let delivery_id = crate::db::deterministic_id(&["webhook", req, r, &hook.id]);
+            match db
+                .claim_webhook_delivery(&delivery_id, req, repo_id, event)
+                .await
+            {
+                Ok(true) => claimed_hooks.push(hook),
+                Ok(false) => continue, // Already claimed
+                // Send fallback (legacy best-effort): a transient claim
+                // error must not drop the delivery while children are
+                // still deleted. The in-task claimer this replaced sent
+                // anyway on claim error; keep that here.
+                Err(e) => {
+                    tracing::warn!(err = %e, hook_id = %hook.id, "webhook claim failed; sending anyway");
+                    claimed_hooks.push(hook);
+                }
+            }
        }
+    }
+
+    Some(claimed_hooks)
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn fire_event_async_occurrence(
+    db: Arc<Db>,
+    http_client: Arc<reqwest::Client>,
+    repo_id: &str,
+    event: &str,
+    payload: serde_json::Value,
+    request_id: Option<&str>,
+    ref_name: Option<&str>,
+    claimed_hooks: Option<Vec<crate::db::Webhook>>,
+) {
+    // A caller that already claimed deliveries hands the claimed hook rows
+    // through; the send path must not re-list, because a list failure here
+    // would silently drop deliveries already claimed in the ledger.
+    let already_claimed = claimed_hooks.is_some();
+    let hooks = match claimed_hooks {
+        Some(h) => h,
+        None => match db.list_webhooks_for_event(repo_id, event).await {
+            Ok(h) => h,
+            Err(e) => {
+                tracing::warn!(err = %e, "failed to list webhooks for event {event}");
+                return;
+            }
+        },
    };

    if hooks.is_empty() {
@@ -76,9 +201,44 @@ async fn fire_event_async(
        let client = Arc::clone(&http_client);
        let event_name = event.to_string();
        let bytes = payload_bytes.clone();
-        let delivery_id = uuid::Uuid::new_v4().to_string();
+        // Stable occurrence delivery key when request context exists;
+        // random UUID otherwise (legacy callers without occurrence).
+        let delivery_id = match (request_id, ref_name) {
+            (Some(req), Some(r)) => crate::db::deterministic_id(&["webhook", req, r, &hook.id]),
+            _ => uuid::Uuid::new_v4().to_string(),
+        };
+        // Claim before spawn so concurrent executors collapse to one
+        // delivery per occurrence. Failures to record fall back to
+        // sending (legacy best-effort) rather than dropping.
+        // Skip claim if already done by caller (durable_outbox path).
+        if !already_claimed && request_id.is_some() {
+            match db
+                .clone()
+                .claim_webhook_delivery(&delivery_id, request_id.unwrap_or(""), repo_id, event)
+                .await
+            {
+                Ok(true) => {}
+                Ok(false) => continue,
+                Err(_) => {}
+            }
+        }

+        let db_for_mark = db.clone();
        tokio::spawn(async move {
+            // Policy skip, not a delivery failure: the stored URL predates the
+            // create-time gate. The "invalid_url" label is documented on
+            // record_webhook_delivery in metrics.rs.
+            if !crate::api::peers::is_public_http_url(&hook.url) {
+                crate::metrics::record_webhook_delivery("invalid_url");
+                tracing::warn!(
+                    url = %hook.url,
+                    event = %event_name,
+                    "webhook delivery skipped: stored URL is not a public http(s) target"
+                );
+                return;
+            }
+
+            let delivery_for_mark = delivery_id.clone();
            let signature = hook.secret.as_deref().map(|s| sign_payload(s, &bytes));

            let mut req = client
@@ -100,6 +260,10 @@ async fn fire_event_async(
                        "http_error"
                    };
                    crate::metrics::record_webhook_delivery(result_label);
+                    // Any HTTP response proves the delivery fired; mark
+                    // sent so the ledger never claims an unsent delivery.
+                    // Network errors stay pending for stale-reclaim.
+                    let _ = db_for_mark.mark_webhook_sent(&delivery_for_mark).await;
                    tracing::info!(
                        url = %hook.url,
                        event = %event_name,
@@ -0,0 +1,59 @@
+use std::io::{Read, Write};
+use std::process::{Command, Stdio};
+use std::time::Duration;
+
+fn main() -> std::io::Result<()> {
+    let executable = std::env::current_exe()?;
+    let config = std::fs::read_to_string(executable.parent().unwrap().join("config"))?;
+    let (mode, parameter) = config.split_once('\n').unwrap_or((&config, ""));
+    let args: Vec<_> = std::env::args_os().skip(1).collect();
+    if mode == "delay" {
+        std::thread::sleep(Duration::from_millis(parameter.parse().unwrap()));
+    }
+    let batch = args.get(1).is_some_and(|arg| arg == "--batch-check");
+    let input = if batch {
+        let mut bytes = Vec::new();
+        std::io::stdin().read_to_end(&mut bytes)?;
+        Some(bytes)
+    } else {
+        None
+    };
+    let cwd = std::env::current_dir()?;
+    let hang = mode == "hang"
+        || (mode == "hang-repo" && cwd.file_name().is_some_and(|name| name == parameter))
+        || (mode == "hang-oid"
+            && (args.iter().any(|arg| arg == parameter)
+                || input
+                    .as_ref()
+                    .is_some_and(|bytes| String::from_utf8_lossy(bytes).trim() == parameter)));
+    if hang {
+        std::thread::sleep(Duration::from_secs(30));
+        std::process::exit(1);
+    }
+    if mode == "log-types" && batch {
+        let mut log = std::fs::OpenOptions::new()
+            .create(true)
+            .append(true)
+            .open(parameter)?;
+        writeln!(
+            log,
+            "{} {}",
+            String::from_utf8_lossy(input.as_ref().unwrap()).trim(),
+            cwd.file_name().unwrap().to_string_lossy()
+        )?;
+    }
+    let mut child = Command::new("git")
+        .args(&args)
+        .stdin(if input.is_some() {
+            Stdio::piped()
+        } else {
+            Stdio::inherit()
+        })
+        .stdout(Stdio::inherit())
+        .stderr(Stdio::inherit())
+        .spawn()?;
+    if let Some(bytes) = input {
+        child.stdin.take().unwrap().write_all(&bytes)?;
+    }
+    std::process::exit(child.wait()?.code().unwrap_or(1));
+}
@@ -29,6 +29,7 @@ fn inv22_concurrency_gates_present_and_not_bypassed() {
    let repos = src("api/repos.rs");
    let smart_http = src("git/smart_http.rs");
    let vis = src("git/visibility_pack.rs");
+    let ipfs = src("api/ipfs.rs");

    // U1 / P1-a: run_bounded_git stands the watchdog down only after confirming the
    // child actually terminated (WNOWAIT), not on the raw stdout-drain EOF — otherwise
@@ -68,6 +69,57 @@ fn inv22_concurrency_gates_present_and_not_bypassed() {
         withheld_recipients_gated, which acquires git_encrypt_semaphore"
    );

+    // U1 / R2 (#173 round-10): the path-scoped filtered-pack serve must thread the
+    // caller's AdmissionGuard through BOTH git stages so read + per-caller admission is
+    // held until the pack-objects group is reaped on disconnect, closing the cap bypass
+    // the plain upload_pack path already fixed. Two load-bearing markers: rev-list hands
+    // the disarmed guard back (its tuple return type), and build_filtered_pack forwards
+    // that guard into the pack-objects stage (the `admission` arg after the
+    // "pack-objects" label). Reverting either — dropping the guard between stages, or
+    // passing `None` to pack-objects — trips this.
+    assert!(
+        smart_http.contains("Result<(Vec<String>, Option<AdmissionGuard>)>")
+            && smart_http.contains("\"pack-objects\",\n        admission,"),
+        "U1/R2 gate missing: build_filtered_pack must thread the AdmissionGuard through \
+         rev-list -> pack-objects so the permits are held until the pack-objects group \
+         is reaped on disconnect (the path-scoped half of #174 P1-a)"
+    );
+
+    // U2 / R1: a cancelled or timed-out `GET /ipfs/{cid}` must release admission only
+    // after the blocking work it admitted has finished, not the instant the handler
+    // future drops (the /ipfs half of #174 P1-a). The mechanism is the shared
+    // `Arc<WalkAdmission>`: both walk permits are moved into it once per request and a
+    // clone rides every `spawn_blocking`, so the permits release when the LAST holder
+    // drops. Reverting to handler-local permits trips this; the per-site clones are
+    // bound separately by `inv22_ipfs_walk_admission_reaches_every_blocking_site`.
+    //
+    // This previously required the permits to be owned by a detached `tokio::spawn`
+    // running the whole pipeline. That closed the same bypass but kept an abandoned
+    // request's full legacy scan running against a held slot, so the merge of #173 and
+    // #174 kept the Arc and dropped the detached task.
+    assert!(
+        ipfs.contains("struct WalkAdmission")
+            && ipfs.contains("let admission = std::sync::Arc::new(WalkAdmission {"),
+        "U2/R1 gate missing: get_by_cid must move both /ipfs admission permits into a \
+         shared Arc<WalkAdmission> whose clones ride the blocking work, so admission is \
+         released only once that work completes (the /ipfs half of #174 P1-a)"
+    );
+
+    // U2 / KTD2 (#173 round-10): the probe/read children on the /ipfs path must be the
+    // duration-bounded twins (process-group teardown via run_bounded_git), not the bare
+    // `store::object_type` / `read_object_content` (or an unbounded `cat-file -s`) a tokio
+    // timeout cannot cancel — otherwise a wedged cat-file lingers and pins the held
+    // admission past the deadline. Reverting any twin call site back to a bare read trips
+    // this.
+    assert!(
+        ipfs.contains("object_type_bounded(")
+            && ipfs.contains("object_size_bounded(")
+            && ipfs.contains("read_object_content_bounded("),
+        "U2/KTD2 gate missing: the /ipfs probe+read must call the run_bounded_git-backed \
+         *_bounded twins so a wedged cat-file is reaped at the deadline, not left to pin \
+         the held /ipfs walk admission"
+    );
+
    // P1-e non-bypass tripwire: the bounded recipients walk is spawn_blocking'd nowhere
    // but inside withheld_recipients_gated. A second call site (count > 1) is a new
    // detached git walk that skips the admission gate — exactly the class U5 closed.
@@ -95,8 +147,12 @@ fn inv22_concurrency_gates_present_and_not_bypassed() {
    // recovery copies are absent until an unrelated later push. Scan only the
    // production half of the file — the u5 tests in its `mod tests` also name the
    // drain call, and matching them would make this check vacuous.
+    // Split at the TEST MODULE, not at the first `#[cfg(test)]`. api/repos.rs carries
+    // test-only items (the drain fault seam, the task entry point) ABOVE the production
+    // code these gates scan for, so splitting on the attribute truncated the production
+    // half above every line being checked and the gate went vacuously blind.
    let repos_production = repos
-        .split("#[cfg(test)]")
+        .split("\nmod tests {")
        .next()
        .expect("split always yields a first chunk");
    assert!(
@@ -169,7 +225,8 @@ fn f4_release_keeps_conn_owned_until_unlock_resolves() {
    let repo_store = src("git/repo_store.rs");

    let rel_start = repo_store
-        .find("pub async fn release(mut self")
+        .find("pub async fn release(self")
+        .or_else(|| repo_store.find("pub async fn release(mut self"))
        .expect("F4 gate: repo_store.rs no longer defines RepoWriteGuard::release");
    let rel_end = repo_store[rel_start..]
        .find("impl Drop for RepoWriteGuard")
@@ -201,10 +258,10 @@ fn f4_release_keeps_conn_owned_until_unlock_resolves() {
    );
 }

-/// F6/KTD-5 (initial IPFS metadata queries deadline-wrapped): `get_by_cid` acquires
-/// the scarce walk permits (RAII, held for the whole request) BEFORE its two initial
-/// metadata queries, and the per-repo loop's first budget gate runs only later. So
-/// both `list_all_repos` and `list_visibility_rules_for_repos` must be clamped to the
+/// F6/KTD-5 (IPFS metadata queries deadline-wrapped): `get_by_cid` acquires
+/// the scarce walk permits (RAII, held for the whole request) BEFORE its metadata
+/// queries, and the per-repo loop's first budget gate runs only later. So
+/// both `list_repos_page_for_scan` and `list_visibility_rules_for_repos` must be clamped to the
 /// remaining request budget — otherwise a query blocked in Postgres pins the walk slot
 /// for the whole stall, past the budget. This scans the PRODUCTION half of `api/ipfs.rs`
 /// (the `mod tests` half names the same calls in its own harness and would make the
@@ -214,31 +271,44 @@ fn f4_release_keeps_conn_owned_until_unlock_resolves() {
 #[test]
 fn f6_ipfs_metadata_queries_are_deadline_wrapped() {
    let ipfs = src("api/ipfs.rs");
+    // Split at the TEST MODULE, not at the first `#[cfg(test)]`: the handler carries
+    // in-line `#[cfg(test)]` query counters, so splitting on the attribute cut the
+    // production half off above the calls this guard exists to check and the scan went
+    // vacuously quiet (found 0 of each rather than failing on a missing wrapper).
    let production = ipfs
-        .split("#[cfg(test)]")
+        .split("\nmod tests {")
        .next()
        .expect("split always yields a first chunk");

-    for call in [".list_all_repos()", ".list_visibility_rules_for_repos("] {
-        assert_eq!(
-            production.matches(call).count(),
-            1,
-            "F6 guard stale: `{call}` must appear exactly once in the production half \
-             of api/ipfs.rs (the deadline-wrapped handler call) — update this guard"
-        );
-        let idx = production
-            .find(call)
-            .unwrap_or_else(|| panic!("F6 gate: api/ipfs.rs no longer calls `{call}`"));
-        // The wrapper opens a few lines above the call (match tokio::time::timeout( ...
-        // remaining budget ..., <query> )); a 240-char lookback covers it without
-        // reaching the previous statement.
-        let window = &production[idx.saturating_sub(240)..idx];
+    // EVERY occurrence must be wrapped, not just the first. The handler now reaches
+    // these queries from two places (the provenance fast path, one repo at a time, and
+    // the legacy scan's preload), and an exact-count check would have to be relaxed
+    // every time a call site is added, which is how a guard quietly stops covering the
+    // site that matters. Requiring all of them scales with the code instead.
+    for call in [
+        ".list_repos_page_for_scan(",
+        ".list_visibility_rules_for_repos(",
+        ".get_repo_by_id(",
+        ".is_repo_quarantined(",
+    ] {
+        let occurrences = production.matches(call).count();
        assert!(
-            window.contains("tokio::time::timeout("),
-            "F6 gate missing: `{call}` must be wrapped in tokio::time::timeout(...) \
-             clamped to the remaining request budget. An unwrapped await pins the held \
-             walk permit for the whole DB stall, past GITLAWB_IPFS_REQUEST_BUDGET_SECS."
+            occurrences >= 1,
+            "F6 guard stale: api/ipfs.rs no longer calls `{call}` — update this guard"
        );
+        for (n, (idx, _)) in production.match_indices(call).enumerate() {
+            // The wrapper opens a few lines above the call (match tokio::time::timeout(
+            // ... remaining budget ..., <query> )); a 240-char lookback covers it
+            // without reaching the previous statement.
+            let window = &production[idx.saturating_sub(240)..idx];
+            assert!(
+                window.contains("tokio::time::timeout("),
+                "F6 gate missing: occurrence {n} of `{call}` (of {occurrences}) is not \
+                 wrapped in tokio::time::timeout(...) clamped to the remaining request \
+                 budget. An unwrapped await pins the held walk permit for the whole DB \
+                 stall, past GITLAWB_IPFS_REQUEST_BUDGET_SECS."
+            );
+        }
    }
 }

@@ -261,8 +331,12 @@ fn f6_ipfs_metadata_queries_are_deadline_wrapped() {
 #[test]
 fn f2_pinata_enqueues_refs_not_retained_object_lists() {
    let repos = src("api/repos.rs");
+    // Split at the TEST MODULE, not at the first `#[cfg(test)]`. api/repos.rs carries
+    // test-only items (the drain fault seam, the task entry point) ABOVE the production
+    // code these gates scan for, so splitting on the attribute truncated the production
+    // half above every line being checked and the gate went vacuously blind.
    let production = repos
-        .split("#[cfg(test)]")
+        .split("\nmod tests {")
        .next()
        .expect("split always yields a first chunk");

@@ -321,8 +395,12 @@ fn f2_pinata_enqueues_refs_not_retained_object_lists() {
 fn f3_second_writer_leased_until_reap() {
    let repos = src("api/repos.rs");
    let smart_http = src("git/smart_http.rs");
+    // Split at the TEST MODULE, not at the first `#[cfg(test)]`. api/repos.rs carries
+    // test-only items (the drain fault seam, the task entry point) ABOVE the production
+    // code these gates scan for, so splitting on the attribute truncated the production
+    // half above every line being checked and the gate went vacuously blind.
    let repos_production = repos
-        .split("#[cfg(test)]")
+        .split("\nmod tests {")
        .next()
        .expect("split always yields a first chunk");

@@ -338,8 +416,8 @@ fn f3_second_writer_leased_until_reap() {
         NOT RepoWriteGuard (which drops at the disconnect instant, reopening F3).",
    );
    let receive = repos_production
-        .find("smart_http::receive_pack(")
-        .expect("F3 gate stale: git_receive_pack no longer calls smart_http::receive_pack");
+        .find("smart_http::receive_pack_raw")
+        .expect("F3 gate stale: git_receive_pack no longer calls smart_http::receive_pack_raw");
    assert!(
        lease_acquire < with_lease && with_lease < receive,
        "F3 gate bypassed: the write lease must be acquired, then carried by the \
@@ -372,7 +450,7 @@ fn f3_second_writer_leased_until_reap() {
 /// exists: it binds all three sites, and any blocking site added to this loop
 /// later, without reversing that deliberate independence.
 ///
-/// MUTATION (RED): delete any one `Arc::clone(&admission)` binding, or drop the
+/// MUTATION (RED): delete any one `Arc::clone(ctx.admission)` binding, or drop the
 /// clone from inside its closure, and the count falls below three.
 #[test]
 fn inv22_ipfs_walk_admission_reaches_every_blocking_site() {
@@ -386,7 +464,10 @@ fn inv22_ipfs_walk_admission_reaches_every_blocking_site() {
    );

    // Every blocking site in the scan takes its own clone...
-    let clones = ipfs.matches("Arc::clone(&admission)").count();
+    // `Arc::clone(ctx.admission)` in the gate, `Arc::clone(&admission)` if a site is
+    // ever added back in the handler body itself; count both spellings.
+    let clones = ipfs.matches("Arc::clone(ctx.admission)").count()
+        + ipfs.matches("Arc::clone(&admission)").count();
    assert!(
        clones >= 3,
        "U1 gate bypassed: expected an admission clone for each of the three \
@@ -428,8 +509,8 @@ fn inv22_ipfs_walk_admission_reaches_every_blocking_site() {
 /// too, so an ungated spawn would fire for a push git rejected, pinning and announcing
 /// a half-applied repo. Above `release` the `?` on `receive_result` can no longer be
 /// what gates it, so the success check is explicit and this gate binds it: the spawn
-/// must sit inside `if push_succeeded`, and `release` must consume the same flag so
-/// the two cannot drift apart.
+/// must sit inside `if disposition.spawn_tail`, and `release` must consume the same
+/// disposition (`disposition.release_ok`) so the two cannot drift apart.
 ///
 /// This is an ordering check rather than a cancellation-race test on purpose: it is
 /// the companion to `receive_pack_tail_survives_a_disconnect_during_release`, which
@@ -438,52 +519,606 @@ fn inv22_ipfs_walk_admission_reaches_every_blocking_site() {
 ///
 /// MUTATION (RED): move the `tokio::spawn(post_receive_replication_tail` call below
 /// `guard.release(` and the ordering assertion fails; take it out of the
-/// `if push_succeeded` block and the failed-push assertion fails.
+/// `if disposition.spawn_tail` block and the failed-push assertion fails; gate
+/// `spawn_tail`/`release_ok` on the outcome commit and
+/// `post_git_disposition_replication_carve_out` fails with it.
 #[test]
 fn inv22_replication_tail_spawns_at_the_durability_boundary() {
    let repos = src("api/repos.rs");
    // Production half only — the tests below name these identifiers too.
+    // Split at the TEST MODULE, not at the first `#[cfg(test)]`. api/repos.rs carries
+    // test-only items (the drain fault seam, the task entry point) ABOVE the production
+    // code these gates scan for, so splitting on the attribute truncated the production
+    // half above every line being checked and the gate went vacuously blind.
    let production = repos
-        .split("#[cfg(test)]")
+        .split("\nmod tests {")
        .next()
        .expect("split always yields a first chunk");

    let success_flag = production
-        .find("let push_succeeded = receive_result.is_ok();")
-        .expect(
-            "U5 gate missing: the tail's success gate must be bound from receive_result.is_ok()",
+        .find("post_git_disposition(")
+        .expect("U5 gate missing: the tail's disposition must be bound before the gate");
+    // One tail per push, on the proven-landed ref set. Either spawn
+    // position (carve-out early, or durability-bound at release) must sit
+    // inside a `disposition.spawn_tail` gate and feed `landed_refs` —
+    // never the unfiltered `ref_updates`.
+    let spawn_count = production
+        .matches("tokio::spawn(post_receive_replication_tail(")
+        .count();
+    assert!(
+        spawn_count >= 1,
+        "U5 gate missing: the replication tail must be spawned by git_receive_pack"
+    );
+    for (i, (pos, _)) in production
+        .match_indices("tokio::spawn(post_receive_replication_tail(")
+        .enumerate()
+    {
+        let gate = production[..pos].rfind("if disposition.spawn_tail").expect(
+            "U5 gate missing: every tail spawn must be gated on `if disposition.spawn_tail`",
        );
-    let gate_open = production
-        .find("if push_succeeded {")
-        .expect("U5 gate missing: the tail spawn must be gated on the push having succeeded");
+        // The spawn must sit inside the gate's block: brace depth at the
+        // spawn position must still be positive measured from the `if`.
+        // An ungated spawn after the block closes fails this, as does a
+        // stray `disposition.spawn_tail` mention in a comment.
+        let mut depth = 0i32;
+        for ch in production[gate..pos].chars() {
+            match ch {
+                '{' => depth += 1,
+                '}' => depth -= 1,
+                _ => {}
+            }
+        }
+        assert!(
+            depth >= 1,
+            "U5 gate bypassed: tail #{i} sits outside the `if disposition.spawn_tail` block"
+        );
+        assert!(
+            !production[gate..pos].contains("post_receive_replication_tail"),
+            "U5 gate bypassed: a second tail spawn sits between the gate and its spawn"
+        );
+        let call_end = production[pos..].find(");").unwrap() + pos;
+        let call = &production[pos..call_end];
+        assert!(
+            call.contains("landed_refs"),
+            "U5 gate bypassed: tail #{i} must carry the proven-landed ref set, not unfiltered ref_updates"
+        );
+    }
    let spawn = production
        .find("tokio::spawn(post_receive_replication_tail(")
        .expect("U5 gate missing: the replication tail must be spawned by git_receive_pack");
    let release = production
-        .find("guard.release(push_succeeded)")
-        .expect("U5 gate stale: release must consume the same success flag as the tail gate");
+        .find(".release(disposition.release_ok)")
+        .expect("U5 gate stale: release must consume the same disposition as the tail gate");
    let touch = production
        .find("state.db.touch_repo(")
        .expect("U5 gate stale: git_receive_pack no longer calls touch_repo");
-    let webhook = production
-        .find("webhooks::fire_event(")
-        .expect("U5 gate stale: git_receive_pack no longer fires push webhooks");
+    // #26 Split PR 1 step 3 — the webhook fan-out moved into
+    // `durable_outbox::apply_request_effects`, which the live
+    // handler calls inline (the recovery drain calls the same
+    // function on the next startup). The gate now pins that the
+    // handler is wired to the executor, not to a per-ref inline
+    // webhook call.
+    let effects_executor = production
+        .find("apply_request_effects(&state, &request_id)")
+        .expect("U5 gate stale: git_receive_pack no longer calls apply_request_effects inline");

    assert!(
-        success_flag < gate_open && gate_open < spawn,
-        "U5 gate bypassed: the tail must be spawned inside `if push_succeeded`, or a \
-         rejected push spawns a tail that pins and announces a half-applied repo"
-    );
-    // Still inside that block: no `}` may close it between the gate and the spawn.
-    assert!(
-        !production[gate_open + "if push_succeeded {".len()..spawn].contains('}'),
-        "U5 gate bypassed: the tail spawn left the `if push_succeeded` block, so a \
-         rejected push now spawns a tail"
+        success_flag < spawn,
+        "U5 gate bypassed: the disposition must be bound before any tail spawn"
    );
    assert!(
-        spawn < release && spawn < touch && spawn < webhook,
+        spawn < release && spawn < touch && spawn < effects_executor,
        "U5 gate bypassed: the tail must be spawned BEFORE guard.release, touch_repo \
-         and the webhook fan-out, so a disconnect in any of those windows cannot drop \
+         and the effect executor, so a disconnect in any of those windows cannot drop \
         this push's pins, recovery copy, and announcements"
    );
 }
+
+/// #26 Split PR 1 step 3 — drain + handler share a single effect
+/// executor. The live handler (api/repos.rs) and the recovery
+/// drain (`durable_outbox::drain_receive_pack_requests_with`) both
+/// call `apply_request_effects`, so the per-ref effects fan-out
+/// lives in exactly one place. The v29 per-ref walk
+/// (`derive_one`, `drain_pending_ref_transitions_all`,
+/// `lookup_accepted_ordinal`) is dead code: any caller reintroduced
+/// would be the per-ref walk the step-3 PR removed. This gate
+/// fails if a call site slips back in or the new seam is bypassed.
+#[test]
+fn inv26_step3_live_and_drain_share_apply_request_effects() {
+    let repos = src("api/repos.rs");
+    let outbox = src("durable_outbox.rs");
+
+    // The live handler calls `apply_request_effects`. Split the
+    // file at the test attribute so test code can't satisfy the
+    // gate by itself.
+    let production_repos = repos
+        .split("\nmod tests {")
+        .next()
+        .expect("split always yields a first chunk");
+
+    assert!(
+        production_repos.contains("apply_request_effects(&state, &request_id)"),
+        "live handler must call `apply_request_effects(&state, &request_id)`; \
+         reverting to a per-ref inline fan-out splits live and recovery"
+    );
+
+    // The durable intent insert closes the pre-outbox crash window this
+    // PR exists to fix. Wrapping it in `if false` must turn this gate
+    // red; drain-only tests cannot prove the handler creates the rows.
+    assert!(
+        production_repos.contains("insert_receive_pack_request_with_children("),
+        "live handler must persist durable intent via \
+         `insert_receive_pack_request_with_children`; disabling it must fail this gate"
+    );
+
+    // The drain's per-request seam calls the same executor. Test
+    // code lives below `mod drain_tests`, so split there too.
+    let production_outbox = outbox
+        .split("\nmod drain_tests {")
+        .next()
+        .expect("split always yields a first chunk");
+
+    assert!(
+        production_outbox.contains("drain_receive_pack_requests_with"),
+        "drain seam `drain_receive_pack_requests_with` must exist; \
+         removing it forces a per-ref walk back into the drain"
+    );
+    assert!(
+        production_outbox.contains("apply_request_effects"),
+        "durable_outbox production code must define `apply_request_effects`; \
+         removing it splits the executor between live and recovery"
+    );
+
+    // The drain's per-request walker is wired to the executor.
+    // The closure body of `drain_receive_pack_requests_with` is the
+    // only call site — if a future change calls `derive_one`
+    // instead, this assertion fires.
+    let drain_seam_open = production_outbox
+        .find("pub async fn drain_receive_pack_requests_with<F, Fut>")
+        .expect("drain seam must be defined in the production half");
+    let drain_seam_close = production_outbox[drain_seam_open..]
+        .find("\n}\n")
+        .expect("drain seam body must close");
+    let drain_seam_body = &production_outbox[drain_seam_open..drain_seam_open + drain_seam_close];
+    assert!(
+        drain_seam_body.contains("apply_request_effects"),
+        "drain_receive_pack_requests_with must call apply_request_effects; \
+         wiring it to a per-ref helper reintroduces the v29 walk"
+    );
+
+    // The deleted per-ref drain must have zero call sites in the
+    // production half. A regression that re-adds a caller would
+    // bring back the per-ref fan-out.
+    assert!(
+        !production_outbox.contains("drain_pending_ref_transitions_all("),
+        "deleted `drain_pending_ref_transitions_all` must have zero production call sites; \
+         a per-ref walk is reintroduced"
+    );
+    assert!(
+        !production_outbox.contains("derive_one("),
+        "deleted `derive_one` must have zero production call sites; \
+         the per-ref fan-out is reintroduced"
+    );
+    assert!(
+        !production_outbox.contains("lookup_accepted_ordinal("),
+        "deleted `lookup_accepted_ordinal` must have zero production call sites; \
+         the per-ref ordinal lookup is reintroduced"
+    );
+    assert!(
+        !production_repos.contains("derive_one("),
+        "deleted `derive_one` must have zero live-handler call sites"
+    );
+    assert!(
+        !production_repos.contains("drain_pending_ref_transitions"),
+        "deleted per-ref drain functions must have zero live-handler call sites"
+    );
+}
+
+/// #26 Split PR 1 step 4 — the periodic queue-lifecycle purge is
+/// wired in `main.rs` and the contract is enforced by the `idx_receive_pack_requests_completed_at`
+/// partial index from v41.
+///
+/// Assertions:
+/// 1. `main.rs` calls `purge_request_queue` once at boot (well, on
+///    the spawn-task interval) and the spawn function exists.
+/// 2. The DB helpers `purge_completed_receive_pack_requests` and
+///    `purge_completed_pending_ref_transitions` exist in `db/mod.rs`.
+/// 3. The config knobs `queue_retention_days` and `queue_purge_batch` exist
+///    on `Config`.
+/// 4. `quarantined` is NOT in the purge WHERE clause (step 4 does not
+///    introduce the state, but the invariant holds for the future).
+#[test]
+fn inv26_step4_queue_lifecycle_purge_is_wired() {
+    let main_src = src("main.rs");
+    assert!(
+        main_src.contains("spawn_queue_lifecycle_sweep"),
+        "main.rs must spawn the periodic queue-lifecycle purge"
+    );
+    assert!(
+        main_src.contains("purge_request_queue"),
+        "main.rs must call purge_request_queue on the periodic sweep"
+    );
+
+    let db_src = src("db/mod.rs");
+    assert!(
+        db_src.contains("purge_completed_receive_pack_requests"),
+        "db/mod.rs must expose purge_completed_receive_pack_requests"
+    );
+    assert!(
+        db_src.contains("purge_completed_pending_ref_transitions"),
+        "db/mod.rs must expose purge_completed_pending_ref_transitions"
+    );
+
+    let config_src = src("config.rs");
+    assert!(
+        config_src.contains("queue_retention_days"),
+        "Config must expose queue_retention_days"
+    );
+    assert!(
+        config_src.contains("queue_purge_batch"),
+        "Config must expose queue_purge_batch"
+    );
+}
+
+/// #26 Split PR 1 step 5 — the marker gate and the retry-bound
+/// quarantine are wired end-to-end. A missing or hash-mismatched
+/// marker on disk quarantines the request; a Retry over the bound
+/// does too. This gate pins every load-bearing seam, against the
+/// production half of each file (the test modules name the same
+/// identifiers in their own harnesses).
+///
+/// Assertions:
+/// 1. `reconcile_prepared_page` calls `mark_request_quarantined`
+///    and `mark_children_rejected_for_quarantined_parent` on
+///    marker-gate failure.
+/// 2. The drain's `EffectsOutcome::Retry` arm checks
+///    `effects_max_attempts`.
+/// 3. `git::store::read_ref` exists in `git/store.rs`.
+/// 4. `db::mark_request_quarantined`,
+///    `db::mark_children_rejected_for_quarantined_parent`,
+///    `db::get_receive_pack_requests_by_ids`, and
+///    `db::request_state::QUARANTINED` all exist.
+/// 5. The handler writes the marker ref BEFORE calling
+///    `smart_http::receive_pack_raw` (the durability window).
+#[test]
+fn inv26_step5_marker_quarantine_and_bound_are_wired() {
+    let outbox = src("durable_outbox.rs");
+    let store = src("git/store.rs");
+    let db = src("db/mod.rs");
+    let repos = src("api/repos.rs");
+
+    // Split at the TEST MODULE for the production-only assertions.
+    let production_outbox = outbox
+        .split("\nmod drain_tests {")
+        .next()
+        .expect("split always yields a first chunk");
+    let production_repos = repos
+        .split("\nmod tests {")
+        .next()
+        .expect("split always yields a first chunk");
+
+    // (1) The reconcile's marker gate quarantines via the DB helpers.
+    assert!(
+        production_outbox.contains("mark_request_quarantined"),
+        "reconcile_prepared_page must call mark_request_quarantined on marker-gate failure"
+    );
+    assert!(
+        production_outbox.contains("mark_children_rejected_for_quarantined_parent"),
+        "reconcile_prepared_page must call mark_children_rejected_for_quarantined_parent \
+         so quarantined parents cancel their children"
+    );
+
+    // The gate reads the marker ref and compares against the parent's
+    // `request_bytes_hash` via `git::store::read_ref` /
+    // `git::store::marker_value_for`. Reverting either reintroduces
+    // the DoS window the marker gate exists to close.
+    assert!(
+        production_outbox.contains("git::store::read_ref"),
+        "reconcile_prepared_page must read the marker ref via git::store::read_ref"
+    );
+    assert!(
+        production_outbox.contains("marker_value_for"),
+        "reconcile_prepared_page must compute the expected marker value via \
+         git::store::marker_value_for"
+    );
+
+    // (2) The drain's `EffectsOutcome::Retry` arm checks the bound.
+    assert!(
+        production_outbox.contains("effects_max_attempts"),
+        "drain_receive_pack_requests_with must consult effects_max_attempts on Retry"
+    );
+
+    // (3) `git::store::read_ref` is the read seam the gate depends on.
+    assert!(
+        store.contains("pub fn read_ref("),
+        "git::store::read_ref must exist; the marker gate reads through it"
+    );
+    assert!(
+        store.contains("pub fn marker_value_for("),
+        "git::store::marker_value_for must exist; the marker gate computes the expected \
+         value with it (and the live handler writes the value via the same helper)"
+    );
+
+    // (4) DB-side seams the gate depends on.
+    assert!(
+        db.contains("pub async fn mark_request_quarantined"),
+        "Db::mark_request_quarantined must exist"
+    );
+    assert!(
+        db.contains("pub async fn mark_children_rejected_for_quarantined_parent"),
+        "Db::mark_children_rejected_for_quarantined_parent must exist"
+    );
+    assert!(
+        db.contains("pub async fn get_receive_pack_requests_by_ids"),
+        "Db::get_receive_pack_requests_by_ids must exist (avoids N+1 in the marker gate)"
+    );
+    assert!(
+        db.contains("pub const QUARANTINED: &str = \"quarantined\""),
+        "request_state::QUARANTINED must be defined"
+    );
+
+    // (5) The handler writes the marker ref BEFORE the durability
+    // boundary (smart_http::receive_pack_raw). Severing the
+    // ordering re-opens the marker-gate DoS window for live pushes.
+    let marker_write = production_repos
+        .find("git::store::marker_value_for")
+        .expect("U5 gate stale: the live handler no longer computes the marker value");
+    let receive_raw = production_repos
+        .find("smart_http::receive_pack_raw")
+        .expect("U5 gate stale: git_receive_pack no longer calls smart_http::receive_pack_raw");
+    assert!(
+        marker_write < receive_raw,
+        "U5 gate bypassed: the marker ref must be written BEFORE receive_pack_raw so the \
+         reconcile's gate has evidence of the live push"
+    );
+}
+
+/// Issue #26 split 1 — namespace gate and pre-git refusal lifecycle.
+///
+/// The push gate (repos.rs) must deny:
+/// 1. Any ref-update line that fails UTF-8 decode (the raw body is passed verbatim
+///    to git receive-pack, and git accepts refnames with bytes >= 0x80; the gate
+///    never sees these refs, no durable child row exists, and under requests/issues/
+///    the lossy for-each-ref decode still matches the narrowed exemption).
+/// 2. The bare `refs/gitlawb` name (the `starts_with("refs/gitlawb/")` check misses it).
+///
+/// Every explicit pre-git refusal must call `refuse_receive_pack_before_git` to
+/// terminalize the intent aggregate in one transaction, or a dropped future between
+/// the intent insert and `receive_pack` strands a `received` + `prepared` aggregate
+/// that is never purged, never promotable, and poisons its tuple.
+///
+/// The visibility exemption (visibility_pack.rs) was narrowed to `requests/` and
+/// `issues/` only. The push gate is the premise that narrowing relies on (the comment
+/// at visibility_pack.rs:340), so tests must prove:
+/// - A `refs/gitlawb/<other>` ref pointing at a blob is denied on push.
+/// - A stray `refs/gitlawb/<other>` blob ref planted in the pack path fails closed.
+///
+/// Source gates below pin the wiring; behavioral proofs live next to the code:
+/// `parse_ref_updates_refuses_non_utf8_command_line` and
+/// `internal_namespace_gate_covers_bare_name_and_subtree` (api/repos.rs),
+/// `planted_non_exempt_gitlawb_blob_ref_fails_closed` (git/visibility_pack.rs),
+/// and the mirror import prune
+/// `prune_removes_non_exempt_gitlawb_refs_and_keeps_exempt` (sync.rs).
+#[test]
+fn issue_26_namespace_gate_and_refusal_wiring() {
+    let repos = src("api/repos.rs");
+    let vis = src("git/visibility_pack.rs");
+
+    // P1: fail the push on ref-update lines the handler cannot decode.
+    // The gate at parse_ref_updates must refuse when any command line in the
+    // ref-update section is undecodable UTF-8. The check is an Err arm on
+    // `from_utf8(data)` that returns BadRequest, so removing it allows non-UTF-8
+    // refs to bypass the namespace gate.
+    assert!(
+        repos.contains("ref-update line is not valid UTF-8; push refused"),
+        "P1 gate missing: parse_ref_updates must refuse when any ref-update line is \
+         non-UTF-8 (git accepts bytes >= 0x80; the gate never sees them, no durable \
+         child exists, and the lossy for-each-ref decode still matches the exemption)"
+    );
+
+    // P1 (cont): the bare `refs/gitlawb` name must also be caught. The
+    // predicate is extracted as `ref_is_internal_namespace` so it is
+    // unit-testable; pin the bare-name arm inside it and its use at the
+    // push gate (a `starts_with("refs/gitlawb/")`-only check misses the
+    // bare name, and it lands on repos with no internal refs yet).
+    assert!(
+        repos.contains("ref_name == \"refs/gitlawb\"")
+            && repos.contains(".find(|r| ref_is_internal_namespace(r))"),
+        "P1 gate missing: ref_is_internal_namespace must match the bare \
+         'refs/gitlawb' name and the push gate must consult it \
+         (starts_with('refs/gitlawb/') misses it, and it lands on repos with \
+         no internal refs yet)"
+    );
+
+    // P1 (cont): every explicit pre-git refusal must call refuse_receive_pack_before_git
+    // to terminalize the intent. Exactly 6 occurrences: the five inline call
+    // sites (lease cap, per-source cap, write pool, acquire timeout, acquire
+    // error) plus the drop-guard's own pre-git spawn. The post-git guard phase
+    // calls mark_receive_pack_interrupted (a different name), so deleting any
+    // one inline site drops the count to 5 and turns this red. The parse and
+    // namespace refusals correctly carry no call: they return before the
+    // intent insert, so there is no aggregate to terminalize.
+    assert_eq!(
+        repos.matches("refuse_receive_pack_before_git(").count(),
+        6,
+        "P1 wiring: exactly 6 refuse_receive_pack_before_git calls expected (5 \
+         inline refusal sites + guard pre-git spawn); deleting any inline site \
+         must turn this red"
+    );
+    // Pin the five inline sites by their reason literals, so a site that stops
+    // terminalizing (or is renumbered away) fails by name, not just by count.
+    for reason in [
+        "repo write-lease waiter cap reached",
+        "admission per-source cap reached",
+        "git write pool saturated",
+        "acquire_write timed out",
+        "acquire_write failed",
+        "handler dropped before git started",
+    ] {
+        assert!(
+            repos.contains(reason),
+            "P1 wiring: refusal reason literal missing: {reason}"
+        );
+    }
+
+    // P1 (cont): the intent drop-guard must be phase-aware. It is armed before
+    // the insert resolves (covering the cancellation window), flipped to
+    // git-started immediately before the receive_pack await (a drop from there
+    // is a mid-git disconnect, terminalized via mark_receive_pack_interrupted
+    // because the Err arm never runs on a drop), and disarmed once the outcome
+    // commit resolves. Assert the ordering, not just presence: the phase flip
+    // must precede the receive_pack call.
+    assert!(
+        repos.contains("IntentDropGuard::new") && repos.contains("notify_git_started()"),
+        "P1 gate missing: IntentDropGuard must exist and flip to git-started phase \
+         before receive_pack (a dropped future between intent insert and git start \
+         strands a received+prepared aggregate)"
+    );
+    let flip_pos = repos
+        .find("notify_git_started()")
+        .expect("notify_git_started call must exist");
+    let git_pos = repos
+        .find("receive_pack_raw_with_reflog(")
+        .expect("receive_pack call must exist");
+    assert!(
+        flip_pos < git_pos,
+        "P1 ordering: notify_git_started() must precede the receive_pack call, \
+         or a mid-git drop is misclassified as a pre-git shed"
+    );
+    assert!(
+        repos.contains("_intent_guard.disarm()"),
+        "P1 gate missing: the guard must disarm once the outcome commit resolves, \
+         or every push spawns a wasteful cleanup txn at scope end"
+    );
+    // A sync Drop cannot await: cleanup is spawned via the current runtime
+    // handle with an off-runtime fallback. A bare tokio::spawn panics
+    // mid-unwind with no runtime (off-runtime drop, process teardown).
+    assert!(
+        repos.contains("Handle::try_current()"),
+        "P1 gate missing: guard Drop must spawn via Handle::try_current with an \
+         off-runtime arm, not a bare tokio::spawn"
+    );
+
+    // P2: the post-git drop arm must exist and be wired. The guard's
+    // GIT_STARTED phase spawns `mark_receive_pack_interrupted` (parent to
+    // `rejected_at_git`, children to `uncertain` — the Err arm never runs on
+    // a drop). Deleting that spawn keeps every count above green, so pin the
+    // method name; its behavior is proved by
+    // `post_git_interruption_marks_uncertain_and_terminalizes_parent`.
+    assert!(
+        repos.contains("mark_receive_pack_interrupted("),
+        "P2 wiring: the guard's post-git phase must spawn \
+         mark_receive_pack_interrupted (a mid-git drop otherwise strands \
+         received+prepared with no Err arm to mark them uncertain)"
+    );
+    // The guard must commit computed fates rather than a blanket uncertain
+    // once the report is parsed, and only disarm on a resolved commit: pin
+    // both the stash call and the `outcome_commit_ok` gate. Making disarm()
+    // unconditional strands the commit-failure path and misfiles proven
+    // rejections as uncertain.
+    assert!(
+        repos.contains("set_fates(ComputedFates"),
+        "P2 wiring: computed fates must be stashed on the guard once parsed, \
+         or a post-commit drop can only apply a context-free uncertain"
+    );
+    assert!(
+        repos.contains("if outcome_commit_ok"),
+        "P2 wiring: disarm must be gated on outcome_commit_ok; unconditional \
+         disarm strands the persistent-commit-failure path"
+    );
+    // The live-handler `AwaitingSiblings` arm is unreachable through the
+    // current outcome builder (documented at the arm itself) and therefore
+    // has no behavioral test; pin its presence so removing it is conscious.
+    // The drain consumer IS behaviorally driven (sibling-wait tests).
+    assert_eq!(
+        repos.matches("EffectsOutcome::AwaitingSiblings").count(),
+        1,
+        "live AwaitingSiblings arm must exist (defensive symmetry with the \
+         drain consumer); deleting it must turn this red"
+    );
+
+    // P2: pin the mirror prune call sites. The prune helper is called from
+    // both `clone_repo` and `fetch_repo` (plus its own definition) — deleting
+    // either call leaves the direct-helper test green while mirrors import
+    // evil refs again. Count occurrences of the call with its argument.
+    let sync_src = src("sync.rs");
+    assert_eq!(
+        sync_src
+            .matches("prune_non_exempt_gitlawb_refs(local_path)")
+            .count(),
+        2,
+        "P2 wiring: prune must be called from both clone_repo and fetch_repo; \
+         deleting either call must turn this red"
+    );
+    // Both call sites must route through `wrap_prune_error`, which preserves
+    // deterministic `PruneInvalid` refusals instead of stringifying every
+    // error into transient `PruneFailed` (an every-tick deferral spin for
+    // permanently invalid mirrors). Pin the exact call shape at both sites.
+    assert_eq!(
+        sync_src.matches(".map_err(wrap_prune_error)").count(),
+        2,
+        "P2 wiring: both prune call sites must use wrap_prune_error; inline \
+         rewrap must turn this red"
+    );
+
+    // P2: the visibility exemption narrowing removed a serving path. The exemption
+    // is now scoped to `requests/` and `issues/` only (visibility_pack.rs:371), and
+    // the comment at :340 states the push gate is the premise. Tests must prove the
+    // gate works, which the P1 findings above covered (non-UTF-8 line, bare name,
+    // refusal wiring). This check verifies the exemption narrowing landed.
+    assert!(
+        vis.contains(
+            r#"!r.starts_with("refs/gitlawb/requests/") && !r.starts_with("refs/gitlawb/issues/")"#
+        ),
+        "P2 gate missing: the visibility exemption must be narrowed to requests/ and \
+         issues/ only (the push gate is the premise; any other refs/gitlawb/* ref is \
+         denied on push and must fail closed in the pack path)"
+    );
+}
+
+/// Claimed-webhook delivery must not re-list hooks: the durable path claims
+/// synchronously, hands the claimed hook rows to the sender, and a list
+/// failure in the send task would silently drop already-claimed deliveries.
+#[test]
+fn inv26_claimed_hook_delivery_never_relists() {
+    let webhooks = src("webhooks.rs");
+    let production = webhooks
+        .split("\nmod tests {")
+        .next()
+        .expect("split always yields a first chunk");
+
+    // The claimed entry point exists and passes hook ROWS (not ids) so the
+    // sender never needs the hook table again.
+    let entry = production
+        .find("fire_event_occurrence_with_claimed_hooks(")
+        .expect("claimed-hooks sender missing");
+    let claimed_rows = production
+        .find("claimed_hooks: Vec<crate::db::Webhook>")
+        .expect("claimed path must carry Webhook rows, not ids");
+    let list_call = production
+        .find("db.list_webhooks_for_event(repo_id, event)")
+        .expect("unclaimed path still lists hooks");
+    assert!(
+        entry < claimed_rows && claimed_rows < list_call,
+        "claimed sender must hand rows through before any list fallback"
+    );
+
+    // Inside the async worker, the claimed branch must bypass the list.
+    // Inside the async worker (which starts after the claim helper), the
+    // claimed branch must be selected before the list fallback can run.
+    let worker = production
+        .find("async fn fire_event_async_occurrence(")
+        .expect("occurrence worker missing");
+    let already_claimed = production[worker..]
+        .find("let already_claimed = claimed_hooks.is_some();")
+        .map(|i| worker + i)
+        .expect("claimed path must short-circuit before re-listing");
+    let worker_list = production[worker..]
+        .find("db.list_webhooks_for_event(repo_id, event)")
+        .map(|i| worker + i)
+        .expect("unclaimed path still lists hooks");
+    assert!(
+        already_claimed < worker_list,
+        "the claimed branch must be selected before list_webhooks_for_event can run"
+    );
+}
@@ -0,0 +1,177 @@
+//! #26 Split PR 4 — `GITLAWB_ENFORCE_OWNER_PUSH` default is consistent across surfaces.
+//!
+//! Reviewer 2 (2026-08-28) closed PR #224 with a P2 finding: SECURITY.md and
+//! README.md both stated the default was `false`, while `Config::enforce_owner_push`
+//! declared `default_value_t = true` and the runtime, .env.example, and the env
+//! var table on the README all agreed with the code. Two operator-facing
+//! surfaces lied about the authorization policy. PR #330 flipped the code
+//! default and updated .env.example / README env table / docs/RUN-A-NODE.md,
+//! but missed these two prose mentions.
+//!
+//! This guard pins the documentation and example surfaces together. Each one
+//! is read off the working tree at test time, not a copy in the binary, so a
+//! revert of the doc edit turns it red:
+//!
+//! 1. `SECURITY.md` says the default is `true`.
+//! 2. `README.md` says the default is `true` in the known-limitations prose.
+//! 3. `README.md` says the default is `true` in the env-var table row.
+//! 4. `.env.example` sets `GITLAWB_ENFORCE_OWNER_PUSH=true`.
+//!
+//! The runtime declared default (`Config::enforce_owner_push`'s
+//! `default_value_t = true`) is protected by the
+//! `enforce_owner_push_is_declared_true_independent_of_the_environment`
+//! unit test in `config.rs`, which reads the parser directly and is
+//! env-independent. The two checks together cover the surfaces that
+//! produced the original doc-vs-code drift; this guard cannot reach into
+//! the binary crate's `Config` type (it is private to the bin target,
+//! which has no `lib.rs`) without a larger refactor, so the contract is
+//! stated as the four documentation/example surfaces above.
+//!
+//! The four are checked independently. A check that just verified "at least
+//! one surface says true" would pass while a stale SECURITY.md lied; a check
+//! that only verified the binary would not catch the doc drift that
+//! produced this finding in the first place.
+
+use std::path::Path;
+
+fn read(rel: &str) -> String {
+    let p = Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("..")
+        .join("..")
+        .join(rel);
+    std::fs::read_to_string(&p).unwrap_or_else(|e| panic!("read {}: {e}", p.display()))
+}
+
+/// Pin the prose mentions of `GITLAWB_ENFORCE_OWNER_PUSH` so a future doc
+/// edit cannot reintroduce the "defaults to false" wording the reviewer
+/// flagged. The wording in the test is what the docs MUST say about the
+/// default; absence of any prose mention is also a fail because operators
+/// rely on the docs to learn the default.
+#[test]
+fn security_md_states_owner_push_default_is_true() {
+    let security = read("SECURITY.md");
+    let section = section_after(&security, "### Repository write authorization defaults")
+        .unwrap_or_else(|| panic!("SECURITY.md is missing the owner-push section"));
+
+    assert!(
+        section.contains("defaults to `true`"),
+        "SECURITY.md owner-push section must state the default is true. \
+         Found:\n{section}"
+    );
+    assert!(
+        !section_claims_default_is_false(&section),
+        "SECURITY.md owner-push section must not claim the default is false. \
+         Found:\n{section}"
+    );
+}
+
+#[test]
+fn readme_known_limitations_states_owner_push_default_is_true() {
+    let readme = read("README.md");
+    let limitations = section_after(&readme, "Known limitations:")
+        .unwrap_or_else(|| panic!("README.md is missing the Known limitations: section"));
+
+    // The owner-push line is one bullet in the limitations list. Pull the
+    // single line that mentions the env var so the assertion is local to
+    // the claim, not a substring of unrelated text.
+    let line = limitations
+        .lines()
+        .find(|l| l.contains("GITLAWB_ENFORCE_OWNER_PUSH"))
+        .unwrap_or_else(|| {
+            panic!(
+                "README.md Known limitations must mention GITLAWB_ENFORCE_OWNER_PUSH.\n\
+                 Got:\n{limitations}"
+            )
+        });
+
+    assert!(
+        line.contains("defaults to `true`") || line.contains("defaults to true"),
+        "README Known-limitations owner-push bullet must state the default is true. \
+         Found:\n{line}"
+    );
+    assert!(
+        !section_claims_default_is_false(line),
+        "README Known-limitations owner-push bullet must not claim the default is false. \
+         Found:\n{line}"
+    );
+}
+
+#[test]
+fn readme_env_var_table_states_owner_push_default_is_true() {
+    let readme = read("README.md");
+    let line = readme
+        .lines()
+        .find(|l| l.contains("GITLAWB_ENFORCE_OWNER_PUSH") && l.contains("|"))
+        .unwrap_or_else(|| {
+            panic!("README.md is missing the env-var-table row for GITLAWB_ENFORCE_OWNER_PUSH")
+        });
+
+    assert!(
+        line.contains("Defaults to `true`") || line.contains("defaults to `true`"),
+        "README env-var-table row for GITLAWB_ENFORCE_OWNER_PUSH must state the default is true. \
+         Found:\n{line}"
+    );
+    assert!(
+        !section_claims_default_is_false(line),
+        "README env-var-table row for GITLAWB_ENFORCE_OWNER_PUSH must not claim the default is false. \
+         Found:\n{line}"
+    );
+}
+
+#[test]
+fn env_example_sets_owner_push_true() {
+    let env = read(".env.example");
+    let line = env
+        .lines()
+        .find(|l| l.trim_start().starts_with("GITLAWB_ENFORCE_OWNER_PUSH"))
+        .unwrap_or_else(|| panic!(".env.example is missing GITLAWB_ENFORCE_OWNER_PUSH"));
+
+    assert!(
+        line.contains("=true"),
+        ".env.example GITLAWB_ENFORCE_OWNER_PUSH must be set to true. Found:\n{line}"
+    );
+}
+
+/// A claim that the default is false can hide in many phrasings. Match the
+/// forms that have appeared in the wild on this branch: "defaults to `false`",
+/// "defaults to false", "default is false", "is `false` for compatibility",
+/// "is false for compatibility". The match is intentionally narrow — false
+/// positives would suppress a real doc edit.
+fn section_claims_default_is_false(section: &str) -> bool {
+    let lower = section.to_ascii_lowercase();
+    lower.contains("defaults to `false`")
+        || lower.contains("defaults to false")
+        || lower.contains("default is `false`")
+        || lower.contains("default is false")
+        || (lower.contains("`false`") && lower.contains("for compatibility"))
+}
+
+/// Return the prose after a marker up to the next heading boundary. The
+/// marker is normally a markdown heading (e.g. `### Section`); if it has
+/// no `#` prefix it is treated as a plain-text marker and the slice is
+/// bounded by the next heading of any level. This matters because a
+/// plain-text marker would otherwise produce `header_level = 0`, making
+/// the same-or-higher-level stop condition unsatisfiable and letting the
+/// slice run to EOF.
+fn section_after(doc: &str, heading: &str) -> Option<String> {
+    let start = doc.find(heading)?;
+    let after = &doc[start..];
+    let header_level = heading.chars().take_while(|c| *c == '#').count();
+
+    // Skip past the heading line itself.
+    let body_start = after.find('\n')? + 1;
+    let body = &after[body_start..];
+
+    // Stop at the next heading of the same or higher level, or at the next
+    // heading of any level when the marker had no `#` prefix.
+    let mut end = body.len();
+    for line in body.lines() {
+        let level = line.chars().take_while(|c| *c == '#').count();
+        if level > 0 && (header_level == 0 || level <= header_level) {
+            end = line.as_ptr() as usize - body.as_ptr() as usize;
+            break;
+        }
+    }
+
+    Some(body[..end].to_string())
+}
@@ -0,0 +1,637 @@
+//! Which p2p key-storage failures stop the node, and which leave HTTP up.
+//!
+//! The two domains are decided by different code at different phases, so the
+//! only honest proof is at the process boundary: spawn the real binary and
+//! watch what it does. A helper's returned error says nothing about whether
+//! `main` treated it as fatal.
+//!
+//! Every row also asserts the key tree is byte-identical afterwards. A refusal
+//! that mutates storage on its way out is its own defect, and a returned error
+//! cannot show that.
+//!
+//! No database: `DATABASE_URL` points at a closed port on purpose, so a row
+//! that reaches the degraded server proves the outcome is observable without
+//! Postgres. That is what makes the port-zero no-I/O guarantee testable at all.
+
+#![cfg(unix)]
+
+use std::io::{BufRead, BufReader, Read, Write};
+use std::os::unix::fs::PermissionsExt;
+use std::path::{Path, PathBuf};
+use std::process::{Child, Command, Stdio};
+use std::time::{Duration, Instant};
+
+/// Wall clock, never an iteration count: a fixed loop count starves on a slow
+/// or loaded machine and turns a real pass into a flake.
+const DEADLINE: Duration = Duration::from_secs(15);
+
+/// Kills the child on every exit path, including a panicking assertion.
+///
+/// Without this a failing row leaves a node behind that retries its database
+/// connection forever, holding the row's tempdir open. The first run of this
+/// file did exactly that, and the leak is invisible until something else in
+/// the suite gets slow.
+struct ChildGuard(Child);
+
+impl ChildGuard {
+    /// The tracing subscriber writes to STDOUT; only anyhow's final error
+    /// print goes to stderr. A row that watches one stream sees half the
+    /// story, so both are read.
+    fn take_stdout(&mut self) -> std::process::ChildStdout {
+        self.0.stdout.take().expect("stdout piped")
+    }
+
+    fn take_stderr(&mut self) -> std::process::ChildStderr {
+        self.0.stderr.take().expect("stderr piped")
+    }
+}
+
+impl Drop for ChildGuard {
+    fn drop(&mut self) {
+        let _ = self.0.kill();
+        let _ = self.0.wait();
+    }
+}
+
+/// `(path, is_dir, is_symlink, mode, len)` for everything under `root`, sorted.
+/// Taken before and after each row so a refusal that creates, chmods, or
+/// deletes anything fails the row.
+fn snapshot(root: &Path) -> Vec<String> {
+    fn walk(dir: &Path, out: &mut Vec<String>) {
+        let Ok(entries) = std::fs::read_dir(dir) else {
+            return;
+        };
+        for e in entries.flatten() {
+            let p = e.path();
+            let Ok(md) = std::fs::symlink_metadata(&p) else {
+                continue;
+            };
+            out.push(format!(
+                "{} dir={} link={} mode={:04o} len={}",
+                p.display(),
+                md.is_dir(),
+                md.is_symlink(),
+                md.permissions().mode() & 0o7777,
+                if md.is_file() { md.len() } else { 0 }
+            ));
+            if md.is_dir() && !md.is_symlink() {
+                walk(&p, out);
+            }
+        }
+    }
+    let mut out = Vec::new();
+    walk(root, &mut out);
+    out.sort();
+    out
+}
+
+struct Row {
+    /// Sandbox the row owns: HOME, repos, cwd and the key tree all live here.
+    home: tempfile::TempDir,
+}
+
+impl Row {
+    /// The only directory the no-mutation assertions watch.
+    fn tree(&self) -> PathBuf {
+        self.home.path().join("keytree")
+    }
+}
+
+impl Row {
+    fn new() -> Row {
+        let home = tempfile::tempdir().expect("tempdir");
+        std::fs::set_permissions(home.path(), std::fs::Permissions::from_mode(0o700))
+            .expect("chmod home");
+        // Created here, before any row takes its `before` snapshot, so the
+        // no-mutation assertions measure what the NODE did and not what this
+        // harness set up.
+        std::fs::create_dir_all(home.path().join("repos")).expect("repos dir");
+        // The p2p key tree lives in its own subdirectory so the no-mutation
+        // assertions can watch exactly it. HOME also holds `.gitlawb`, which
+        // the node creates for its NODE identity on every start; measuring the
+        // whole home would read that legitimate write as a p2p storage
+        // mutation.
+        let tree = home.path().join("keytree");
+        std::fs::create_dir(&tree).expect("key tree");
+        std::fs::set_permissions(&tree, std::fs::Permissions::from_mode(0o700))
+            .expect("chmod key tree");
+        Row { home }
+    }
+
+    fn spawn(&self, p2p_key: &str, p2p_port: &str, cwd: &Path) -> ChildGuard {
+        self.spawn_env(p2p_key, p2p_port, cwd, &[])
+    }
+
+    fn spawn_env(
+        &self,
+        p2p_key: &str,
+        p2p_port: &str,
+        cwd: &Path,
+        extra: &[(&str, &str)],
+    ) -> ChildGuard {
+        let repos = self.home.path().join("repos");
+        let mut cmd = Command::new(env!("CARGO_BIN_EXE_gitlawb-node"));
+        cmd.env_clear()
+            .env("PATH", std::env::var("PATH").unwrap_or_default())
+            // The subscriber is built from the default env filter, so with a
+            // cleared environment the node logs nothing at all and every
+            // assertion here reads as a timeout with empty output. The rows
+            // are about which log lines appear, so the filter is part of the
+            // fixture, not incidental setup.
+            .env("RUST_LOG", "info")
+            .env("HOME", self.home.path())
+            .env("GITLAWB_REPOS_DIR", &repos)
+            .env("GITLAWB_HOST", "127.0.0.1")
+            // Ephemeral: the real port is read back from the ready log.
+            .env("GITLAWB_PORT", "0")
+            // A closed port, so the node stays on the degraded path for the
+            // whole row instead of ever reaching a database.
+            .env("DATABASE_URL", "postgres://127.0.0.1:1/nonexistent")
+            .env("GITLAWB_P2P_PORT", p2p_port)
+            .env("GITLAWB_P2P_KEY", p2p_key)
+            .env("GITLAWB_METRICS_ADDR", "")
+            .current_dir(cwd)
+            .stdout(Stdio::piped())
+            .stderr(Stdio::piped());
+        for (k, v) in extra {
+            cmd.env(k, v);
+        }
+        cmd.spawn().map(ChildGuard).expect("spawn gitlawb-node")
+    }
+}
+
+/// Read the child's stderr until `needle` appears, the stream ends, or the
+/// deadline passes. Returns everything read.
+///
+/// The read runs on its own thread feeding a channel, and the deadline is
+/// enforced with `recv_timeout`. Checking elapsed time between blocking
+/// `read_line` calls does NOT work: a child that goes quiet blocks the read
+/// forever and the check never runs, so the deadline looks present and cannot
+/// fire. The first version of this file did that and hung for minutes on a row
+/// that was supposed to fail in fifteen seconds.
+fn read_until(child: &mut ChildGuard, needle: &str) -> (bool, String) {
+    read_until_all(child, &[needle])
+}
+
+/// Read until EVERY needle has appeared, the stream ends, or the deadline
+/// passes.
+///
+/// Waiting on a single line is a race here: the degraded server logs "ready"
+/// from a spawned task while the p2p gate logs its verdict from the main task,
+/// so whichever the reader stops at first can leave the other uncaptured. That
+/// made this file pass or fail depending on scheduling.
+fn read_until_all(child: &mut ChildGuard, needles: &[&str]) -> (bool, String) {
+    let stdout = child.take_stdout();
+    let (tx, rx) = std::sync::mpsc::channel::<String>();
+    std::thread::spawn(move || {
+        let mut reader = BufReader::new(stdout);
+        loop {
+            let mut line = String::new();
+            match reader.read_line(&mut line) {
+                Ok(0) | Err(_) => break,
+                Ok(_) => {
+                    if tx.send(line).is_err() {
+                        break;
+                    }
+                }
+            }
+        }
+    });
+
+    let mut acc = String::new();
+    let start = Instant::now();
+    loop {
+        if needles.iter().all(|n| acc.contains(n)) {
+            return (true, acc);
+        }
+        let left = match DEADLINE.checked_sub(start.elapsed()) {
+            Some(d) if !d.is_zero() => d,
+            _ => return (false, acc),
+        };
+        match rx.recv_timeout(left) {
+            Ok(line) => acc.push_str(&line),
+            // Sender gone: the stream ended. One last check, then give up.
+            Err(std::sync::mpsc::RecvTimeoutError::Disconnected) => {
+                return (needles.iter().all(|n| acc.contains(n)), acc)
+            }
+            Err(std::sync::mpsc::RecvTimeoutError::Timeout) => return (false, acc),
+        }
+    }
+}
+
+fn wait_for_exit(child: &mut ChildGuard) -> (std::process::ExitStatus, String) {
+    // Both streams: the refusal reason is anyhow's stderr print, while
+    // "binding HTTP listener" is a tracing line on stdout, and the fatal rows
+    // assert on one of each.
+    let mut out = String::new();
+    let mut o = child.take_stdout();
+    let mut e = child.take_stderr();
+    let reader = std::thread::spawn(move || {
+        let mut buf = String::new();
+        let _ = o.read_to_string(&mut buf);
+        buf
+    });
+    let _ = e.read_to_string(&mut out);
+    let status = child.0.wait().expect("wait");
+    out.push_str(&reader.join().unwrap_or_default());
+    (status, out)
+}
+
+/// Strip ANSI escapes: the subscriber colourises even when piped, which would
+/// otherwise sit between `addr=` and the value.
+fn strip_ansi(s: &str) -> String {
+    let mut out = String::with_capacity(s.len());
+    let mut chars = s.chars();
+    while let Some(c) = chars.next() {
+        if c == '\u{1b}' {
+            for c2 in chars.by_ref() {
+                if c2.is_ascii_alphabetic() {
+                    break;
+                }
+            }
+        } else {
+            out.push(c);
+        }
+    }
+    out
+}
+
+/// A lexically invalid value must stop the process before the listener binds.
+fn assert_fatal_before_bind(row: &Row, key: &str, cwd: &Path) {
+    let before = snapshot(&row.tree());
+    let mut child = row.spawn(key, "7546", cwd);
+    let (status, err) = wait_for_exit(&mut child);
+
+    assert!(
+        !status.success(),
+        "key={key:?} must exit non-zero\n--- stderr ---\n{err}"
+    );
+    assert!(
+        err.contains("invalid configuration"),
+        "key={key:?} must fail as invalid configuration\n--- stderr ---\n{err}"
+    );
+    assert!(
+        !err.contains("binding HTTP listener"),
+        "key={key:?} must be refused BEFORE the listener binds\n--- stderr ---\n{err}"
+    );
+    assert_eq!(
+        snapshot(&row.tree()),
+        before,
+        "key={key:?}: a configuration refusal must not touch the key tree"
+    );
+}
+
+/// A live storage fault must leave HTTP serving with p2p off, and must not
+/// mutate the tree on its way to that verdict.
+fn assert_degrades_with_http_up(row: &Row, key: &str, cwd: &Path, label: &str) {
+    let before = snapshot(&row.tree());
+    let mut child = row.spawn(key, "7546", cwd);
+    let (found, log) = read_until_all(
+        &mut child,
+        &[
+            "degraded HTTP server ready",
+            "failed to load p2p identity key",
+        ],
+    );
+    assert!(
+        found,
+        "{label}: the node must bind and serve while p2p is off\n--- stderr ---\n{log}"
+    );
+    // The operator docs promise a specific signal, so pin it here: the level,
+    // the stable event name, and the message. A degrade that logs at warn or
+    // renames the event silently breaks the alert those docs tell operators to
+    // build, and asserting only the prose would not notice.
+    let clean_log = strip_ansi(&log);
+    let signal = clean_log
+        .lines()
+        .find(|l| l.contains("failed to load p2p identity key"))
+        .unwrap_or_else(|| panic!("{label}: the p2p failure must be logged\n{clean_log}"));
+    assert!(
+        signal.contains("ERROR"),
+        "{label}: the p2p failure must be logged at error level, got: {signal}"
+    );
+    assert!(
+        signal.contains("p2p_identity_key_load_failed"),
+        "{label}: the p2p failure must carry the documented event name, got: {signal}"
+    );
+
+    let clean = strip_ansi(&log);
+    let addr = clean
+        .lines()
+        .find(|l| l.contains("degraded HTTP server ready"))
+        .and_then(|l| l.rsplit("addr=").next())
+        .map(|a| {
+            a.trim()
+                .trim_start_matches("Some(")
+                .trim_matches(|c| c == '"' || c == ')')
+        })
+        .and_then(|a| a.parse::<std::net::SocketAddr>().ok())
+        .unwrap_or_else(|| panic!("{label}: could not parse the bound address from {clean}"));
+
+    // A served response, any status: the degraded server answers 503, which is
+    // still proof the port is up and answering.
+    let mut sock = std::net::TcpStream::connect(addr).expect("connect to the degraded server");
+    sock.set_read_timeout(Some(DEADLINE)).unwrap();
+    sock.write_all(b"GET /health HTTP/1.1\r\nHost: localhost\r\nConnection: close\r\n\r\n")
+        .expect("send request");
+    let mut resp = String::new();
+    let _ = sock.read_to_string(&mut resp);
+    assert!(
+        resp.starts_with("HTTP/1.1 "),
+        "{label}: the degraded server must answer, got {resp:?}"
+    );
+
+    drop(child);
+    assert_eq!(
+        snapshot(&row.tree()),
+        before,
+        "{label}: a storage refusal must not mutate the key tree"
+    );
+}
+
+#[test]
+fn lexically_invalid_key_paths_stop_the_node_before_it_binds() {
+    for key in [
+        "p2p.key",
+        "./p2p.key",
+        "a/../p2p.key",
+        "../p2p.key",
+        "/p2p.key",
+        "keys/",
+        "~/",
+        "~//etc/p2p.key",
+        "~/../x/p2p.key",
+    ] {
+        let row = Row::new();
+        let cwd = row.home.path().to_path_buf();
+        assert_fatal_before_bind(&row, key, &cwd);
+    }
+}
+
+#[test]
+fn live_storage_faults_leave_http_up_with_p2p_off() {
+    // Symlinked final parent.
+    {
+        let row = Row::new();
+        let real = row.tree().join("real");
+        std::fs::create_dir(&real).unwrap();
+        std::fs::set_permissions(&real, std::fs::Permissions::from_mode(0o700)).unwrap();
+        let link = row.tree().join("keys");
+        std::os::unix::fs::symlink(&real, &link).unwrap();
+        let key = link.join("p2p.key");
+        let cwd = row.home.path().to_path_buf();
+        assert_degrades_with_http_up(&row, key.to_str().unwrap(), &cwd, "symlinked parent");
+    }
+
+    // Regular file where the key directory should be.
+    {
+        let row = Row::new();
+        let file = row.tree().join("keys");
+        std::fs::write(&file, b"not a directory").unwrap();
+        let key = file.join("p2p.key");
+        let cwd = row.home.path().to_path_buf();
+        assert_degrades_with_http_up(&row, key.to_str().unwrap(), &cwd, "regular-file parent");
+    }
+
+    // Unreadable parent: an inspection error, the purest instance of a live
+    // fault that used to be reported as invalid configuration.
+    {
+        let row = Row::new();
+        let parent = row.tree().join("locked");
+        std::fs::create_dir(&parent).unwrap();
+        std::fs::set_permissions(&parent, std::fs::Permissions::from_mode(0o000)).unwrap();
+        let key = parent.join("keys").join("p2p.key");
+        let cwd = row.home.path().to_path_buf();
+        assert_degrades_with_http_up(&row, key.to_str().unwrap(), &cwd, "unreadable parent");
+        // Restore so the tempdir can be cleaned up.
+        let _ = std::fs::set_permissions(&parent, std::fs::Permissions::from_mode(0o700));
+    }
+
+    // World-writable non-sticky ancestor: already the degrade class today, kept
+    // so the row set covers both sides of the policy.
+    {
+        let row = Row::new();
+        let anc = row.tree().join("open");
+        std::fs::create_dir(&anc).unwrap();
+        std::fs::set_permissions(&anc, std::fs::Permissions::from_mode(0o777)).unwrap();
+        let key = anc.join("keys").join("p2p.key");
+        let cwd = row.home.path().to_path_buf();
+        assert_degrades_with_http_up(&row, key.to_str().unwrap(), &cwd, "world-writable ancestor");
+    }
+}
+
+#[test]
+fn port_zero_does_no_key_storage_io_at_all() {
+    // A lexically invalid path that would be fatal with p2p enabled.
+    {
+        let row = Row::new();
+        let before = snapshot(&row.tree());
+        let cwd = row.home.path().to_path_buf();
+        let mut child = row.spawn("p2p.key", "0", &cwd);
+        let (found, log) = read_until(&mut child, "p2p disabled");
+        assert!(
+            found,
+            "port zero must bypass key handling entirely\n--- stderr ---\n{log}"
+        );
+        drop(child);
+        assert_eq!(
+            snapshot(&row.tree()),
+            before,
+            "port zero must not touch the key tree"
+        );
+    }
+
+    // A hostile tree that would degrade with p2p enabled.
+    {
+        let row = Row::new();
+        let real = row.tree().join("real");
+        std::fs::create_dir(&real).unwrap();
+        let link = row.tree().join("keys");
+        std::os::unix::fs::symlink(&real, &link).unwrap();
+        let before = snapshot(&row.tree());
+        let cwd = row.home.path().to_path_buf();
+        let key: PathBuf = link.join("p2p.key");
+        let mut child = row.spawn(key.to_str().unwrap(), "0", &cwd);
+        let (found, log) = read_until(&mut child, "p2p disabled");
+        assert!(
+            found,
+            "port zero must bypass a hostile tree too\n--- stderr ---\n{log}"
+        );
+        drop(child);
+        assert_eq!(
+            snapshot(&row.tree()),
+            before,
+            "port zero must not touch a hostile key tree"
+        );
+    }
+}
+
+/// A terminal `.` is a directory-valued spelling. Path drops it, so the load
+/// path would otherwise chmod the parent and publish a file named `keys`.
+#[test]
+fn terminal_dot_key_path_is_fatal_and_does_not_mutate() {
+    // Absent leaf: `/data/keys/.` must not create `keys` as a file or chmod `/data`.
+    {
+        let row = Row::new();
+        let data = row.tree().join("data");
+        std::fs::create_dir(&data).unwrap();
+        std::fs::set_permissions(&data, std::fs::Permissions::from_mode(0o755)).unwrap();
+        let key = format!("{}/keys/.", data.display());
+        let cwd = row.home.path().to_path_buf();
+        assert_fatal_before_bind(&row, &key, &cwd);
+        let mode = std::fs::metadata(&data).unwrap().permissions().mode() & 0o7777;
+        assert_eq!(mode, 0o755, "rejection must not chmod the parent");
+        assert!(
+            !data.join("keys").exists(),
+            "rejection must not create a file named keys"
+        );
+    }
+
+    // Existing directory at the would-be leaf: still boot-fatal, directory untouched.
+    {
+        let row = Row::new();
+        let data = row.tree().join("data");
+        let keys = data.join("keys");
+        std::fs::create_dir_all(&keys).unwrap();
+        std::fs::set_permissions(&data, std::fs::Permissions::from_mode(0o755)).unwrap();
+        std::fs::set_permissions(&keys, std::fs::Permissions::from_mode(0o700)).unwrap();
+        let key = format!("{}/.", keys.display());
+        let cwd = row.home.path().to_path_buf();
+        assert_fatal_before_bind(&row, &key, &cwd);
+        assert!(
+            keys.is_dir(),
+            "the existing keys directory must remain a directory"
+        );
+        let data_mode = std::fs::metadata(&data).unwrap().permissions().mode() & 0o7777;
+        let keys_mode = std::fs::metadata(&keys).unwrap().permissions().mode() & 0o7777;
+        assert_eq!(data_mode, 0o755, "rejection must not chmod the parent");
+        assert_eq!(
+            keys_mode, 0o700,
+            "rejection must not chmod the existing keys dir"
+        );
+    }
+}
+
+/// GITLAWB_KEY has the same Path-retarget hole as GITLAWB_P2P_KEY: a terminal
+/// `.` would publish a file named `keys` under `data`. Fatal before bind, and
+/// the parent tree is unchanged. p2p is off so it cannot chmod a shared dir.
+#[test]
+fn identity_terminal_dot_is_fatal_and_does_not_mutate() {
+    let row = Row::new();
+    let data = row.home.path().join("data");
+    std::fs::create_dir(&data).unwrap();
+    std::fs::set_permissions(&data, std::fs::Permissions::from_mode(0o755)).unwrap();
+    let key = format!("{}/keys/.", data.display());
+    let cwd = row.home.path().to_path_buf();
+    let mut child = row.spawn_env("p2p.key", "0", &cwd, &[("GITLAWB_KEY", key.as_str())]);
+    let (status, err) = wait_for_exit(&mut child);
+    assert!(
+        !status.success(),
+        "GITLAWB_KEY={key:?} must exit non-zero\n--- stderr ---\n{err}"
+    );
+    assert!(
+        err.contains("invalid configuration"),
+        "GITLAWB_KEY={key:?} must fail as invalid configuration\n--- stderr ---\n{err}"
+    );
+    assert!(
+        !err.contains("binding HTTP listener"),
+        "GITLAWB_KEY={key:?} must be refused BEFORE the listener binds\n--- stderr ---\n{err}"
+    );
+    let mode = std::fs::metadata(&data).unwrap().permissions().mode() & 0o7777;
+    assert_eq!(mode, 0o755, "rejection must not chmod the parent");
+    assert!(
+        !data.join("keys").exists(),
+        "rejection must not create a file named keys"
+    );
+}
+
+/// An existing 0755 identity parent is usable and must not be chmodded. p2p is
+/// off so a default `~/.gitlawb/p2p.key` cannot tighten the same directory.
+#[test]
+fn identity_existing_0755_parent_is_not_chmodded_on_boot() {
+    let row = Row::new();
+    let iddir = row.home.path().join("idparent");
+    std::fs::create_dir(&iddir).unwrap();
+    std::fs::set_permissions(&iddir, std::fs::Permissions::from_mode(0o755)).unwrap();
+    let key = iddir.join("identity.pem");
+    let cwd = row.home.path().to_path_buf();
+    let mut child = row.spawn_env(
+        "p2p.key",
+        "0",
+        &cwd,
+        &[("GITLAWB_KEY", key.to_str().unwrap())],
+    );
+    let (found, log) = read_until(&mut child, "degraded HTTP server ready");
+    assert!(
+        found,
+        "a 0755 identity parent must still boot\n--- stderr ---\n{log}"
+    );
+    drop(child);
+    let mode = std::fs::metadata(&iddir).unwrap().permissions().mode() & 0o7777;
+    assert_eq!(
+        mode, 0o755,
+        "boot must not chmod the existing identity parent"
+    );
+    assert_eq!(
+        std::fs::metadata(&key).unwrap().permissions().mode() & 0o7777,
+        0o600,
+        "the identity key must be owner-only"
+    );
+}
+
+/// Creating into a world-writable identity parent is refused before bind, and
+/// the directory is left untouched.
+#[test]
+fn identity_writable_parent_is_fatal_and_unchanged() {
+    let row = Row::new();
+    let iddir = row.home.path().join("idparent");
+    std::fs::create_dir(&iddir).unwrap();
+    std::fs::set_permissions(&iddir, std::fs::Permissions::from_mode(0o777)).unwrap();
+    let key = iddir.join("identity.pem");
+    let cwd = row.home.path().to_path_buf();
+    let mut child = row.spawn_env(
+        "p2p.key",
+        "0",
+        &cwd,
+        &[("GITLAWB_KEY", key.to_str().unwrap())],
+    );
+    let (status, err) = wait_for_exit(&mut child);
+    assert!(
+        !status.success(),
+        "a writable identity parent must exit non-zero\n--- stderr ---\n{err}"
+    );
+    assert!(
+        !err.contains("binding HTTP listener"),
+        "a writable identity parent must be refused BEFORE the listener binds\n--- stderr ---\n{err}"
+    );
+    let mode = std::fs::metadata(&iddir).unwrap().permissions().mode() & 0o7777;
+    assert_eq!(mode, 0o777, "refusal must not chmod the parent");
+    assert!(!key.exists(), "refusal must not publish the identity key");
+}
+
+/// A 2700 (setgid, owner rwx) p2p key directory is repairable: boot, tighten
+/// to 0700, HTTP still comes up.
+#[test]
+fn p2p_setgid_key_directory_is_tightened_and_http_comes_up() {
+    let row = Row::new();
+    let keys = row.tree().join("keys");
+    std::fs::create_dir(&keys).unwrap();
+    std::fs::set_permissions(&keys, std::fs::Permissions::from_mode(0o2700)).unwrap();
+    let key = keys.join("p2p.key");
+    let cwd = row.home.path().to_path_buf();
+    let mut child = row.spawn(key.to_str().unwrap(), "7546", &cwd);
+    let (found, log) = read_until_all(
+        &mut child,
+        &["degraded HTTP server ready", "generated new p2p identity"],
+    );
+    assert!(
+        found,
+        "a 2700 key directory must boot and generate a key\n--- stderr ---\n{log}"
+    );
+    drop(child);
+    let mode = std::fs::metadata(&keys).unwrap().permissions().mode() & 0o7777;
+    assert_eq!(
+        mode, 0o700,
+        "2700 must be normalized to 0700, found {mode:04o}"
+    );
+}
@@ -1,6 +1,6 @@
 [package]
 name = "gl"
-description = "The gitlawb CLI — identity management, node control, MCP server"
+description = "The Twigpine CLI — identity management, node control, MCP server"
 version = "0.7.1" # x-release-please-version
 edition.workspace = true
 rust-version.workspace = true
@@ -11,7 +11,7 @@ name = "gl"
 path = "src/main.rs"

 [dependencies]
-gitlawb-core = { path = "../gitlawb-core" }
+gitlawb-core = { path = "../gitlawb-core", features = ["redirect"] }
 icaptcha-client = { path = "../icaptcha-client" }
 base64 = { workspace = true }
 tokio = { workspace = true }
@@ -26,7 +26,10 @@ clap = { version = "4", features = ["derive", "env"] }
 dirs = "5"
 reqwest = { workspace = true }
 uuid = { workspace = true }
+url = { workspace = true }
+percent-encoding = { workspace = true }
 urlencoding = "2"
+libc = "0.2"
 alloy = { version = "1", default-features = false, features = [
    "contract",
    "provider-http",
@@ -42,3 +45,4 @@ alloy = { version = "1", default-features = false, features = [
 mockito = "1"
 tempfile = "3"
 tokio = { workspace = true }
+ed25519-dalek = { workspace = true }
@@ -2,7 +2,6 @@

 use anyhow::{Context, Result};
 use clap::{Args, Subcommand};
-use serde_json::Value;
 use std::path::PathBuf;

 use crate::http::NodeClient;
@@ -17,7 +16,11 @@ pub struct AgentArgs {
 pub enum AgentCmd {
    /// List agents registered on a node
    List {
-        #[arg(long, default_value = "https://node.gitlawb.com", env = "GITLAWB_NODE")]
+        #[arg(
+            long,
+            default_value = "https://node.twigpine.com",
+            env = "GITLAWB_NODE"
+        )]
        node: String,
        /// Filter by capability (e.g. git:push)
        #[arg(long)]
@@ -29,7 +32,11 @@ pub enum AgentCmd {
    Show {
        /// Agent DID or short key segment
        did: String,
-        #[arg(long, default_value = "https://node.gitlawb.com", env = "GITLAWB_NODE")]
+        #[arg(
+            long,
+            default_value = "https://node.twigpine.com",
+            env = "GITLAWB_NODE"
+        )]
        node: String,
    },
 }
@@ -57,24 +64,16 @@ async fn cmd_list(node: String, capability: Option<String>) -> Result<()> {
        .get(&path)
        .await
        .context("failed to connect to node")?;
-    let status = resp.status();
-
-    if status == reqwest::StatusCode::NOT_FOUND {
+    // Keep the bespoke 404 hint (older node without the agents API) ahead of
+    // read_json; every other status routes through it (capped + sanitized error).
+    if resp.status() == reqwest::StatusCode::NOT_FOUND {
        anyhow::bail!(
            "this node does not yet support the agents API (v0.3+)\n\
             upgrade the node or check GITLAWB_NODE is pointing to the right server"
        );
    }

-    let body: Value = resp
-        .json()
-        .await
-        .context("invalid JSON from node — is GITLAWB_NODE correct?")?;
-
-    if !status.is_success() {
-        let msg = body["message"].as_str().unwrap_or("unknown error");
-        anyhow::bail!("list agents failed ({status}): {msg}");
-    }
+    let body = crate::http::read_json(resp, "list agents").await?;

    let agents = body["agents"].as_array().cloned().unwrap_or_default();

@@ -89,7 +88,7 @@ async fn cmd_list(node: String, capability: Option<String>) -> Result<()> {
        let short = did
            .split(':')
            .next_back()
-            .map(|s| &s[..s.len().min(16)])
+            .map(|s| crate::text::truncate(s, 16))
            .unwrap_or("?");
        let trust = agent["trust_score"].as_f64().unwrap_or(0.0);
        let caps = agent["capabilities"]
@@ -117,24 +116,15 @@ async fn cmd_show(did: String, node: String) -> Result<()> {
        .get(&format!("/api/v1/agents/{did}"))
        .await
        .context("failed to connect to node")?;
-    let status = resp.status();
-
-    if status == reqwest::StatusCode::NOT_FOUND {
+    // Keep the bespoke 404 hint ahead of read_json; other statuses route through it.
+    if resp.status() == reqwest::StatusCode::NOT_FOUND {
        anyhow::bail!(
            "agent not found, or this node does not yet support the agents API (v0.3+)\n\
             upgrade the node or check GITLAWB_NODE is pointing to the right server"
        );
    }

-    let agent: Value = resp
-        .json()
-        .await
-        .context("invalid JSON from node — is GITLAWB_NODE correct?")?;
-
-    if !status.is_success() {
-        let msg = agent["message"].as_str().unwrap_or("unknown error");
-        anyhow::bail!("agent not found ({status}): {msg}");
-    }
+    let agent = crate::http::read_json(resp, "agent").await?;

    let full_did = agent["did"].as_str().unwrap_or("?");
    let trust = agent["trust_score"].as_f64().unwrap_or(0.0);
@@ -215,11 +205,14 @@ mod tests {
            .with_status(404)
            .with_header("content-type", "application/json")
            .with_body(r#"{"message":"not found"}"#)
+            .expect(1)
            .create_async()
            .await;

        let err = cmd_list(server.url(), None).await.unwrap_err();
        assert!(err.to_string().contains("agents API"));
+        // Prove the mocked route was actually requested; a non-matching request (mockito's 501, also non-2xx) would otherwise satisfy the error assertion vacuously.
+        _m.assert_async().await;
    }

    #[tokio::test]
@@ -230,11 +223,14 @@ mod tests {
            .with_status(500)
            .with_header("content-type", "application/json")
            .with_body(r#"{"message":"internal error"}"#)
+            .expect(1)
            .create_async()
            .await;

        let err = cmd_list(server.url(), None).await.unwrap_err();
        assert!(err.to_string().contains("list agents failed"));
+        // Prove the mocked route was actually requested; a non-matching request (mockito's 501, also non-2xx) would otherwise satisfy the error assertion vacuously.
+        _m.assert_async().await;
    }

    #[tokio::test]
@@ -261,6 +257,7 @@ mod tests {
            .with_status(404)
            .with_header("content-type", "application/json")
            .with_body(r#"{"message":"agent not found"}"#)
+            .expect(1)
            .create_async()
            .await;

@@ -268,6 +265,8 @@ mod tests {
            .await
            .unwrap_err();
        assert!(err.to_string().contains("agents API") || err.to_string().contains("not found"));
+        // Prove the mocked route was actually requested; a non-matching request (mockito's 501, also non-2xx) would otherwise satisfy the error assertion vacuously.
+        _m.assert_async().await;
    }

    #[tokio::test]