fix(studio): a thought the reader set to Open is not capped to the live six-line window

The cap keyed on data-live alone, so Open and Open-while-running looked the same while the model thought. SayCard now marks data-full when the thinking mode is open and the cap skips it.
fix(studio): the pin button hides until its row is hovered, a pinned row keeps a marker, and the model menu's image badge no longer wraps
2026-11-07 10:45:59 -10:00 · 2026-11-07 10:15:47 -08:00 · 2026-11-07 09:19:30 -08:00 · 2026-11-07 08:25:49 -08:00 · 2026-11-07 07:11:37 -08:00 · 2026-11-07 07:11:24 -08:00
2163 changed files with 44132 additions and 111985 deletions
@@ -0,0 +1,7 @@
+# Logins allowed to open pull requests, one per line, matched case-insensitively.
+# Repository owners, members and collaborators, and bots, are always allowed.
+lancy69
+chuiwenwei
+xxoingr
+kevin9327
+imcreazy
@@ -2,6 +2,9 @@
 Target branch: 2.x features and fixes go to `studio`. 1.x takes bug fixes,
 provider/API compatibility, release/updater, security and platform stability
 only, on `main-v2`.
+
+Pull requests are limited to maintainers and invited contributors; one from
+anyone else is closed automatically. Open an issue instead.
 -->

 ## Summary
@@ -1,24 +0,0 @@
-[
-  {
-    "package": "reasonix/internal/runtime/delegation",
-    "tests": [
-      "TestAncestorCopyAdmissionCoversUnpublishedRun",
-      "TestAncestorCopyAdmissionCoversUnpublishedRun/continue_then_continue",
-      "TestAncestorCopyAdmissionCoversUnpublishedRun/continue_then_legacy",
-      "TestAncestorCopyAdmissionCoversUnpublishedRun/legacy_then_continue",
-      "TestAncestorCopyAdmissionCoversUnpublishedRun/legacy_then_legacy",
-      "TestAncestorCopyAdmissionAllowsDifferentParents",
-      "TestAncestorCopyAdmissionReleasesAbandonedPreparation",
-      "TestAncestorCopyAdmissionReleasesAfterInvalidCopy",
-      "TestAncestorCopyAdmissionAllowsDifferentSources",
-      "TestAncestorCopyAdmissionUsesNearestSourceIdentity",
-      "TestAncestorCopyAdmissionReleasesAfterBusyCurrentReference",
-      "TestTaskAncestorCopyAdmissionBeforeRunningMetadata",
-      "TestTaskAncestorCopyAdmissionBeforeRunningMetadata/continue_from",
-      "TestTaskAncestorCopyAdmissionBeforeRunningMetadata/fork_from",
-      "TestTaskAncestorCopyAdmissionReleasesUnstartedBackground",
-      "TestTaskAncestorCopyAdmissionReleasesUnstartedBackground/continue_from",
-      "TestTaskAncestorCopyAdmissionReleasesUnstartedBackground/fork_from"
-    ]
-  }
-]
@@ -1,26 +0,0 @@
-[
-  {
-    "package": "reasonix/internal/runtime/agent",
-    "tests": [
-      "TestCarriedProofInvalidatesPermissionChange",
-      "TestCarriedProofFileStateCompatibility",
-      "TestCarriedProofFileStateCompatibility/legacy_file",
-      "TestCarriedProofFileStateCompatibility/legacy_absent",
-      "TestCarriedProofFileStateCompatibility/future",
-      "TestCarriedProofFileStateCompatibility/current_file",
-      "TestCarriedProofFileStateCompatibility/current_absent",
-      "TestCarriedProofRejectsDirectoryState",
-      "TestCarriedProofStandsOnlyForUnchangedContent",
-      "TestCarriedProofSeesADeletion"
-    ]
-  },
-  {
-    "package": "reasonix/internal/assembly/boot",
-    "tests": [
-      "TestEffectCarriedProofTracksFileState",
-      "TestEffectCarriedProofTracksFileState/unchanged",
-      "TestEffectCarriedProofTracksFileState/content",
-      "TestEffectCarriedProofTracksFileState/mode"
-    ]
-  }
-]
@@ -15,7 +15,7 @@ permissions:

 concurrency:
  group: ci-${{ github.ref }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' || github.ref == 'refs/heads/studio' }}

 jobs:
  # Cheap path gate for pull requests. PRs confined to docs/release-notes
@@ -170,23 +170,6 @@ jobs:
        shell: pwsh
        run: Get-CimInstance Win32_Processor | Select-Object -ExpandProperty Name

-      - name: test (Windows concurrency contracts)
-        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        timeout-minutes: 5
-        shell: bash
-        run: |
-          node --test scripts/windows-concurrency-tests.test.mjs
-          node scripts/windows-concurrency-tests.mjs "$RUNNER_TEMP/concurrency-windows.jsonl"
-
-      - name: upload Windows concurrency inventory
-        if: always() && env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7
-        with:
-          name: concurrency-windows-${{ github.run_attempt }}
-          path: ${{ runner.temp }}/concurrency-windows.jsonl
-          if-no-files-found: error
-          retention-days: 7
-
      - name: test
        if: env.RUN_STEPS == 'true' && runner.os != 'Windows'
        env:
@@ -196,35 +179,6 @@ jobs:
          REASONIX_RELEASE_CACHE_GUARD: "1"
        run: go test ./...

-      - name: test (Windows subagent status admission)
-        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        timeout-minutes: 5
-        shell: bash
-        run: |
-          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^TestSubagent(PersistedStatusAdmission|UnknownStatusStillReadableForDiagnosis)$' ./internal/runtime/delegation | tee "$RUNNER_TEMP/subagent-status-admission.jsonl"
-          jq -se '
-            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
-            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
-            ["future-paused", "missing", "empty", "COMPLETED", "completed", "cancelled", "failed", "interrupted", "running"] as $states |
-            ["continue", "copy", "ancestor", "legacy_ancestor", "current_copy"] as $routes |
-            (["TestSubagentPersistedStatusAdmission", "TestSubagentUnknownStatusStillReadableForDiagnosis"] +
-              [$states[] | "TestSubagentPersistedStatusAdmission/" + .] +
-              [$states[] as $state | $routes[] | "TestSubagentPersistedStatusAdmission/" + $state + "/" + .]) as $required |
-            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/runtime/delegation"] - $packages | length) == 0 and
-            ($required - $tests | length) == 0
-          ' "$RUNNER_TEMP/subagent-status-admission.jsonl"
-
-      - name: upload Windows subagent status admission inventory
-        if: always() && env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7
-        with:
-          name: subagent-status-admission-windows-${{ github.run_attempt }}
-          path: ${{ runner.temp }}/subagent-status-admission.jsonl
-          if-no-files-found: ignore
-          retention-days: 7
-
      - name: test (Windows composition links)
        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
        timeout-minutes: 5
@@ -301,40 +255,21 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -p 4 -timeout=18m -json ./internal/runtime/isolation/... ./internal/platform/worktree/... ./internal/state/mutation/... | tee "$RUNNER_TEMP/isolation-owners.jsonl" |
+          go test -count=1 -p 4 -timeout=18m -json ./internal/platform/worktree/... ./internal/state/mutation/... | tee "$RUNNER_TEMP/isolation-owners.jsonl" |
            jq -rj --unbuffered 'select(.Action == "output" and .Test == null) | .Output'
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $passed |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
-            (["reasonix/internal/runtime/isolation", "reasonix/internal/platform/worktree", "reasonix/internal/state/mutation"] - $passed | length) == 0 and
-            (["TestResultOperationUsesCrossProcessOwnership",
-              "TestActualResultOperationHoldsOwnershipThroughWriteAndCleanup/restore",
-              "TestActualResultOperationHoldsOwnershipThroughWriteAndCleanup/remove",
-              "TestApplyPathsWithUntrustedModesPreservesTheIndex/content",
-              "TestApplyPathsWithUntrustedModesPreservesTheIndex/source_index_mode",
-              "TestApplyPathsWithUntrustedModesPreservesTheIndex/candidate_mode",
-              "TestApplyPathsWithUntrustedModesPreservesTheIndex/candidate_type",
-              "TestApplyPathsEmulatedLinksUseSnapshotFilters",
-              "TestHeldPathsCoverOnlyTheSameFileAcrossCaseSpellings",
+            (["reasonix/internal/platform/worktree", "reasonix/internal/state/mutation"] - $passed | length) == 0 and
+            (["TestHeldPathsCoverOnlyTheSameFileAcrossCaseSpellings",
              "TestLegacyWaitDoesNotConsumeExecutionCapacity"] - $tests | length) == 0
          ' "$RUNNER_TEMP/isolation-owners.jsonl"
          jq -se '
            . as $events |
            {"reasonix/internal/platform/worktree": [
-               "TestApplyPathsRejectsRedirectedAncestors",
-               "TestApplyPathsRejectsRedirectedAncestors/outside_delete",
-               "TestApplyPathsRejectsRedirectedAncestors/inside_delete",
-               "TestApplyPathsRejectsRedirectedAncestors/dangling_delete",
-               "TestApplyPathsRejectsRedirectedAncestors/outside_write",
-               "TestApplyPathsRejectsRedirectedAncestors/inside_write",
-               "TestApplyPathsRejectsRedirectedAncestors/regular_parent",
-               "TestApplyPathsRejectsRedirectedAncestors/unchanged",
-               "TestApplyPathsRejectsRedirectedAncestors/missing_creation",
               "TestWriteChangesRejectsRedirectedDeletion",
-               "TestWriteChangesUnlinksLeafDirectoryLink",
-               "TestWholeWorkspaceApplyPreservesFileToDirectoryTransition"],
-             "reasonix/internal/runtime/isolation": [
-               "TestApplyToolRetainsResultAfterAncestorReplacement"]} |
+               "TestWriteChangesUnlinksLeafDirectoryLink"
+               ]} |
            to_entries | all(.[]; . as $package |
              ([ $events[] | select(.Package == $package.key and .Test == null and
                 (.Action == "pass" or .Action == "fail" or .Action == "skip")) ] |
@@ -354,60 +289,15 @@ jobs:
          if-no-files-found: ignore
          retention-days: 7

-      - name: test (Windows background resource owners)
-        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        timeout-minutes: 25
-        shell: bash
-        run: |
-          set -o pipefail
-          go test -count=1 -p 2 -timeout=18m -json ./internal/tools/jobs/... ./internal/runtime/delegation/... | tee "$RUNNER_TEMP/background-resources.jsonl" |
-            jq -rj --unbuffered 'select(.Action == "output" and .Test == null) | .Output'
-          jq -se '
-            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
-            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
-            (["reasonix/internal/tools/jobs", "reasonix/internal/runtime/delegation"] - $packages | length) == 0 and
-            (["TestJobFinalizationCoversInvokedAndUninvoked/success",
-              "TestJobFinalizationCoversInvokedAndUninvoked/error",
-              "TestJobFinalizationCoversInvokedAndUninvoked/panic",
-              "TestJobFinalizationCoversInvokedAndUninvoked/closed",
-              "TestJobFinalizationCoversInvokedAndUninvoked/invalid_session",
-              "TestJobFinalizationCoversInvokedAndUninvoked/invalid_kind",
-              "TestJobFinalizationFailuresAttemptAllCallbacksAndKeepCauses/invoked",
-              "TestJobFinalizationFailuresAttemptAllCallbacksAndKeepCauses/uninvoked",
-              "TestJobFinalizationPrecedesCompletionPublication",
-              "TestJobFinalizationCloseBetweenPublicationAndLaunch",
-              "TestJobFinalizationWaitsForClosureAfterLogicalCancellation",
-              "TestBackgroundNoLaunchFinalizesOwnedResources/task/closed_manager",
-              "TestBackgroundNoLaunchFinalizesOwnedResources/task/invalid_session",
-              "TestBackgroundNoLaunchFinalizesOwnedResources/fleet/closed_manager",
-              "TestBackgroundNoLaunchFinalizesOwnedResources/fleet/invalid_session"] - $tests | length) == 0
-          ' "$RUNNER_TEMP/background-resources.jsonl"
-
-      - name: upload Windows background resource inventory
-        if: always() && env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7
-        with:
-          name: background-resources-windows-${{ github.run_attempt }}
-          path: ${{ runner.temp }}/background-resources.jsonl
-          if-no-files-found: ignore
-          retention-days: 7
-
      - name: test (Windows delegated Goal usage scope)
        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(UsageScope.*|DelegatedUsageKeepsGoalOwner|EffectDelegatedUsageKeepsGoalOwner)$' ./internal/contract/event ./internal/session/control ./internal/assembly/boot | tee "$RUNNER_TEMP/goal-usage-scope.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(UsageScope.*)$' ./internal/contract/event ./internal/session/control ./internal/assembly/boot | tee "$RUNNER_TEMP/goal-usage-scope.jsonl"
          jq -se '
-            ["TestUsageScopePreservesCapturedOwnerAndWire", "TestUsageScopeKeepsEarlierEventOwner",
-             "TestUsageScopeWithoutCapability", "TestDelegatedUsageKeepsGoalOwner",
-             "TestDelegatedUsageKeepsGoalOwner/chat_to_goal", "TestDelegatedUsageKeepsGoalOwner/replaced_goal",
-             "TestDelegatedUsageKeepsGoalOwner/cleared_goal", "TestDelegatedUsageKeepsGoalOwner/resumed_goal",
-             "TestDelegatedUsageKeepsGoalOwner/same_goal_next_turn", "TestDelegatedUsageKeepsGoalOwner/between_turns",
-             "TestDelegatedUsageKeepsGoalOwner/nested_after_replacement", "TestEffectDelegatedUsageKeepsGoalOwner",
-             "TestEffectDelegatedUsageKeepsGoalOwner/chat_to_goal", "TestEffectDelegatedUsageKeepsGoalOwner/replaced_goal",
-             "TestEffectDelegatedUsageKeepsGoalOwner/nested_chat_to_goal", "TestEffectDelegatedUsageKeepsGoalOwner/same_goal",
-             "TestEffectDelegatedUsageKeepsGoalOwner/completed_before_goal"] as $required |
+            ["TestUsageScopeKeepsEarlierEventOwner",
+             "TestUsageScopeWithoutCapability" ] as $required |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $passed |
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            all(.[]; .Action != "fail" and .Action != "skip") and
@@ -424,60 +314,19 @@ jobs:
          if-no-files-found: error
          retention-days: 7

-      - name: test (Windows planner local budget)
-        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        timeout-minutes: 5
-        shell: bash
-        run: |
-          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^TestEffectPlannerConfiguredTaskBudget$' ./internal/assembly/boot | tee "$RUNNER_TEMP/planner-local-budget.jsonl"
-          jq -se '
-            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
-            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
-            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/assembly/boot"] - $packages | length) == 0 and
-            (["TestEffectPlannerConfiguredTaskBudget", "TestEffectPlannerConfiguredTaskBudget/tokens",
-              "TestEffectPlannerConfiguredTaskBudget/cost", "TestEffectPlannerConfiguredTaskBudget/wall",
-              "TestEffectPlannerConfiguredTaskBudget/unconfigured", "TestEffectPlannerConfiguredTaskBudget/explicit_zero",
-              "TestEffectPlannerConfiguredTaskBudget/unpriced_cost", "TestEffectPlannerConfiguredTaskBudget/missing_usage_cost",
-              "TestEffectPlannerConfiguredTaskBudget/context_bound", "TestEffectPlannerConfiguredTaskBudget/context_disables",
-              "TestEffectPlannerConfiguredTaskBudget/executor_only"] - $tests | length) == 0
-          ' "$RUNNER_TEMP/planner-local-budget.jsonl"
-
-      - name: upload Windows planner local budget inventory
-        if: always() && env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7
-        with:
-          name: planner-local-budget-windows-${{ github.run_attempt }}
-          path: ${{ runner.temp }}/planner-local-budget.jsonl
-          if-no-files-found: ignore
-          retention-days: 7
-
      - name: test (Windows Controller routing root budget)
        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
        timeout-minutes: 5
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(SemanticRootUsage|EffectRoutingRootBudget)$' ./internal/runtime/capability ./internal/assembly/boot | tee "$RUNNER_TEMP/controller-routing-budget.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(SemanticRootUsage)$' ./internal/runtime/capability | tee "$RUNNER_TEMP/controller-routing-budget.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/runtime/capability", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
+            (["reasonix/internal/runtime/capability"] - $packages | length) == 0 and
            ([
-              "TestEffectRoutingRootBudget",
-              "TestEffectRoutingRootBudget/balanced",
-              "TestEffectRoutingRootBudget/cached_continuation",
-              "TestEffectRoutingRootBudget/cached_new_scope",
-              "TestEffectRoutingRootBudget/cached_new_task",
-              "TestEffectRoutingRootBudget/cached_reset",
-              "TestEffectRoutingRootBudget/context_zero",
-              "TestEffectRoutingRootBudget/dual",
-              "TestEffectRoutingRootBudget/headless",
-              "TestEffectRoutingRootBudget/interactive",
-              "TestEffectRoutingRootBudget/missing",
-              "TestEffectRoutingRootBudget/unconfigured",
              "TestSemanticRootUsage",
              "TestSemanticRootUsage/billed_error",
              "TestSemanticRootUsage/cache",
@@ -505,22 +354,17 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(PreparedRootBudget|CoordinatorRootBudget|EffectCoordinatorRoot)' ./internal/runtime/agent ./internal/runtime/coordinator ./internal/assembly/boot | tee "$RUNNER_TEMP/coordinator-root-budget.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(PreparedRootBudget|CoordinatorRootBudget|EffectCoordinatorRoot)' ./internal/runtime/agent ./internal/runtime/coordinator | tee "$RUNNER_TEMP/coordinator-root-budget.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/runtime/agent", "reasonix/internal/runtime/coordinator", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
-            (["TestPreparedRootBudgetLifetime", "TestPreparedRootBudgetLifetime/continuation",
-              "TestPreparedRootBudgetLifetime/new_scope", "TestPreparedRootBudgetLifetime/new_task",
-              "TestPreparedRootBudgetLifetime/reset", "TestPreparedRootBudgetLifetime/new_session",
-              "TestPreparedRootBudgetProjectionAndPreservedTurn", "TestCoordinatorRootBudgetRoutes",
+            (["reasonix/internal/runtime/agent", "reasonix/internal/runtime/coordinator"] - $packages | length) == 0 and
+            (["TestCoordinatorRootBudgetRoutes",
              "TestCoordinatorRootBudgetRoutes/raw", "TestCoordinatorRootBudgetRoutes/tool_planner",
              "TestCoordinatorRootBudgetRoutes/failure_fallback", "TestCoordinatorRootBudgetRoutes/plan_only_continuation",
-              "TestCoordinatorRootBudgetRoutes/approval_declined_continuation", "TestCoordinatorRootBudgetRoutes/approval_allowed",
-              "TestEffectCoordinatorRootTokenBudget", "TestEffectCoordinatorRootTokenBudget/shared",
-              "TestEffectCoordinatorRootTokenBudget/unconfigured", "TestEffectCoordinatorRootTokenBudget/context_disables",
-              "TestEffectCoordinatorRootTokenBudget/missing_planner", "TestEffectCoordinatorRootTokenBudget/executor_only"] - $tests | length) == 0
+              "TestCoordinatorRootBudgetRoutes/approval_declined_continuation", "TestCoordinatorRootBudgetRoutes/approval_allowed"
+              ] - $tests | length) == 0
          ' "$RUNNER_TEMP/coordinator-root-budget.jsonl"

      - name: upload Windows Coordinator root budget inventory
@@ -578,10 +422,8 @@ jobs:
              "TestRootWallBudgetReachesDelegatedAgents/context", "TestRootWallBudgetReachesDelegatedAgents/unconfigured",
              "TestRootWallBudgetReachesDelegatedAgents/explicit_zero", "TestRootWallBudgetContinuationAndReset",
              "TestRootWallBudgetContinuationAndReset/continuation", "TestRootWallBudgetContinuationAndReset/new_task",
-              "TestRootWallBudgetContinuationAndReset/reset", "TestRootWallBudgetContinuationAndReset/override",
-              "TestRootWallBudgetProjectionKeepsJobDeadline", "TestEffectRootWallBudgetReachesManagedJobs",
-              "TestEffectRootWallBudgetReachesManagedJobs/task", "TestEffectRootWallBudgetReachesManagedJobs/capability",
-              "TestEffectRootWallBudgetReachesManagedJobs/fleet"] - $tests | length) == 0
+              "TestRootWallBudgetContinuationAndReset/reset", "TestRootWallBudgetContinuationAndReset/override"
+              ] - $tests | length) == 0
          ' "$RUNNER_TEMP/root-wall-budget.jsonl"

      - name: upload Windows root wall budget inventory
@@ -599,22 +441,19 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(RootToken|EffectRootToken|BudgetCountsKnownReportedTokens|ReportedToken|EffectReportedTotalTokenBudget)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-token-budget.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(RootToken|EffectRootToken|BudgetCountsKnownReportedTokens|ReportedToken)' ./internal/runtime/agent | tee "$RUNNER_TEMP/root-token-budget.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
+            (["reasonix/internal/runtime/agent"] - $packages | length) == 0 and
            (["TestRootTokenBudgetIncludesChildProviderUsage", "TestRootTokenBudgetIncludesChildProviderUsage/configured",
              "TestRootTokenBudgetIncludesChildProviderUsage/context", "TestRootTokenBudgetIncludesChildProviderUsage/unconfigured",
              "TestRootTokenBudgetIncludesChildProviderUsage/explicit_zero", "TestRootTokenBudgetSurvivesNestedAndConcurrentChildren",
              "TestRootTokenBudgetSurvivesNestedAndConcurrentChildren/nested", "TestRootTokenBudgetSurvivesNestedAndConcurrentChildren/siblings",
              "TestRootTokenBudgetSurvivesNestedAndConcurrentChildren/child_override", "TestRootTokenBudgetContinuationAndReset",
              "TestRootTokenBudgetContinuationAndReset/continuation", "TestRootTokenBudgetContinuationAndReset/new_task",
-              "TestRootTokenBudgetContinuationAndReset/reset", "TestRootTokenScopeAtomicAccountingAndJobCancellation",
-              "TestEffectRootTokenBudgetReachesDelegatedJobs", "TestEffectRootTokenBudgetReachesDelegatedJobs/foreground",
-              "TestEffectRootTokenBudgetReachesDelegatedJobs/foreground_capability", "TestEffectRootTokenBudgetReachesDelegatedJobs/background",
-              "TestEffectRootTokenBudgetReachesDelegatedJobs/background_capability", "TestEffectRootTokenBudgetReachesDelegatedJobs/background_fleet"] - $tests | length) == 0
+              "TestRootTokenBudgetContinuationAndReset/reset" ] - $tests | length) == 0
          ' "$RUNNER_TEMP/root-token-budget.jsonl"
          jq -se '
            . as $events |
@@ -628,17 +467,7 @@ jobs:
               "TestBudgetCountsKnownReportedTokens/negative_split",
               "TestBudgetCountsKnownReportedTokens/zero",
               "TestBudgetCountsKnownReportedTokens/negative",
-               "TestReportedTokensEnforceLocalCapOnInheritedRoot",
-               "TestReportedTokensFollowOriginalAuxiliaryRoot",
-               "TestReportedTokenCountersSaturate"],
-             "reasonix/internal/assembly/boot": [
-               "TestEffectReportedTotalTokenBudget",
-               "TestEffectReportedTotalTokenBudget/split",
-               "TestEffectReportedTotalTokenBudget/total_only",
-               "TestEffectReportedTotalTokenBudget/partial_split",
-               "TestEffectReportedTotalTokenBudget/lower_total",
-               "TestEffectReportedTotalTokenBudget/missing_total",
-               "TestEffectReportedTotalTokenBudget/disabled"]} |
+               "TestReportedTokenCountersSaturate"]} |
            to_entries | all(.[]; . as $package |
              ([ $events[] | select(.Package == $package.key and .Test == null and
                 (.Action == "pass" or .Action == "fail" or .Action == "skip")) ] |
@@ -697,7 +526,7 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(HostStorageEnvironment|HostObjectStorage|HostExplicitStorageControls|DetachedStorageEnvironment|WorktreePreparationOwnsItsIndex|EffectIsolatedTaskOwnsGitStorage)$' ./internal/platform/gitcmd ./internal/platform/worktree ./internal/assembly/boot | tee "$RUNNER_TEMP/git-storage-environment.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(HostStorageEnvironment|HostObjectStorage|HostExplicitStorageControls|DetachedStorageEnvironment)$' ./internal/platform/gitcmd ./internal/platform/worktree ./internal/assembly/boot | tee "$RUNNER_TEMP/git-storage-environment.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
@@ -707,10 +536,6 @@ jobs:
              "TestDetachedStorageEnvironment",
              "TestDetachedStorageEnvironment/explicit_env",
              "TestDetachedStorageEnvironment/inherited_env",
-              "TestEffectIsolatedTaskOwnsGitStorage",
-              "TestEffectIsolatedTaskOwnsGitStorage/ambient_existing",
-              "TestEffectIsolatedTaskOwnsGitStorage/ambient_missing",
-              "TestEffectIsolatedTaskOwnsGitStorage/control",
              "TestHostExplicitStorageControls",
              "TestHostExplicitStorageControls/explicit_object_directory",
              "TestHostExplicitStorageControls/private_index",
@@ -721,15 +546,8 @@ jobs:
              "TestHostStorageEnvironment",
              "TestHostStorageEnvironment/GIT_ALTERNATE_OBJECT_DIRECTORIES",
              "TestHostStorageEnvironment/GIT_INDEX_FILE",
-              "TestHostStorageEnvironment/GIT_OBJECT_DIRECTORY",
-              "TestWorktreePreparationOwnsItsIndex",
-              "TestWorktreePreparationOwnsItsIndex/candidate_ambient_existing",
-              "TestWorktreePreparationOwnsItsIndex/candidate_ambient_missing",
-              "TestWorktreePreparationOwnsItsIndex/candidate_control",
-              "TestWorktreePreparationOwnsItsIndex/delivery_ambient_existing",
-              "TestWorktreePreparationOwnsItsIndex/delivery_ambient_missing",
-              "TestWorktreePreparationOwnsItsIndex/delivery_control"
-            ] - $tests | length) == 0
+              "TestHostStorageEnvironment/GIT_OBJECT_DIRECTORY"
+              ] - $tests | length) == 0
          ' "$RUNNER_TEMP/git-storage-environment.jsonl"

      - name: upload Windows Git storage environment inventory
@@ -779,20 +597,19 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(RootWebSearch|EffectRootWebSearch)' ./internal/tools/websearch ./internal/assembly/boot | tee "$RUNNER_TEMP/root-websearch-budget.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(RootWebSearch|EffectRootWebSearch)' ./internal/tools/websearch | tee "$RUNNER_TEMP/root-websearch-budget.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/tools/websearch", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
+            (["reasonix/internal/tools/websearch"] - $packages | length) == 0 and
            (["TestRootWebSearchUsageTerminal", "TestRootWebSearchUsageTerminal/reported",
              "TestRootWebSearchUsageTerminal/no_reporter", "TestRootWebSearchUsageTerminal/missing",
              "TestRootWebSearchUsageTerminal/no_observer", "TestRootWebSearchUsageTerminal/input_refused",
              "TestRootWebSearchUsageTerminal/no_factory", "TestRootWebSearchUsageTerminal/factory_error",
              "TestRootWebSearchUsageTerminal/billed_failure", "TestRootWebSearchUsageTerminal/cancel_after_usage",
-              "TestRootWebSearchUsageTerminal/cancel_before_start", "TestRootWebSearchUsageTerminal/stream_error",
-              "TestEffectRootWebSearchTokenBudget", "TestEffectRootWebSearchTokenBudget/reported",
-              "TestEffectRootWebSearchTokenBudget/disabled", "TestEffectRootWebSearchTokenBudget/missing"] - $tests | length) == 0
+              "TestRootWebSearchUsageTerminal/cancel_before_start", "TestRootWebSearchUsageTerminal/stream_error"
+              ] - $tests | length) == 0
          ' "$RUNNER_TEMP/root-websearch-budget.jsonl"

      - name: upload Windows root websearch budget inventory
@@ -810,21 +627,13 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^(TestProfileRootBudgetLifetime|TestProfileBudgetPreparationUsesTurnAdmission|TestEffectExplicitSkillRootBudget)$' ./internal/session/control ./internal/assembly/boot | tee "$RUNNER_TEMP/explicit-skill-root-budget.jsonl"
+          go test -count=1 -timeout=3m -json -run '^(TestProfileRootBudgetLifetime|TestProfileBudgetPreparationUsesTurnAdmission|TestEffectExplicitSkillRootBudget)$' ./internal/session/control | tee "$RUNNER_TEMP/explicit-skill-root-budget.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/session/control", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
-            (["TestProfileRootBudgetLifetime", "TestProfileRootBudgetLifetime/new_task",
-              "TestProfileRootBudgetLifetime/same_scope", "TestProfileRootBudgetLifetime/new_scope",
-              "TestProfileRootBudgetLifetime/reset", "TestProfileRootBudgetLifetime/new_session",
-              "TestProfileRootBudgetLifetime/inherited", "TestProfileRootBudgetLifetime/context_zero",
-              "TestProfileBudgetPreparationUsesTurnAdmission", "TestEffectExplicitSkillRootBudget",
-              "TestEffectExplicitSkillRootBudget/profile", "TestEffectExplicitSkillRootBudget/preview",
-              "TestEffectExplicitSkillRootBudget/slash", "TestEffectExplicitSkillRootBudget/chips",
-              "TestEffectExplicitSkillRootBudget/goal", "TestEffectExplicitSkillRootBudget/context_zero",
-              "TestEffectExplicitSkillRootBudget/unconfigured"] - $tests | length) == 0
+            (["reasonix/internal/session/control"] - $packages | length) == 0 and
+            (["TestProfileBudgetPreparationUsesTurnAdmission" ] - $tests | length) == 0
          ' "$RUNNER_TEMP/explicit-skill-root-budget.jsonl"

      - name: upload Windows explicit skill root budget inventory
@@ -842,19 +651,16 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(RootTriage|EffectRootTriage)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-triage-budget.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(RootTriage|EffectRootTriage)' ./internal/runtime/agent | tee "$RUNNER_TEMP/root-triage-budget.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
+            (["reasonix/internal/runtime/agent"] - $packages | length) == 0 and
            (["TestRootTriageTokenBudget", "TestRootTriageTokenBudget/classifier", "TestRootTriageTokenBudget/screen",
              "TestRootTriageTokenBudget/billed_failure", "TestRootTriageTokenBudget/missing", "TestRootTriageTokenBudget/disabled",
              "TestRootTriageTokenBudget/no_sink", "TestRootTriageTokenBudget/counts_without_total",
-              "TestRootTriageOutsideRunDoesNotCreateScope", "TestEffectRootTriageScreenBudget",
-              "TestEffectRootTriageScreenBudget/reported", "TestEffectRootTriageScreenBudget/flagged",
-              "TestEffectRootTriageScreenBudget/disabled", "TestEffectRootTriageScreenBudget/missing",
-              "TestEffectRootTriageScreenBudget/screen_disabled"] - $tests | length) == 0
+              "TestRootTriageOutsideRunDoesNotCreateScope" ] - $tests | length) == 0
          ' "$RUNNER_TEMP/root-triage-budget.jsonl"

      - name: upload Windows root triage budget inventory
@@ -866,54 +672,22 @@ jobs:
          if-no-files-found: ignore
          retention-days: 7

-      - name: test (Windows root auxiliary budget)
-        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        timeout-minutes: 5
-        shell: bash
-        run: |
-          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(RootAuxiliary|EffectRootAuxiliary)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-auxiliary-budget.jsonl"
-          jq -se '
-            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
-            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
-            all(.[]; .Action != "skip" and .Action != "fail") and
-            (["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
-            (["TestRootAuxiliaryAutomaticCompaction", "TestRootAuxiliaryTokenBudget",
-              "TestRootAuxiliaryTokenBudget/summary", "TestRootAuxiliaryTokenBudget/summary_failed",
-              "TestRootAuxiliaryTokenBudget/summary_missing", "TestRootAuxiliaryTokenBudget/summary_disabled",
-              "TestRootAuxiliaryTokenBudget/diagnosis", "TestRootAuxiliaryTokenBudget/diagnosis_failed",
-              "TestRootAuxiliaryTokenBudget/diagnosis_missing", "TestRootAuxiliaryTokenBudget/diagnosis_disabled",
-              "TestRootAuxiliaryManualCompactionOutsideRun", "TestRootAuxiliaryBudgetProjectionKeepsOldTask",
-              "TestEffectRootAuxiliarySummaryBudget", "TestEffectRootAuxiliarySummaryBudget/configured",
-              "TestEffectRootAuxiliarySummaryBudget/disabled", "TestEffectRootAuxiliarySummaryBudget/missing_usage"] - $tests | length) == 0
-          ' "$RUNNER_TEMP/root-auxiliary-budget.jsonl"
-
-      - name: upload Windows root auxiliary budget inventory
-        if: always() && env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
-        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7
-        with:
-          name: root-auxiliary-budget-windows-${{ github.run_attempt }}
-          path: ${{ runner.temp }}/root-auxiliary-budget.jsonl
-          if-no-files-found: ignore
-          retention-days: 7
-
      - name: test (Windows local cost quote)
        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(LocalBudgetUsesOccurrenceQuote|LocalCost.*|UsageQuote.*|EffectLocalCostUsesOccurrenceQuote)$' ./internal/contract/event ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/local-cost-quote.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(LocalBudgetUsesOccurrenceQuote|LocalCost.*|UsageQuote.*)$' ./internal/contract/event ./internal/runtime/agent | tee "$RUNNER_TEMP/local-cost-quote.jsonl"
          jq -se '
            ["TestLocalBudgetUsesOccurrenceQuote", "TestLocalBudgetUsesOccurrenceQuote/peak", "TestLocalBudgetUsesOccurrenceQuote/offpeak",
             "TestLocalCostKeepsOriginalCurrencyBuckets", "TestLocalCostFixedPointThresholdAndUnknown",
-             "TestUsageQuoteUsesOneOwnerAndOccurrence", "TestUsageQuoteLegacyAndNilSinks",
-             "TestEffectLocalCostUsesOccurrenceQuote", "TestEffectLocalCostUsesOccurrenceQuote/peak",
-             "TestEffectLocalCostUsesOccurrenceQuote/offpeak", "TestEffectLocalCostUsesOccurrenceQuote/relay"] as $required |
+             "TestUsageQuoteUsesOneOwnerAndOccurrence", "TestUsageQuoteLegacyAndNilSinks"
+             ] as $required |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $passed |
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            all(.[]; .Action != "fail" and .Action != "skip") and
            (($required - $passed) | length == 0) and
-            ((["reasonix/internal/contract/event", "reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages) | length == 0)
+            ((["reasonix/internal/contract/event", "reasonix/internal/runtime/agent"] - $packages) | length == 0)
          ' "$RUNNER_TEMP/local-cost-quote.jsonl"

      - name: upload Windows local cost quote inventory
@@ -963,24 +737,18 @@ jobs:
        shell: bash
        run: |
          set -o pipefail
-          go test -count=1 -timeout=3m -json -run '^Test(AuxiliaryUsageObserver|BoundedUsageObserver|RootBounded|EffectRootBounded)' ./internal/contract/provider ./internal/model/boundedllm ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/bounded-root-budget.jsonl"
+          go test -count=1 -timeout=3m -json -run '^Test(AuxiliaryUsageObserver|BoundedUsageObserver|RootBounded|EffectRootBounded)' ./internal/contract/provider ./internal/model/boundedllm ./internal/runtime/agent | tee "$RUNNER_TEMP/bounded-root-budget.jsonl"
          jq -se '
            [.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
            [.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
            all(.[]; .Action != "skip" and .Action != "fail") and
            (["reasonix/internal/contract/provider", "reasonix/internal/model/boundedllm",
-              "reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
+              "reasonix/internal/runtime/agent"] - $packages | length) == 0 and
            (["TestAuxiliaryUsageObserverReplacesAndCopies",
              "TestBoundedUsageObserverOwnsOneTerminal", "TestBoundedUsageObserverOwnsOneTerminal/reported",
              "TestBoundedUsageObserverOwnsOneTerminal/billed_failure", "TestBoundedUsageObserverOwnsOneTerminal/without_sink",
-              "TestBoundedUsageObserverOwnsOneTerminal/missing", "TestBoundedUsageObserverOwnsOneTerminal/input_refused",
-              "TestRootBoundedReviewerBudget", "TestRootBoundedReviewerBudget/reported",
-              "TestRootBoundedReviewerBudget/billed_failure", "TestRootBoundedReviewerBudget/missing_usage",
-              "TestRootBoundedReviewerBudget/disabled", "TestRootBoundedReviewerBudget/no_emission",
-              "TestRootBoundedReviewerBudget/projected_job_context", "TestRootBoundedNestedBudgetChargesOnce",
-              "TestRootBoundedLateProjectionStaysWithOldTask", "TestEffectRootBoundedAdvisorBudget",
-              "TestEffectRootBoundedAdvisorBudget/reported", "TestEffectRootBoundedAdvisorBudget/disabled",
-              "TestEffectRootBoundedAdvisorBudget/missing_usage"] - $tests | length) == 0
+              "TestBoundedUsageObserverOwnsOneTerminal/missing", "TestBoundedUsageObserverOwnsOneTerminal/input_refused"
+              ] - $tests | length) == 0
          ' "$RUNNER_TEMP/bounded-root-budget.jsonl"

      - name: upload Windows bounded root budget inventory
@@ -1009,42 +777,6 @@ jobs:
        timeout-minutes: 5
        run: go test -count=1 -timeout=3m -run '^TestBashPowerShellExecuteDetailedContract$|^TestBashPowerShell51PreflightRejectsAndAndDetailed$|^TestBashPowerShellOutputIsUTF8$|^TestBashPowerShellSurfacesNonZeroExit$|^TestBashPowerShellRejectsChaining$' ./internal/tools/builtin

-      # Full push suite: a 36m package timeout inside a 45m step budget. -count=1
-      # as at the smoke step. A failure prints the failed tests' output and the 15
-      # slowest boot tests.
-      - name: test (full)
-        if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name != 'pull_request'
-        timeout-minutes: 45
-        env:
-          # Run the prompt-cache prefix-stability guard (TestCacheHit*) in CI: a
-          # regression there silently tanks the cache hit rate the project is
-          # built around.
-          REASONIX_RELEASE_CACHE_GUARD: "1"
-          # Bound sandbox helper children in Windows tests so a failed OS-level
-          # launch cannot pin the Actions step after Go's package timeout fires.
-          WINDOWS_SANDBOX_WAIT_MS: "20000"
-        shell: bash
-        run: |
-          # internal/assembly/boot alone can take over 25 minutes on a Windows runner;
-          # the package limit leaves headroom inside the step limit above.
-          go test -count=1 -p 4 -timeout=36m -json ./... | tee "$RUNNER_TEMP/full.jsonl" |
-            jq -rj --unbuffered 'select(.Action == "output" and .Test == null) | .Output'
-
-      - name: test (full) failure detail
-        if: failure() && env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name != 'pull_request'
-        shell: bash
-        run: |
-          f="$RUNNER_TEMP/full.jsonl"
-          [ -s "$f" ] || exit 0
-          echo "::group::output of failed tests"
-          jq -rn '[inputs | select(.Test != null)] | group_by(.Package + "/" + .Test)
-            | map(select(any(.[]; .Action == "fail")))[]
-            | map(select(.Action == "output") | .Output) | join("")' "$f" || true
-          echo "::endgroup::"
-          echo "slowest internal/assembly/boot tests (seconds, name):"
-          jq -rn '[inputs | select((.Package | endswith("/internal/assembly/boot")) and .Test != null and .Action == "pass")]
-            | sort_by(-.Elapsed)[:15][] | "\(.Elapsed)\t\(.Test)"' "$f" || true
-
  race:
    needs: changes
    if: always() && (github.event_name != 'pull_request' || needs.changes.outputs.code != 'false')
@@ -1079,15 +811,130 @@ jobs:
        if: github.event_name == 'pull_request'
        env:
          REASONIX_RELEASE_CACHE_GUARD: "1"
-        run: go test -race ./internal/runtime/agent/... ./internal/runtime/delegation/... ./internal/runtime/isolation/... ./internal/platform/worktree/... ./internal/state/mutation/... ./internal/ext/plugin/... ./internal/tools/jobs/... ./internal/base/proc/... ./internal/safety/sandbox/... ./internal/base/filelock/... ./internal/contract/eventwire/... ./internal/platform/remote/... ./internal/ext/extension/... ./internal/assembly/boot/... ./internal/session/control/... ./internal/contract/tool/...
+        run: go test -race ./internal/runtime/agent/... ./internal/runtime/delegation/... ./internal/platform/worktree/... ./internal/state/mutation/... ./internal/ext/plugin/... ./internal/tools/jobs/... ./internal/base/proc/... ./internal/safety/sandbox/... ./internal/base/filelock/... ./internal/contract/eventwire/... ./internal/platform/remote/... ./internal/ext/extension/... ./internal/assembly/boot/... ./internal/session/control/... ./internal/contract/tool/...

-      - name: test -race (full)
-        if: github.event_name != 'pull_request'
+  # Push-only full suites. internal/frontend/serve and internal/assembly/boot run
+  # hundreds of serial tests and set the wall time of one whole-tree `go test`,
+  # so each is cut by test name into shards that run on their own runners. The
+  # -run / -skip patterns partition the names exactly: every test lands in one
+  # shard. `rest` is every other package. Pull requests keep the focused steps
+  # in the `test` and `race` jobs above.
+  test-windows-shards:
+    if: github.event_name != 'pull_request'
+    name: test (windows-latest, ${{ matrix.shard }})
+    runs-on: windows-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { shard: rest }
+          - { shard: serve-1, pkg: ./internal/frontend/serve, flag: -run, filter: '^Test[A-E]' }
+          - { shard: serve-2, pkg: ./internal/frontend/serve, flag: -run, filter: '^Test[F-R]' }
+          - { shard: serve-3, pkg: ./internal/frontend/serve, flag: -skip, filter: '^Test[A-R]' }
+          - { shard: boot-1, pkg: ./internal/assembly/boot, flag: -skip, filter: '^TestEffect([^0-9A-C]|$)' }
+          - { shard: boot-2, pkg: ./internal/assembly/boot, flag: -run, filter: '^TestEffect[D-O]' }
+          - { shard: boot-3, pkg: ./internal/assembly/boot, flag: -run, filter: '^TestEffect([^0-9A-O]|$)' }
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
+
+      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
+        with:
+          go-version-file: go.mod
+          cache: true
+
+      - name: test (shard)
+        timeout-minutes: 40
+        env:
+          # Run the prompt-cache prefix-stability guard (TestCacheHit*) in CI: a
+          # regression there silently tanks the cache hit rate the project is
+          # built around.
+          REASONIX_RELEASE_CACHE_GUARD: "1"
+          # Bound sandbox helper children in Windows tests so a failed OS-level
+          # launch cannot pin the Actions step after Go's package timeout fires.
+          WINDOWS_SANDBOX_WAIT_MS: "20000"
+          SHARD: ${{ matrix.shard }}
+          PKG: ${{ matrix.pkg }}
+          FLAG: ${{ matrix.flag }}
+          FILTER: ${{ matrix.filter }}
+        shell: bash
+        run: |
+          if [ "$SHARD" = rest ]; then
+            mapfile -t pkgs < <(go list ./... | tr -d '\r' | grep -v -E '/internal/(frontend/serve|assembly/boot)$')
+          else
+            pkgs=("$PKG")
+          fi
+          filter=()
+          if [ -n "$FILTER" ]; then filter=("$FLAG" "$FILTER"); fi
+          go test -count=1 -p 4 -timeout=36m -json "${filter[@]}" "${pkgs[@]}" | tee "$RUNNER_TEMP/shard.jsonl" |
+            jq -rj --unbuffered 'select(.Action == "output" and .Test == null) | .Output'
+
+      - name: test (shard) failure detail
+        if: failure()
+        shell: bash
+        run: |
+          f="$RUNNER_TEMP/shard.jsonl"
+          [ -s "$f" ] || exit 0
+          echo "::group::output of failed tests"
+          jq -rn '[inputs | select(.Test != null)] | group_by(.Package + "/" + .Test)
+            | map(select(any(.[]; .Action == "fail")))[]
+            | map(select(.Action == "output") | .Output) | join("")' "$f" || true
+          echo "::endgroup::"
+          echo "slowest tests (seconds, package, name):"
+          jq -rn '[inputs | select(.Test != null and .Action == "pass")]
+            | sort_by(-.Elapsed)[:15][] | "\(.Elapsed)	\(.Package)	\(.Test)"' "$f" || true
+
+  race-shards:
+    if: github.event_name != 'pull_request'
+    name: race (${{ matrix.shard }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { shard: rest }
+          - { shard: serve-1, pkg: ./internal/frontend/serve, flag: -run, filter: '^Test[A-E]' }
+          - { shard: serve-2, pkg: ./internal/frontend/serve, flag: -run, filter: '^Test[F-R]' }
+          - { shard: serve-3, pkg: ./internal/frontend/serve, flag: -skip, filter: '^Test[A-R]' }
+          - { shard: boot-1, pkg: ./internal/assembly/boot, flag: -skip, filter: '^TestEffect([^0-9A-C]|$)' }
+          - { shard: boot-2, pkg: ./internal/assembly/boot, flag: -run, filter: '^TestEffect[D-O]' }
+          - { shard: boot-3, pkg: ./internal/assembly/boot, flag: -run, filter: '^TestEffect([^0-9A-O]|$)' }
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
+
+      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
+        with:
+          go-version-file: go.mod
+          cache: true
+
+      - name: Install and verify Linux sandbox backend
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y bubblewrap
+          if sysctl -n kernel.unprivileged_userns_clone >/dev/null 2>&1; then
+            sudo sysctl -w kernel.unprivileged_userns_clone=1
+          fi
+          if sysctl -n kernel.apparmor_restrict_unprivileged_userns >/dev/null 2>&1; then
+            sudo sysctl -w kernel.apparmor_restrict_unprivileged_userns=0
+          fi
+          bwrap --ro-bind / / --dev /dev --proc /proc -- true ||             echo "::warning::bubblewrap is installed but this runner denies user namespaces; unavailable-backend tests will run fail-closed"
+
+      - name: test -race (shard)
+        timeout-minutes: 30
        env:
          REASONIX_RELEASE_CACHE_GUARD: "1"
-        # internal/assembly/boot under -race runs past ten minutes on a shared
-        # runner; the package limit leaves headroom over the default.
-        run: go test -race -timeout=25m ./...
+          SHARD: ${{ matrix.shard }}
+          PKG: ${{ matrix.pkg }}
+          FLAG: ${{ matrix.flag }}
+          FILTER: ${{ matrix.filter }}
+        run: |
+          if [ "$SHARD" = rest ]; then
+            mapfile -t pkgs < <(go list ./... | grep -v -E '/internal/(frontend/serve|assembly/boot)$')
+          else
+            pkgs=("$PKG")
+          fi
+          filter=()
+          if [ -n "$FILTER" ]; then filter=("$FLAG" "$FILTER"); fi
+          go test -race -timeout=25m "${filter[@]}" "${pkgs[@]}"

  # sdk/go is a nested stdlib-only module invisible to root `go test ./...`.
  # Its DTOs are generated from internal/ext/extension/protocol, and the
@@ -0,0 +1,27 @@
+name: PR author gate
+
+# Pull requests are limited to maintainers and the logins in .github/pr-authors.txt.
+# This runs in the base repository's context, checks out the base branch only, and
+# never executes pull-request code; the allowlist it reads is the base branch's.
+on:
+  pull_request_target:
+    types: [opened]
+
+permissions:
+  contents: read
+  pull-requests: write
+
+concurrency:
+  group: pr-author-gate-${{ github.event.pull_request.number }}
+  cancel-in-progress: false
+
+jobs:
+  gate:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
+        with:
+          persist-credentials: false
+      - run: node scripts/pr-author-gate.mjs
+        env:
+          GH_TOKEN: ${{ github.token }}
@@ -16,7 +16,7 @@ permissions:

 concurrency:
  group: studio-${{ github.ref }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' || github.ref == 'refs/heads/studio' }}

 jobs:
  frontend:
@@ -103,10 +103,6 @@ func newHarness(t *testingDir, p provider.Provider, window int, rec *callRecorde
 		ContextWindow: window,
 		ArchiveDir:    filepath.Join(t.dir, "archive"),
 		SessionPath:   path,
-		RecentKeep:    4,
-		// boot's default when cfg.Agent.Keep is unset; without it the bench
-		// would measure a configuration no real session runs.
-		KeepPolicy: agent.KeepErrors,
 	}, rec.sink())
 	return &harness{sess: sess, agentA: a, path: path, calls: rec}
 }
@@ -108,7 +108,6 @@ func newAgent(p provider.Provider, sess *sessionstore.Session, path string, sink
 	return agent.New(p, tool.NewRegistry(), sess, agent.Options{
 		ContextWindow: windowTokens,
 		CompactRatio:  compactRatio,
-		RecentKeep:    2,
 		SessionPath:   path,
 		WorkspaceID:   "cm-e2e",
 		ModelRef:      "deepseek/" + model,
@@ -1,180 +0,0 @@
-# Context retrieval bench
-
-Two questions about long-horizon memory, and the harness that makes their
-answers trustworthy.
-
-    -mode=preflight     build all 18 tasks under every arm and check the contract
-    -mode=snippet       does a rank-1 hit hand over the whole answer?
-    -mode=adversarial   send a model after the answer through every host surface
-    -mode=run-search    Experiment R: is searchable canonical a real capability?
-    -mode=run-boundary  Experiment I, paired: the 6 efficiency tasks across their boundary
-    -mode=run-index     Experiment I, full: the 6 efficiency tasks x 4 budgets
-    -mode=calibrate     find each index task's PlantAfterGen for its cue tier
-
-`-dry` drives any run mode with a scripted provider, for nothing. Real runs need
-`DEEPSEEK_API_KEY`.
-
-## Status
-
-**Searchable canonical recall: proven.** Six of six recovered with search on,
-three of six with it off, and the successes without it came from enumerating
-addresses by hand at roughly nine times the page-in cost (139 vs 1292
-tokens/task). Measured after three separate leaks were closed, so the numbers
-are from a clean environment.
-
-**Fold index marginal utility: direction reproduced, size still conditional.** On the v2
-substrate the direction holds in six of six tasks and in the pooled paired batch: with the
-cue visible the index addresses the target, so the same task is searched less and pages in
-less (0.67 searches and 70 recall tokens per task, against 1.50 and 120 with the cue
-absent).
-
-A size is visible but is not yet quotable. On the four tasks whose absent-cue arm stayed
-out of a loop, the index cost 64,952 prompt tokens against 115,551 without, 44% less
-(`i106-budgetseal` -35,147, `i70-ingestfloor` -8,222, `i37-dispatchlag` -7,599, and
-`i112-lagquota` +369, the prefix rent a task that never searches pays). That is a
-**post-hoc subset**: `i118-fencelag` and `i47-tenantlease` are excluded because their
-absent-cue arm looped in this batch (five and three retrieval rounds, three escapes each;
-together 80% of the absent-cue total), and the qualification above does not screen for
-stopping behaviour — holding on both boundary arms across three passes shows a stable cue
-boundary, not a stopping arm. Quoting 44% as the index's effect would mean choosing the
-sample after seeing the result; a batch screened on stopping stability is what would make
-a size quotable, and one arm also reverses either way (`i70-ingestfloor` default, +11,135
-prompt tokens over three escapes).
-
-Batch facts, because the counts differ: 36 runs were executed (12 paired, 24 index), and
-the token totals here cover the 24-run index batch — the paired batch ran before the
-per-run usage line existed.
-
-**Things that turned out not to be problems**, each after being measured rather
-than argued about:
-
-| Suspected | Measured |
-| --- | --- |
-| Model writes poor queries (2.17 searches/task) | 22/23 found the target on query 1; the extra searches are one round's parallel fan-out |
-| Model reaches for the workspace before memory (7/11) | Linear event order misread; by model round it is MemoryFirst 4/6 |
-| Snippets too short for multi-value answers | 240 runes covers 12/12 tasks at 100%; records are 30-179 runes |
-| A general stopping failure | 37/44 runs issued no search after the answer was in hand; all 7 that did belong to 2 tasks |
-
-## Measurement distinctions this bench had to learn
-
-Each of these reversed a conclusion once. They are enforced in the schema now,
-not left to whoever reads the numbers next.
-
- **TargetHit is not EvidenceSufficient.** A rank-1 hit can return a window too
-  short to answer with. `FirstHitCoverage` records how much of the scored answer
-  the first hit actually handed over.
- **Linear event order is not model-round order.** Two calls in one round are a
-  parallel fan-out, not a preference. Routing compares rounds.
- **An isolated workdir is not an isolated host.** The sandbox mounts the host
-  read-only by design. Three leaks were found and closed: the corpus source, the
-  fixture transcript, and its event log sidecar.
- **A count is not its evidence.** Query text, snippet contents and trajectories
-  are all persisted now, because three analyses in a row needed a paid re-run to
-  ask a question of data already collected.
-
-## How a run stays honest
-
-The corpus holds templates, never answers: every scored literal is a `{{var}}`
-instantiated per run, so grepping this directory reveals the question and not
-the answer. `TestNoAnswerLiteralExistsInTheRepository` asserts it with
-`git grep`. Preflight rebuilds every arm and checks the answer is unreadable in
-the real `provider.Request`, the probe query ranks the target within five, and
-an index task's cue sits at exactly the scales its tier names. The fixture is
-removed from disk once the agent holds it in memory, and the directory is
-scanned for answer literals before the first provider request and again after
-the turn. Any answer reaching the model through a tool that is not `recall`
-marks the run contaminated and takes it out of every statistic.
-
-## Index corpus v2
-
-The efficiency substrate is six tasks that hold on both arms: `i37-dispatchlag`
-and `i118-fencelag` (quarter), `i106-budgetseal` and `i70-ingestfloor` (half),
-`i47-tenantlease` and `i112-lagquota` (default).
-
-They came out of five batches of candidates, each screened the same way: one
-`-mode=run-index` pass over the batch, then two more only for the candidates that
-were clean on both boundary arms, and only the candidates that held on both arms
-in all three passes stayed. A candidate that does not hold is dropped, not
-promoted. The numbers are candidates, per tier:
-
-| batch | quarter | half | default |
-| --- | --- | --- | --- |
-| first pool, `i31`-`i54` | 1 of 8 | 0 of 8 | 2 of 8 |
-| second pool, `i61`-`i84` | 1 of 4 screened | 4 of 8 | 0 of 8 |
-| third pool, `i85`-`i108` | 2 of 8 | 8 of 16 | none authored |
-| replacement pool, `i109`-`i116` | - | - | 4 of 8 |
-| replacement pool, `i92`, `i117`-`i124` | 3 of 9 | - | - |
-
-That table counts candidates that held on both arms in all three passes, which is
-not the size of the substrate. 89 candidates were authored, 77 were screened (the
-second pool's quarter tier stopped at 4 of 8 and its default tier was never run),
-and 25 held: 7 quarter, 12 half, 6 default. The substrate keeps two per tier, so
-the other 19 survivors are spares rather than corpus members. Candidates from the
-earlier round are not counted here.
-
-| task | tier | cue present | cue absent |
-| --- | --- | ---: | ---: |
-| `i37-dispatchlag` | quarter | 3/3 | 3/3 |
-| `i118-fencelag` | quarter | 3/3 | 3/3 |
-| `i106-budgetseal` | half | 3/3 | 3/3 |
-| `i70-ingestfloor` | half | 3/3 | 3/3 |
-| `i47-tenantlease` | default | 3/3 | 3/3 |
-| `i112-lagquota` | default | 3/3 | 3/3 |
-
-Across their 36 judged cells: `SnippetStop` 23, `SnippetThenRead` 10 and
-`DefensiveRead` 3 - every cell clean, nothing post-sufficient and nothing
-`NeverSufficient`.
-
-Four of the first six were flagged in review for sharing answer values, probe
-queries and prompt wording within a tier. Separating them and re-screening those
-four changed two of them: `i37` and `i47` held, `i62` and `i51` did not, so they
-were replaced - `i90-recoverylag` was tried first and did not hold either, and
-`i118-fencelag` and `i112-lagquota` are what did. The separation was not
-cosmetic: the shared wording had been carrying two of the six.
-
-Nothing in the task shape separates the survivors from the rest: within the half
-tier, plain-number answers survived 0 of 8 in the first pool and 4 of 8 in the
-second, and codename answers survived 8 of 16 in the third. An earlier claim here
-that the answer kind decides whether a candidate holds does not survive this
-table and is withdrawn. The dirty class was `PostSufficientRetrieval` in most
-failures - the model re-checking a value it already had - but which tasks hit it
-was not predictable from their shape at this sample.
-
-`i01`-`i06` are the frozen study's six tasks, kept as `roleStopping`. Measured
-once each with `-mode=run-index -task <id>`, per task and per arm:
-
-| task | tier | cue present | cue absent |
-| --- | --- | --- | --- |
-| `i01-transport-fallback` | quarter | SnippetStop | DefensiveRead |
-| `i02-scheduler-handoff` | quarter | PostSufficientRetrieval | PostSufficientRetrieval |
-| `i03-coalescing` | half | PostSufficientRetrieval | PostSufficientRunaway |
-| `i04-recovery-fence` | half | DefensiveRead | PostSufficientRunaway |
-| `i05-cache-probe` | default | SnippetThenRead | PostSufficientRetrieval |
-| `i06-dispatch-handoff` | default | SnippetStop | PostSufficientRetrieval |
-
-Five of the six show post-sufficient behaviour in at least one arm (`i02`-`i06`),
-and `i01` is clean on both. The cue-present arm is clean for four of them
-(`i01`, `i04`, `i05`, `i06`), which is why they stay in the corpus as the
-stopping side rather than being retitled. All six are excluded from `run-index`
-and `run-boundary` batches and stay measurable one at a time with `-task <id>`. A
-task added to the corpus declares a role, so it cannot arrive without one.
-
-A task qualifies only if it is clean on both of its boundary arms - the
-cue-present run and the cue-absent run - in every run of a screening batch. Clean
-means a `stopping_class` other than `PostSufficientRetrieval` or
-`PostSufficientRunaway`: `SnippetStop`, `SnippetThenRead` and `DefensiveRead`
-all qualify, and a defensive read is deliberately not counted as waste. The
-class comes from the three counts in that report — `searches_after_sufficient`,
-`reads_after_sufficient` and `escapes_after_sufficient` — together with how
-sufficiency was reached and how many rounds followed it. The remaining
-requirements still apply — successful tool call, answer in the output rather
-than the cue, rank 1, `FirstHitCoverage` complete, one or two simple values,
-nothing in the workspace.
-
-`-mode=calibrate` is not a screening tool: it builds and reads back the
-fixtures, and reads nothing else, so its cue-visibility profile is the same for
-every task.
-
-Then repeats rather than levels: cue-present against cue-absent on each task's
-own boundary, three runs per cell. What is unknown is the variance, and four
-budgets sampled once each cannot show it.
@@ -1,86 +0,0 @@
-package main
-
-import (
-	"encoding/json"
-	"os"
-	"path/filepath"
-	"strings"
-	"testing"
-)
-
-// runBlockedBatch runs one dry index batch where every arm's runOne fails
-// before any model turn: the task's own directory is a regular file, so the
-// fixture cannot be built. In production the same early return happens when the
-// fixture build, the session load, or the answer-on-disk guard fails.
-func runBlockedBatch(t *testing.T) string {
-	t.Helper()
-	tk := indexTasks()[0]
-	root := t.TempDir()
-	if err := os.WriteFile(filepath.Join(root, tk.ID), []byte("not a directory\n"), 0o644); err != nil {
-		t.Fatal(err)
-	}
-	if code := runExperiment("run-index", experimentIndex, root, true, []contextTask{tk}); code != 0 {
-		t.Fatalf("runExperiment exit = %d, want 0 (every run failed, but the batch finished)", code)
-	}
-	return root
-}
-
-func readResultLines(t *testing.T, root string) ([]contextMetrics, []string) {
-	t.Helper()
-	raw, err := os.ReadFile(filepath.Join(root, "results.jsonl"))
-	if err != nil {
-		t.Fatal(err)
-	}
-	var runs []contextMetrics
-	var lines []string
-	for text := range strings.SplitSeq(strings.TrimSpace(string(raw)), "\n") {
-		var m contextMetrics
-		if err := json.Unmarshal([]byte(text), &m); err != nil {
-			t.Fatalf("results.jsonl: %v", err)
-		}
-		runs = append(runs, m)
-		lines = append(lines, text)
-	}
-	return runs, lines
-}
-
-// A run that failed before the model turn produced no measurement. Recording it
-// as a zero-valued contextMetrics puts it in the funnel's Scored denominator,
-// in "Where it broke", and in the dose curve as a real zero.
-func TestBatchDoesNotRecordAFailedRunAsAResult(t *testing.T) {
-	root := runBlockedBatch(t)
-	runs, lines := readResultLines(t, root)
-	if len(runs) == 0 {
-		t.Fatal("no runs recorded at all")
-	}
-	for i, m := range runs {
-		if m.Task == "" || m.Arm == "" || m.FailureStage == "" {
-			t.Fatalf("results.jsonl line %d records a run that never ran as a result "+
-				"(task=%q arm=%q failure_stage=%q):\n  %s", i+1, m.Task, m.Arm, m.FailureStage, lines[i])
-		}
-	}
-}
-
-// The manifest is the artefact that says which runs counted, so a failed run
-// must not be added to CompletedRuns or to the task id list.
-func TestBatchManifestDoesNotReportFailedRunsAsCompleted(t *testing.T) {
-	root := runBlockedBatch(t)
-	var manifest batchManifest
-	data, err := os.ReadFile(filepath.Join(root, "batch.json"))
-	if err != nil {
-		t.Fatal(err)
-	}
-	if err := json.Unmarshal(data, &manifest); err != nil {
-		t.Fatal(err)
-	}
-	for _, id := range manifest.TaskIDs {
-		if id == "" {
-			t.Fatalf("manifest lists an empty task id: task_ids=%q (completed=%d planned=%d)",
-				manifest.TaskIDs, manifest.CompletedRuns, manifest.PlannedRuns)
-		}
-	}
-	if manifest.CompletedRuns == manifest.PlannedRuns {
-		t.Fatalf("manifest reports every planned run completed although none produced evidence: "+
-			"completed=%d planned=%d task_ids=%q", manifest.CompletedRuns, manifest.PlannedRuns, manifest.TaskIDs)
-	}
-}
@@ -1,155 +0,0 @@
-package main
-
-import (
-	"crypto/sha256"
-	"encoding/json"
-	"fmt"
-	"os"
-	"os/exec"
-	"path/filepath"
-	"sort"
-	"strings"
-	"time"
-)
-
-// batchInfo is what the mode itself knows before any run happens. Everything
-// else in the manifest is derived from the runs that were written, so a figure
-// in it cannot drift from the artifact beside it.
-type batchInfo struct {
-	Mode    string
-	Planned int
-	Started time.Time
-	Dry     bool
-	// Repo is the checkout the batch ran from, captured before the harness
-	// isolates its workspace: git cannot describe a directory that is gone.
-	Repo string
-}
-
-// batchManifest is written beside results.jsonl so a later reader does not have
-// to reconstruct what the batch was from a terminal scrollback.
-type batchManifest struct {
-	SchemaVersion int      `json:"schema_version"`
-	Mode          string   `json:"mode"`
-	StartedAt     string   `json:"started_at"`
-	FinishedAt    string   `json:"finished_at"`
-	GitCommit     string   `json:"git_commit"`
-	GitDirty      *bool    `json:"git_dirty,omitempty"`
-	Model         string   `json:"model"`
-	Window        int      `json:"window"`
-	Generations   int      `json:"generations"`
-	TaskIDs       []string `json:"task_ids"`
-	Arms          []string `json:"arms"`
-	PlannedRuns   int      `json:"planned_runs"`
-	CompletedRuns int      `json:"completed_runs"`
-	// FailedRuns counts the planned runs that ended before a transcript existed.
-	// They are rows in results.jsonl, never part of CompletedRuns, and the field
-	// is absent when there were none so an intact batch is unchanged.
-	FailedRuns int `json:"failed_runs,omitempty"`
-	// Usage totals are absent when nothing reported any: an unmeasured cost must
-	// not read as a measured zero. UsageReportedRuns stays, because zero there
-	// is a fact about the batch.
-	UsageReportedRuns     int    `json:"usage_reported_runs"`
-	UsageReportedRequests int    `json:"usage_reported_requests,omitempty"`
-	PromptTokens          int    `json:"prompt_tokens,omitempty"`
-	CompletionTokens      int    `json:"completion_tokens,omitempty"`
-	CacheHitTokens        int    `json:"cache_hit_tokens,omitempty"`
-	CacheMissTokens       int    `json:"cache_miss_tokens,omitempty"`
-	UsageEstimated        bool   `json:"usage_estimated,omitempty"`
-	ResultsSHA256         string `json:"results_sha256"`
-	TrajectoriesSHA256    string `json:"trajectories_sha256"`
-}
-
-// gitCommitAndDirty reports the checkout the batch ran from. Unknown is written
-// rather than omitted, so an absent field is never mistaken for a clean tree.
-func gitCommitAndDirty(repo string) (commit string, dirty *bool) {
-	commit = "unknown"
-	if repo == "" {
-		return commit, nil
-	}
-	in := func(args ...string) ([]byte, error) {
-		cmd := exec.Command("git", args...)
-		cmd.Dir = repo
-		return cmd.Output()
-	}
-	if out, err := in("rev-parse", "--short", "HEAD"); err == nil {
-		commit = strings.TrimSpace(string(out))
-	}
-	if out, err := in("status", "--porcelain"); err == nil {
-		state := strings.TrimSpace(string(out)) != ""
-		dirty = &state
-	}
-	return commit, dirty
-}
-
-func fileSHA256(path string) (string, error) {
-	data, err := os.ReadFile(path)
-	if err != nil {
-		return "", err
-	}
-	return fmt.Sprintf("%x", sha256.Sum256(data)), nil
-}
-
-func sortedKeys(set map[string]bool) []string {
-	out := make([]string, 0, len(set))
-	for k := range set {
-		out = append(out, k)
-	}
-	sort.Strings(out)
-	return out
-}
-
-// writeBatchManifest records the batch beside the runs it describes. It runs
-// last, so its digests are of what is already on disk.
-func writeBatchManifest(root string, info batchInfo, all []contextMetrics) error {
-	ids, arms := map[string]bool{}, map[string]bool{}
-	var prompt, completion, cacheHit, cacheMiss, requests, reported, failed int
-	estimated := false
-	for _, m := range all {
-		// A failed run has no task, arm or evidence to contribute: it is counted
-		// as a failure below and is not part of what the batch measured.
-		if m.Failed {
-			failed++
-			continue
-		}
-		ids[m.Task] = true
-		arms[m.Arm] = true
-		if m.UsageReportedRequests > 0 {
-			reported++
-			requests += m.UsageReportedRequests
-			prompt += m.PromptTokens
-			completion += m.CompletionTokens
-			cacheHit += m.CacheHitTokens
-			cacheMiss += m.CacheMissTokens
-			estimated = estimated || m.UsageEstimated
-		}
-	}
-	commit, dirty := gitCommitAndDirty(info.Repo)
-	model := runModel
-	if info.Dry {
-		model = "scripted"
-	}
-	manifest := batchManifest{
-		SchemaVersion: 1, Mode: info.Mode,
-		StartedAt:  info.Started.UTC().Format(time.RFC3339),
-		FinishedAt: time.Now().UTC().Format(time.RFC3339),
-		GitCommit:  commit, GitDirty: dirty,
-		Model: model, Window: fixtureWindow, Generations: fixtureGenerations,
-		TaskIDs: sortedKeys(ids), Arms: sortedKeys(arms),
-		PlannedRuns: info.Planned, CompletedRuns: len(all) - failed, FailedRuns: failed,
-		UsageReportedRuns: reported, UsageReportedRequests: requests,
-		PromptTokens: prompt, CompletionTokens: completion,
-		CacheHitTokens: cacheHit, CacheMissTokens: cacheMiss, UsageEstimated: estimated,
-	}
-	var err error
-	if manifest.ResultsSHA256, err = fileSHA256(filepath.Join(root, "results.jsonl")); err != nil {
-		return err
-	}
-	if manifest.TrajectoriesSHA256, err = fileSHA256(filepath.Join(root, "trajectories.jsonl")); err != nil {
-		return err
-	}
-	data, err := json.MarshalIndent(manifest, "", "  ")
-	if err != nil {
-		return err
-	}
-	return os.WriteFile(filepath.Join(root, "batch.json"), append(data, '\n'), 0o644)
-}
@@ -1,91 +0,0 @@
-package main
-
-import (
-	"crypto/sha256"
-	"encoding/json"
-	"fmt"
-	"os"
-	"path/filepath"
-	"strings"
-	"testing"
-	"time"
-)
-
-func seedArtifacts(t *testing.T, root string) {
-	t.Helper()
-	for _, name := range []string{"results.jsonl", "trajectories.jsonl"} {
-		if err := os.WriteFile(filepath.Join(root, name), []byte("{}\n"), 0o644); err != nil {
-			t.Fatal(err)
-		}
-	}
-}
-
-func readManifest(t *testing.T, root string) (batchManifest, string) {
-	t.Helper()
-	data, err := os.ReadFile(filepath.Join(root, "batch.json"))
-	if err != nil {
-		t.Fatal(err)
-	}
-	var got batchManifest
-	if err := json.Unmarshal(data, &got); err != nil {
-		t.Fatal(err)
-	}
-	return got, string(data)
-}
-
-func TestBatchManifestRecordsWhatTheBatchWas(t *testing.T) {
-	root := t.TempDir()
-	seedArtifacts(t, root)
-	all := []contextMetrics{
-		{Task: "b", Arm: "index-off"},
-		{Task: "a", Arm: "index-off"},
-		{Task: "a", Arm: "index-half", UsageReportedRequests: 2, PromptTokens: 100,
-			CompletionTokens: 10, CacheHitTokens: 60, CacheMissTokens: 40},
-	}
-	if err := writeBatchManifest(root, batchInfo{Mode: "run-index", Planned: 4, Started: time.Now().Add(-time.Minute)}, all); err != nil {
-		t.Fatal(err)
-	}
-	got, _ := readManifest(t, root)
-	if got.SchemaVersion != 1 || got.Mode != "run-index" || got.PlannedRuns != 4 || got.CompletedRuns != 3 {
-		t.Fatalf("manifest header = %+v", got)
-	}
-	if strings.Join(got.TaskIDs, ",") != "a,b" || strings.Join(got.Arms, ",") != "index-half,index-off" {
-		t.Fatalf("ids/arms = %v / %v, want sorted and unique", got.TaskIDs, got.Arms)
-	}
-	if got.UsageReportedRuns != 1 || got.UsageReportedRequests != 2 || got.PromptTokens != 100 ||
-		got.CompletionTokens != 10 || got.CacheHitTokens != 60 || got.CacheMissTokens != 40 {
-		t.Fatalf("usage totals = %+v", got)
-	}
-	// No repo means git could not describe the checkout: unknown commit, and no
-	// dirty flag at all, because "clean" is a claim this batch cannot make.
-	if got.GitCommit != "unknown" || got.GitDirty != nil {
-		t.Fatalf("git state = %q / %v, want unknown and no dirty flag", got.GitCommit, got.GitDirty)
-	}
-	want := fmt.Sprintf("%x", sha256.Sum256([]byte("{}\n")))
-	if got.ResultsSHA256 != want || got.TrajectoriesSHA256 != want {
-		t.Fatalf("digests = %q / %q, want %q", got.ResultsSHA256, got.TrajectoriesSHA256, want)
-	}
-}
-
-// A dry batch pays nothing, so its cost fields must be absent rather than zero:
-// an unmeasured spend must not read as a measured one.
-func TestBatchManifestLeavesUnreportedUsageAbsent(t *testing.T) {
-	root := t.TempDir()
-	seedArtifacts(t, root)
-	all := []contextMetrics{{Task: "a", Arm: "index-off"}, {Task: "a", Arm: "index-half"}}
-	if err := writeBatchManifest(root, batchInfo{Mode: "run-index", Planned: 2, Started: time.Now(), Dry: true}, all); err != nil {
-		t.Fatal(err)
-	}
-	got, raw := readManifest(t, root)
-	if got.UsageReportedRuns != 0 {
-		t.Fatalf("usage_reported_runs = %d, want 0 reported", got.UsageReportedRuns)
-	}
-	if got.Model != "scripted" {
-		t.Fatalf("dry batch model = %q, want scripted", got.Model)
-	}
-	for _, absent := range []string{"usage_reported_requests", "prompt_tokens", "completion_tokens", "cache_hit_tokens"} {
-		if strings.Contains(raw, absent) {
-			t.Fatalf("a dry batch wrote %q:\n%s", absent, raw)
-		}
-	}
-}
@@ -1,69 +0,0 @@
-package main
-
-import (
-	"strings"
-	"testing"
-)
-
-// boundaryRuns builds the input reportBoundaries reads: one scored run for an
-// index task's cue side, and one failed run for another index task's cue side.
-// The runs are constructed because the report is a pure function of the slice;
-// the batch path that produces them is covered by the failed-batch tests.
-func boundaryRuns(t *testing.T) (all []contextMetrics, scoredTask, failedTask string) {
-	t.Helper()
-	index := indexTasks()
-	if len(index) < 2 {
-		t.Fatalf("need at least two index tasks, have %d", len(index))
-	}
-	scored, failed := index[0], index[1]
-	cue, noCue := boundaryPair(scored.CueTier)
-	if cue == "" || noCue == "" {
-		t.Fatalf("task %s has an empty boundary pair (%q, %q)", scored.ID, cue, noCue)
-	}
-	failedCue, _ := boundaryPair(failed.CueTier)
-	if failedCue == "" {
-		t.Fatalf("task %s has no cue side", failed.ID)
-	}
-	all = []contextMetrics{
-		{Task: scored.ID, Arm: "index-" + cue, AnswerRecovered: true, SearchCalls: 1, RecallReturnedTokens: 318},
-		{Task: scored.ID, Arm: "index-" + noCue, AnswerRecovered: false},
-		{Task: failed.ID, Arm: "index-" + failedCue, Failed: true, FailureStage: stageNoEvidence},
-	}
-	return all, scored.ID, failed.ID
-}
-
-// TestBoundaryReportExcludesFailedRuns: a run that never produced a transcript
-// must not appear in the cue-boundary table, which was the last reporting exit
-// that still folded it in as a zero measurement.
-func TestBoundaryReportExcludesFailedRuns(t *testing.T) {
-	all, scoredTask, failedTask := boundaryRuns(t)
-	out := captureStdout(t, func() { reportBoundaries(all) })
-	if strings.Contains(out, failedTask) {
-		t.Fatalf("the boundary table printed the failed task %q as a measurement:\n%s", failedTask, out)
-	}
-	if !strings.Contains(out, scoredTask) {
-		t.Fatalf("the boundary table dropped the scored task %q, so the exclusion is not specific to failures:\n%s", scoredTask, out)
-	}
-	if !strings.Contains(out, "cue-present") || !strings.Contains(out, "cue-absent") {
-		t.Fatalf("the pooled comparison disappeared:\n%s", out)
-	}
-}
-
-// TestBoundaryReportSaysNothingToCompareWhenEveryRunFailed: with no scored run
-// the report must say so, rather than printing a comparison of zeros.
-func TestBoundaryReportSaysNothingToCompareWhenEveryRunFailed(t *testing.T) {
-	all, _, _ := boundaryRuns(t)
-	failedOnly := make([]contextMetrics, 0, len(all))
-	for _, m := range all {
-		m.Failed = true
-		m.FailureStage = stageNoEvidence
-		failedOnly = append(failedOnly, m)
-	}
-	out := captureStdout(t, func() { reportBoundaries(failedOnly) })
-	if !strings.Contains(out, "scored=0") || !strings.Contains(out, "failed=") {
-		t.Fatalf("an all-failed batch did not report scored=0 with a failed count:\n%s", out)
-	}
-	if strings.Contains(out, "cue-present") || strings.Contains(out, "cue-absent") {
-		t.Fatalf("an all-failed batch still printed a pooled comparison of zeros:\n%s", out)
-	}
-}
@@ -1,171 +0,0 @@
-// capture.go — what the agent actually hands a provider.
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"reasonix/internal/state/sessionstore"
-	"strings"
-
-	"reasonix/internal/assembly/boot"
-	"reasonix/internal/contract/ablation"
-	"reasonix/internal/contract/event"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/contract/tool"
-	"reasonix/internal/runtime/agent"
-	_ "reasonix/internal/tools/builtin"
-)
-
-// A projection is what the agent believes; a provider.Request is what the model
-// is told. Rendering, sanitization and tool-schema construction sit between
-// them, so the arm's contract is checked on the request.
-
-type captureProvider struct {
-	requests []provider.Request
-}
-
-func (p *captureProvider) Name() string { return "contextbench-capture" }
-
-// Stream records the request and answers with a clean terminal response. An
-// empty answer would buy an empty-final retry and a second request, which is
-// not the one round this is here to inspect.
-func (p *captureProvider) Stream(_ context.Context, req provider.Request) (<-chan provider.Chunk, error) {
-	p.requests = append(p.requests, cloneRequest(req))
-	ch := make(chan provider.Chunk, 2)
-	ch <- provider.Chunk{Type: provider.ChunkText, Text: "contextbench preflight complete"}
-	ch <- provider.Chunk{Type: provider.ChunkDone}
-	close(ch)
-	return ch, nil
-}
-
-func cloneRequest(req provider.Request) provider.Request {
-	out := req
-	out.Messages = append([]provider.Message(nil), req.Messages...)
-	out.Tools = append([]provider.ToolSchema(nil), req.Tools...)
-	return out
-}
-
-// captureRequest rebuilds the session the way a resume does — transcript from
-// disk, projection from its sidecar, the shipped tool surface at this arm —
-// and returns the request the first round hands the provider.
-func captureRequest(f builtFixture, arm ablation.Set, prompt string) (provider.Request, error) {
-	sess, err := sessionstore.LoadSession(f.Path)
-	if err != nil {
-		return provider.Request{}, fmt.Errorf("load session: %w", err)
-	}
-	reg := tool.NewRegistry()
-	for _, t := range tool.Builtins() {
-		reg.Add(t)
-	}
-	boot.ApplyUnifiedProviderToolSurface(reg, false, arm)
-
-	cap := &captureProvider{}
-	a := agent.New(cap, reg, sess, agent.Options{
-		ContextWindow: fixtureWindow, CompactRatio: 0.5, RecentKeep: 2,
-		SessionPath: f.Path, KeepPolicy: agent.KeepErrors, Ablation: arm,
-	}, event.Discard)
-	a.LoadProjectionSidecar(f.Path)
-
-	if err := a.Run(context.Background(), prompt); err != nil {
-		return provider.Request{}, fmt.Errorf("run: %w", err)
-	}
-	// The turn's own request is the one carrying tools. A fold on the way in
-	// asks the same provider for a digest, and that request has none — taking
-	// the first request blindly would inspect the summarizer's prompt.
-	for _, req := range cap.requests {
-		if len(req.Tools) > 0 {
-			return req, nil
-		}
-	}
-	return provider.Request{}, fmt.Errorf("no request carried tools (%d captured)", len(cap.requests))
-}
-
-// surfaceField is one place a model could read a string, named so a leak
-// report says where rather than only that.
-type surfaceField struct {
-	Path string
-	Text string
-}
-
-// providerSurface enumerates every provider-visible string in a request.
-// Enumerated rather than marshalled whole: a Message also carries local-only
-// fields, and scanning those would fail on bytes no model ever sees.
-func providerSurface(req provider.Request) []surfaceField {
-	var out []surfaceField
-	add := func(path, text string) {
-		if strings.TrimSpace(text) != "" {
-			out = append(out, surfaceField{Path: path, Text: text})
-		}
-	}
-	for i, m := range req.Messages {
-		add(fmt.Sprintf("messages[%d].content", i), m.Content)
-		add(fmt.Sprintf("messages[%d].reasoning_content", i), m.ReasoningContent)
-		add(fmt.Sprintf("messages[%d].name", i), m.Name)
-		for j, tc := range m.ToolCalls {
-			add(fmt.Sprintf("messages[%d].tool_calls[%d].name", i, j), tc.Name)
-			add(fmt.Sprintf("messages[%d].tool_calls[%d].arguments", i, j), tc.Arguments)
-		}
-		for j, item := range m.ResponsesItems {
-			add(fmt.Sprintf("messages[%d].responses_items[%d]", i, j), string(item))
-		}
-	}
-	for i, t := range req.Tools {
-		add(fmt.Sprintf("tools[%d].name", i), t.Name)
-		add(fmt.Sprintf("tools[%d].description", i), t.Description)
-		if len(t.Parameters) > 0 {
-			b, err := json.Marshal(t.Parameters)
-			if err == nil {
-				add(fmt.Sprintf("tools[%d].parameters", i), string(b))
-			}
-		}
-	}
-	return out
-}
-
-// findInSurface reports every place a marker is readable.
-func findInSurface(req provider.Request, marker string) []string {
-	var hits []string
-	for _, f := range providerSurface(req) {
-		if strings.Contains(f.Text, marker) {
-			hits = append(hits, f.Path)
-		}
-	}
-	return hits
-}
-
-// recallToolSchema returns the recall tool as the provider sees it.
-func recallToolSchema(req provider.Request) (provider.ToolSchema, bool) {
-	for _, t := range req.Tools {
-		if t.Name == "recall" {
-			return t, true
-		}
-	}
-	return provider.ToolSchema{}, false
-}
-
-// localOnlyLeaks reports host-side fields that should never survive the
-// provider boundary. A hit here is a boundary regression, marker or no marker.
-func localOnlyLeaks(req provider.Request) []string {
-	var out []string
-	for i, m := range req.Messages {
-		switch {
-		case m.RawContent != "":
-			out = append(out, fmt.Sprintf("messages[%d].raw_content", i))
-		case m.LocalOnly:
-			out = append(out, fmt.Sprintf("messages[%d].local_only", i))
-		case m.DecisionReceipt != nil || len(m.DecisionReceipts) > 0:
-			out = append(out, fmt.Sprintf("messages[%d].decision_receipt", i))
-		case m.ToolExecution != nil:
-			out = append(out, fmt.Sprintf("messages[%d].tool_execution", i))
-		case len(m.MemoryCitations) > 0:
-			out = append(out, fmt.Sprintf("messages[%d].memory_citations", i))
-		}
-	}
-	return out
-}
-
-// searchWording is what a search-off arm must never be told. The arm exists to
-// reproduce the runtime before search; an arm offered a capability it does not
-// have is a different experiment.
-var searchWording = []string{"recall with a query", "search the whole folded region", "query still finds"}
@@ -1,671 +0,0 @@
-// contextbench.go — what the model did with the affordance, read off structure.
-package main
-
-import (
-	"encoding/json"
-	"fmt"
-	"io"
-	"regexp"
-	"slices"
-	"strconv"
-	"strings"
-
-	"reasonix/internal/base/tokencount"
-	"reasonix/internal/contract/provider"
-)
-
-// Every judgement here reads a tool call's name and its JSON arguments, or a
-// canonical position, and never the model's prose. "I will search the history"
-// is not a search; a recall call with a query is.
-
-// failure stages, in funnel order. Exactly one is reported per run, except for
-// NoEvidence: that stage belongs to a run with no measurement at all.
-const (
-	stageReadMissed   = "NoSearch"           // never asked, and never read the target either
-	stageSearchMiss   = "SearchMiss"         // searched, target never came back
-	stageHitNotRead   = "HitNotRead"         // target was returned and never opened
-	stageAnswerWrong  = "ReadButAnswerWrong" // opened it and still answered wrong
-	stageRecovered    = "Recovered"
-	stageContaminated = "Contaminated"
-	stageNoEvidence   = "NoEvidence"
-
-	// Routing is judged by model round, not by the linear order of tool events:
-	// two calls in one round are one parallel fan-out, and calling that
-	// "workspace first" would report a choice the model never made.
-	routeMemoryFirst    = "MemoryFirst"
-	routeWorkspaceFirst = "WorkspaceFirst"
-	routeParallel       = "Parallel"
-	routeNeither        = "Neither"
-	stageCueRead        = "DirectRead"  // read the target from an index address, no search needed
-	stageNoRetrieval    = "NoRetrieval" // answered without touching recall at all
-	toolRecall          = "recall"
-)
-
-// searchAttempt is one query the model wrote, and what it got back. Kept per
-// attempt rather than aggregated: the question is why a retrieval that the host
-// probe ranks first takes the model more than one try.
-type searchAttempt struct {
-	Ordinal     int    `json:"ordinal"`
-	Round       int    `json:"round"`
-	Query       string `json:"query"`
-	TargetRank  int    `json:"target_rank"` // 0 = the target did not come back
-	ResultCount int    `json:"result_count"`
-	// Reformulation classifies this query against the one before it, and
-	// MissingProbeTerms are the host probe's terms this query left out. Both
-	// are derived from token sets: no judge, and nothing read out of prose.
-	Reformulation     string   `json:"reformulation,omitempty"`
-	MissingProbeTerms []string `json:"missing_probe_terms,omitempty"`
-}
-
-// contextMetrics is one run's behaviour.
-type contextMetrics struct {
-	Task string `json:"task"`
-	Arm  string `json:"arm"`
-
-	SearchCalls int `json:"search_calls"`
-	// SearchHits counts searches that returned anything; TargetSearchHits
-	// counts searches that returned the target. The gap is query quality.
-	SearchHits       int `json:"search_hits"`
-	TargetSearchHits int `json:"target_search_hits"`
-	FirstTargetRank  int `json:"first_target_rank"` // 0 = never returned
-	// FirstHitCoverage: how much of the answer the first hit handed over.
-	// rank=1 with coverage 1/3 is a window too short to answer with, not a
-	// model that would not stop.
-	FirstHitCoverage int `json:"first_hit_coverage"`
-	AnswerValues     int `json:"answer_values"`
-
-	ReadCalls    int  `json:"read_calls"`
-	TargetRead   bool `json:"target_read"`
-	ReadAfterHit bool `json:"read_after_hit"`
-	// RecallReadWithoutSearch: the index addressed the target, so no search was
-	// needed. Not "direct read" — read_file is a read too, and in an audit the
-	// two mean opposite things.
-	RecallReadWithoutSearch bool `json:"recall_read_without_search"`
-	SearchThenRecallRead    bool `json:"search_then_recall_read"`
-
-	FirstSearchRound int `json:"first_search_round"`
-	FirstReadRound   int `json:"first_read_round"`
-	RetrievalRounds  int `json:"retrieval_rounds"`
-
-	// RecallReturnedTokens is what recall paged back in. The index saves
-	// resident prefix; this is what is paid dynamically instead.
-	RecallReturnedTokens int `json:"recall_returned_tokens"`
-
-	// Provider usage as reported on the stream, or empty for a dry run. The count
-	// is completions that reported usage, so it is not a count of attempts.
-	UsageReportedRequests int  `json:"usage_reported_requests,omitempty"`
-	PromptTokens          int  `json:"prompt_tokens,omitempty"`
-	CompletionTokens      int  `json:"completion_tokens,omitempty"`
-	CacheHitTokens        int  `json:"cache_hit_tokens,omitempty"`
-	CacheMissTokens       int  `json:"cache_miss_tokens,omitempty"`
-	UsageEstimated        bool `json:"usage_estimated,omitempty"`
-
-	Searches []searchAttempt `json:"searches,omitempty"`
-	// PostHitSearches are queries issued after the target had already been
-	// returned: not a query problem, a result-to-read problem.
-	PostHitSearches int `json:"post_hit_searches"`
-
-	AnswerRecovered bool `json:"answer_recovered"`
-	// AnswerAmbiguous marks an answer that offered more than one candidate for
-	// a value. Markers matching proves the right fact appears; it does not
-	// prove it was submitted as the answer. Recorded, never acted on.
-	AnswerAmbiguous bool     `json:"answer_ambiguous,omitempty"`
-	AmbiguousVar    string   `json:"ambiguous_var,omitempty"`
-	MissingMarkers  []string `json:"missing_markers,omitempty"`
-	// FinalAnswer is kept bounded so a surprising run can be read back. A
-	// metric nobody can audit is a claim, not a measurement.
-	FinalAnswer   string `json:"final_answer,omitempty"`
-	FailureStage  string `json:"failure_stage"`
-	TimeoutReason string `json:"timeout_reason,omitempty"`
-
-	// Failed marks a run that ended before a transcript existed: it has no score
-	// and no place in the statistics. The row stays in results.jsonl with its
-	// task, arm and stage, so a failure is counted rather than dropped.
-	Failed bool   `json:"failed,omitempty"`
-	Reason string `json:"reason,omitempty"`
-
-	// UnexpectedWorkTools are investigation tools used instead of recall.
-	// Reaching for the workspace is behaviour, not noise: how a model routes
-	// its uncertainty when memory is missing is part of what is measured.
-	UnexpectedWorkTools     []string `json:"unexpected_work_tools,omitempty"`
-	EscapeCalls             int      `json:"escape_calls"`
-	EscapeBeforeFirstRecall int      `json:"escape_before_first_recall"`
-	EscapeAfterFirstRecall  int      `json:"escape_after_first_recall"`
-	// Contaminated: the answer was reachable through something other than
-	// recall. Asserting this beats trusting the isolation, and a run that trips
-	// it leaves the statistics rather than inflating one.
-	Contaminated bool   `json:"contaminated,omitempty"`
-	LeakedVia    string `json:"leaked_via,omitempty"`
-	LeakedMarker string `json:"leaked_marker,omitempty"`
-	LeakedArgs   string `json:"leaked_args,omitempty"`
-	LeakedOutput string `json:"leaked_output,omitempty"`
-	// EscapeFoundAnswer separates "looked at the workspace" from "found the
-	// answer there": different facts about the same call.
-	EscapeFoundAnswer bool `json:"escape_found_answer,omitempty"`
-
-	CueVisible bool `json:"cue_visible"`
-
-	// trajectory is kept out of the metric line and written beside it, so a
-	// future question can be asked of this batch instead of a new one.
-	trajectory []provider.Message
-
-	// CueDirectRead: the cue was addressed, the first memory action was a
-	// positions read, and it covered the target. Crediting any positions call
-	// would hand the index a search's work.
-	CueDirectRead bool `json:"cue_direct_read"`
-
-	// Sufficiency, not hit. A search that returns the target at rank 1 has not
-	// necessarily returned the answer: the snippet is bounded. The stopping
-	// boundary is the round where everything needed was readable.
-	FirstSufficientRound        int    `json:"first_sufficient_round"`
-	SufficientVia               string `json:"sufficient_via"` // snippet | read | never
-	RoundsAfterSufficient       int    `json:"rounds_after_sufficient"`
-	SearchesAfterSufficient     int    `json:"searches_after_sufficient"`
-	ReadsAfterSufficient        int    `json:"reads_after_sufficient"`
-	EscapesAfterSufficient      int    `json:"escapes_after_sufficient"`
-	RecallTokensAfterSufficient int    `json:"recall_tokens_after_sufficient"`
-	StoppingClass               string `json:"stopping_class"`
-
-	// What happened after the answer was already in hand. A timeout that spent
-	// seven rounds past its target is a stopping problem, and extending the
-	// budget would only buy the loop more time.
-	TargetFirstSeenRound    int  `json:"target_first_seen_round"`
-	RoundsAfterTarget       int  `json:"rounds_after_target"`
-	SearchesAfterTarget     int  `json:"searches_after_target"`
-	RecallTokensAfterTarget int  `json:"recall_tokens_after_target"`
-	TargetHeldAtEnd         bool `json:"target_held_at_end"`
-
-	// Routing is where the model looked first when its memory was missing.
-	Routing          string `json:"routing"`
-	FirstRecallRound int    `json:"first_recall_round"`
-	FirstEscapeRound int    `json:"first_escape_round"`
-}
-
-// hitPosition matches the address recall prints for a search hit. Reading our
-// own rendered format, not guessing at the model's words.
-var hitPosition = regexp.MustCompile(`(?m)^#(\d+) `)
-
-// scoreRun reads the messages one turn appended and reports what happened.
-// runScorer walks one turn's messages, holding the state a judgement needs:
-// which calls are open, whether the target has come back, and which search
-// each result belongs to.
-type runScorer struct {
-	m            contextMetrics
-	inst         fixtureInstance
-	target       int
-	calls        map[string]provider.ToolCall
-	pending      map[string]int
-	round        int
-	sawTargetHit bool
-	// seen accumulates everything recall has shown the model, so sufficiency is
-	// judged on what it could read rather than on which call produced it.
-	seen       strings.Builder
-	sufficient bool
-}
-
-// scoreRun reads the messages one turn appended and reports what happened.
-func scoreRun(msgs []provider.Message, inst fixtureInstance, target int, arm string, cueVisible bool) contextMetrics {
-	s := &runScorer{
-		m:    contextMetrics{Task: inst.Task.ID, Arm: arm, CueVisible: cueVisible},
-		inst: inst, target: target,
-		calls: map[string]provider.ToolCall{}, pending: map[string]int{},
-	}
-	for _, msg := range msgs {
-		if len(msg.ToolCalls) > 0 {
-			s.round++
-		}
-		for _, tc := range msg.ToolCalls {
-			s.calls[tc.ID] = tc
-			s.observeCall(tc)
-		}
-		if msg.Role == provider.RoleTool {
-			s.observeResult(msg)
-		}
-	}
-	return s.finish(cueVisible)
-}
-
-// observeCall records one tool call: a recall search, a recall read, or a
-// reach for the workspace.
-func (s *runScorer) observeCall(tc provider.ToolCall) {
-	m := &s.m
-	if tc.Name != toolRecall {
-		if !containsString(m.UnexpectedWorkTools, tc.Name) {
-			m.UnexpectedWorkTools = append(m.UnexpectedWorkTools, tc.Name)
-		}
-		m.EscapeCalls++
-		if s.sufficient {
-			m.EscapesAfterSufficient++
-		}
-		if m.FirstEscapeRound == 0 {
-			m.FirstEscapeRound = s.round
-		}
-		if m.SearchCalls+m.ReadCalls == 0 {
-			m.EscapeBeforeFirstRecall++
-		} else {
-			m.EscapeAfterFirstRecall++
-		}
-		return
-	}
-	if m.FirstRecallRound == 0 {
-		m.FirstRecallRound = s.round
-	}
-	query, positions := recallArgs(tc.Arguments)
-	switch {
-	case query != "":
-		m.SearchCalls++
-		if m.FirstSearchRound == 0 {
-			m.FirstSearchRound = s.round
-		}
-		if s.sawTargetHit {
-			m.PostHitSearches++
-			m.SearchesAfterTarget++
-		}
-		if s.sufficient {
-			m.SearchesAfterSufficient++
-		}
-		m.Searches = append(m.Searches, searchAttempt{Ordinal: m.SearchCalls, Round: s.round, Query: query})
-		s.pending[tc.ID] = len(m.Searches) - 1
-	case len(positions) > 0:
-		m.ReadCalls++
-		if s.sufficient {
-			m.ReadsAfterSufficient++
-		}
-		if m.FirstReadRound == 0 {
-			m.FirstReadRound = s.round
-		}
-		// A read covers the target when it names its position: recall returns
-		// the tool results under the caller's own address.
-		if containsInt(positions, s.target) {
-			m.TargetRead = true
-			switch {
-			case s.sawTargetHit:
-				m.ReadAfterHit, m.SearchThenRecallRead = true, true
-			case m.SearchCalls == 0:
-				m.RecallReadWithoutSearch = true
-			}
-		}
-	}
-}
-
-// observeResult reads what a call returned: contamination, page-in cost, and
-// whether this search brought the target back.
-func (s *runScorer) observeResult(msg provider.Message) {
-	m := &s.m
-	call, ok := s.calls[msg.ToolCallID]
-	if !ok {
-		return
-	}
-	if call.Name != toolRecall {
-		// Codenames only: a three-digit epoch appears in any long file by
-		// chance, and one did. A model holding the codename holds the numbers
-		// beside it anyway.
-		for _, marker := range s.inst.AnswerMarkers {
-			if len(marker) >= 6 && strings.Contains(msg.Content, marker) {
-				m.Contaminated, m.LeakedVia, m.LeakedMarker = true, call.Name, marker
-				m.EscapeFoundAnswer = true
-				m.LeakedArgs, m.LeakedOutput = truncate(call.Arguments, 200), truncate(msg.Content, 300)
-			}
-		}
-		return
-	}
-	cost := tokencount.Text(msg.Content)
-	m.RecallReturnedTokens += cost
-	if s.sawTargetHit {
-		m.RecallTokensAfterTarget += cost
-	}
-	if s.sufficient {
-		m.RecallTokensAfterSufficient += cost
-	}
-	// Everything recall showed the model, judged as a whole: the answer may
-	// arrive across a snippet and a read, and either way it is readable.
-	s.seen.WriteString(msg.Content)
-	s.seen.WriteString("\n")
-	if !s.sufficient && allMarkersIn(s.seen.String(), s.inst.AnswerMarkers) {
-		s.sufficient = true
-		m.FirstSufficientRound = s.round
-		query, _ := recallArgs(call.Arguments)
-		m.SufficientVia = "read"
-		if query != "" {
-			m.SufficientVia = "snippet"
-		}
-	}
-	if query, _ := recallArgs(call.Arguments); query == "" {
-		return
-	}
-	positions := searchHitPositions(msg.Content)
-	if len(positions) > 0 {
-		m.SearchHits++
-	}
-	rank := indexOfInt(positions, s.target)
-	if i, ok := s.pending[msg.ToolCallID]; ok {
-		m.Searches[i].TargetRank, m.Searches[i].ResultCount = rank, len(positions)
-	}
-	if rank == 0 {
-		return
-	}
-	m.TargetSearchHits++
-	if !s.sawTargetHit {
-		m.TargetFirstSeenRound = s.round
-		m.FirstHitCoverage = answerCoverage(msg.Content, s.inst.AnswerMarkers)
-	}
-	s.sawTargetHit = true
-	if m.FirstTargetRank == 0 {
-		m.FirstTargetRank = rank
-	}
-}
-
-// finish derives what only the whole walk can say.
-func (s *runScorer) finish(cueVisible bool) contextMetrics {
-	m := s.m
-	m.RetrievalRounds = m.SearchCalls + m.ReadCalls
-	m.TargetHeldAtEnd = s.sawTargetHit || m.TargetRead
-	if m.TargetFirstSeenRound > 0 {
-		m.RoundsAfterTarget = s.round - m.TargetFirstSeenRound
-	}
-	m.Routing = routing(m.FirstRecallRound, m.FirstEscapeRound)
-	m.CueDirectRead = cueVisible && m.RecallReadWithoutSearch
-	if m.SufficientVia == "" {
-		m.SufficientVia = "never"
-	}
-	if m.FirstSufficientRound > 0 {
-		m.RoundsAfterSufficient = s.round - m.FirstSufficientRound
-	}
-	m.AnswerValues = len(s.inst.AnswerMarkers)
-	m.StoppingClass = stoppingClass(m)
-	classifySearches(&m, s.inst)
-	return m
-}
-
-// scoreAnswer grades the final text on the task's markers. Nonces, matched
-// exactly: no judge model, and nothing a general answer could satisfy.
-func (m *contextMetrics) scoreAnswer(final string, inst fixtureInstance) {
-	lower := strings.ToLower(final)
-	for _, marker := range inst.AnswerMarkers {
-		if !strings.Contains(lower, strings.ToLower(marker)) {
-			m.MissingMarkers = append(m.MissingMarkers, marker)
-		}
-	}
-	m.AnswerRecovered = len(m.MissingMarkers) == 0
-	m.AnswerAmbiguous, m.AmbiguousVar = ambiguousAnswer(final, inst)
-	m.FailureStage = m.stage()
-	if len(final) > 1200 {
-		final = final[:1200] + "…"
-	}
-	m.FinalAnswer = final
-}
-
-// stage places the run in the funnel. One label, chosen at the first step that
-// did not happen, so a table of stages reads as where the retrieval broke.
-func (m contextMetrics) stage() string {
-	switch {
-	case m.Contaminated:
-		return stageContaminated
-	case m.AnswerRecovered && m.RecallReadWithoutSearch:
-		return stageCueRead
-	case m.AnswerRecovered:
-		return stageRecovered
-	case m.TargetRead:
-		return stageAnswerWrong
-	case m.TargetSearchHits > 0:
-		return stageHitNotRead
-	case m.SearchCalls > 0:
-		return stageSearchMiss
-	case m.ReadCalls > 0:
-		return stageReadMissed
-	default:
-		return stageNoRetrieval
-	}
-}
-
-// recallArgs pulls the two fields that say which operation a recall call is.
-func recallArgs(args string) (query string, positions []int) {
-	var parsed struct {
-		Query     string `json:"query"`
-		Positions []int  `json:"positions"`
-	}
-	if err := json.Unmarshal([]byte(args), &parsed); err != nil {
-		return "", nil
-	}
-	return strings.TrimSpace(parsed.Query), parsed.Positions
-}
-
-// searchHitPositions reads the addresses out of a search result, in rank order.
-func searchHitPositions(text string) []int {
-	var out []int
-	for _, match := range hitPosition.FindAllStringSubmatch(text, -1) {
-		if n, err := strconv.Atoi(match[1]); err == nil {
-			out = append(out, n)
-		}
-	}
-	return out
-}
-
-func containsInt(list []int, want int) bool { return slices.Contains(list, want) }
-
-// indexOfInt returns a 1-based rank, or 0 when absent.
-func indexOfInt(list []int, want int) int {
-	for i, v := range list {
-		if v == want {
-			return i + 1
-		}
-	}
-	return 0
-}
-
-func containsString(list []string, want string) bool { return slices.Contains(list, want) }
-
-// funnel summarizes one arm's runs the way the experiment is read.
-type funnel struct {
-	Arm                                       string
-	Runs                                      int
-	Scored, Contaminated, Errored             int
-	Searched, TargetHit, TargetRead, Answered int
-	CueReads                                  int
-	SearchCalls, ReadCalls, RecallTokens      int
-	Escapes, EscapeCalls                      int
-	Stages, Routes                            map[string]int
-}
-
-func summarize(arm string, runs []contextMetrics) funnel {
-	f := funnel{Arm: arm, Runs: len(runs), Stages: map[string]int{}, Routes: map[string]int{}}
-	for _, r := range runs {
-		// A run that never produced a transcript is not a scored run and has no
-		// stage to place: counting it would put an empty name in the funnel, the
-		// stage table and the routes.
-		if r.Failed {
-			continue
-		}
-		if r.Contaminated {
-			f.Contaminated++
-			f.Stages[r.FailureStage]++
-			continue
-		}
-		if r.FailureStage == "RunError" {
-			f.Errored++
-			f.Stages[r.FailureStage]++
-			continue
-		}
-		f.Scored++
-		if r.SearchCalls > 0 {
-			f.Searched++
-		}
-		if r.TargetSearchHits > 0 {
-			f.TargetHit++
-		}
-		if r.TargetRead {
-			f.TargetRead++
-		}
-		if r.AnswerRecovered {
-			f.Answered++
-		}
-		if r.RecallReadWithoutSearch {
-			f.CueReads++
-		}
-		if len(r.UnexpectedWorkTools) > 0 {
-			f.Escapes++
-		}
-		f.EscapeCalls += r.EscapeCalls
-		f.SearchCalls += r.SearchCalls
-		f.ReadCalls += r.ReadCalls
-		f.RecallTokens += r.RecallReturnedTokens
-		f.Stages[r.FailureStage]++
-		if r.FailureStage != "RunError" {
-			f.Routes[r.Routing]++
-		}
-	}
-	return f
-}
-
-func (f funnel) line() string {
-	per := func(n int) float64 {
-		if f.Scored == 0 {
-			return 0
-		}
-		return float64(n) / float64(f.Scored)
-	}
-	line := fmt.Sprintf("%-14s runs=%-3d answered=%d/%d  searched=%d  target-hit=%d  cue-read=%d  search/task=%.2f  recall-tok/task=%.0f  escape/task=%.2f",
-		f.Arm, f.Scored, f.Answered, f.Scored, f.Searched, f.TargetHit, f.CueReads,
-		per(f.SearchCalls), per(f.RecallTokens), per(f.EscapeCalls))
-	if f.Contaminated > 0 {
-		line += fmt.Sprintf("  CONTAMINATED=%d", f.Contaminated)
-	}
-	if f.Errored > 0 {
-		line += fmt.Sprintf("  ERRORED=%d", f.Errored)
-	}
-	return line
-}
-
-func writeJSONLine(w io.Writer, v any) error {
-	b, err := json.Marshal(v)
-	if err != nil {
-		return err
-	}
-	_, err = fmt.Fprintln(w, string(b))
-	return err
-}
-
-func truncate(s string, n int) string {
-	if len(s) <= n {
-		return s
-	}
-	return s[:n] + "…"
-}
-
-// routing says where the model looked first, comparing rounds rather than
-// event order. Same round is a parallel fan-out, not a preference.
-func routing(recallRound, escapeRound int) string {
-	switch {
-	case recallRound == 0 && escapeRound == 0:
-		return routeNeither
-	case recallRound == 0:
-		return routeWorkspaceFirst
-	case escapeRound == 0:
-		return routeMemoryFirst
-	case recallRound < escapeRound:
-		return routeMemoryFirst
-	case escapeRound < recallRound:
-		return routeWorkspaceFirst
-	default:
-		return routeParallel
-	}
-}
-
-// ambiguousAnswer reports an answer naming more than one candidate of a
-// codename's shape. Only codenames are checked: they carry a stem this run
-// declared, so two distinct "fence-…" values in one answer is one answer too
-// many. Numbers have no such shape and are left alone.
-func ambiguousAnswer(final string, inst fixtureInstance) (bool, string) {
-	for _, spec := range inst.Task.Vars {
-		if spec.Kind != varCodename {
-			continue
-		}
-		seen := map[string]bool{}
-		for _, match := range regexp.MustCompile(regexp.QuoteMeta(spec.Word)+`-[a-z0-9]{3,}`).FindAllString(final, -1) {
-			seen[match] = true
-		}
-		if len(seen) > 1 {
-			return true, spec.Name
-		}
-	}
-	return false, ""
-}
-
-func allMarkersIn(text string, markers []string) bool {
-	if len(markers) == 0 {
-		return false
-	}
-	lower := strings.ToLower(text)
-	for _, marker := range markers {
-		if !strings.Contains(lower, strings.ToLower(marker)) {
-			return false
-		}
-	}
-	return true
-}
-
-// Stopping classes. The boundary is sufficiency, not the target coming back:
-// a rank-1 hit whose snippet was too short to answer leaves a read entirely
-// justified, and counting that as waste would blame the model for the bound.
-const (
-	stopSnippet       = "SnippetStop"             // enough arrived, and the turn ended
-	stopSnippetRead   = "SnippetThenRead"         // the snippet was short; the read finished it
-	stopDefensiveRead = "DefensiveRead"           // enough had arrived, one read confirmed it
-	stopPostRetrieval = "PostSufficientRetrieval" // enough had arrived, and retrieval continued
-	stopRunaway       = "PostSufficientRunaway"   // it continued for rounds, or ran out of budget
-	stopNeverEnough   = "NeverSufficient"
-)
-
-// runawayRounds is where continuing stops looking like confirmation.
-const runawayRounds = 3
-
-func stoppingClass(m contextMetrics) string {
-	if m.FirstSufficientRound == 0 {
-		return stopNeverEnough
-	}
-	after := m.SearchesAfterSufficient + m.ReadsAfterSufficient + m.EscapesAfterSufficient
-	switch {
-	case after == 0 && m.SufficientVia == "snippet":
-		return stopSnippet
-	case after == 0:
-		return stopSnippetRead
-	case m.RoundsAfterSufficient >= runawayRounds || m.FailureStage == "RunError":
-		return stopRunaway
-	case after == 1 && m.ReadsAfterSufficient == 1:
-		return stopDefensiveRead
-	default:
-		return stopPostRetrieval
-	}
-}
-
-// reportStopping is the funnel that matters once retrieval succeeds: what the
-// turn did after it already had everything it needed.
-func reportStopping(all []contextMetrics) string {
-	classes := map[string]int{}
-	var extraRounds, extraTokens []float64
-	sufficient, answeredNext := 0, 0
-	for _, m := range all {
-		if m.Failed || m.Contaminated {
-			continue
-		}
-		classes[m.StoppingClass]++
-		if m.FirstSufficientRound == 0 {
-			continue
-		}
-		sufficient++
-		extraRounds = append(extraRounds, float64(m.RoundsAfterSufficient))
-		extraTokens = append(extraTokens, float64(m.RecallTokensAfterSufficient))
-		if m.RoundsAfterSufficient <= 1 {
-			answeredNext++
-		}
-	}
-	var b strings.Builder
-	b.WriteString("\n## After the evidence was already enough\n")
-	fmt.Fprintf(&b, "  sufficient in %d runs; ended within one round in %d\n", sufficient, answeredNext)
-	fmt.Fprintf(&b, "  extra rounds  mean %.1f  median %.1f  p90 %.0f  max %.0f\n",
-		mean(extraRounds), median(extraRounds), percentile(extraRounds, 0.9), maxOf(extraRounds))
-	fmt.Fprintf(&b, "  extra recall tokens  mean %.0f  median %.0f  p90 %.0f  max %.0f\n",
-		mean(extraTokens), median(extraTokens), percentile(extraTokens, 0.9), maxOf(extraTokens))
-	b.WriteString("  " + countsLine(classes) + "\n")
-	return b.String()
-}
@@ -1,337 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"strings"
-	"testing"
-
-	"reasonix/internal/contract/provider"
-)
-
-// The scorer decides what the pilot concludes, so it is held to the same rule
-// the corpus is: judgements read structure. A model that says it will search
-// has not searched; a recall call carrying a query has.
-
-const scorerTarget = 42
-
-func recallCall(id, args string) provider.Message {
-	return provider.Message{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-		{ID: id, Name: "recall", Arguments: args},
-	}}
-}
-
-func toolResult(id, text string) provider.Message {
-	return provider.Message{Role: provider.RoleTool, ToolCallID: id, Name: "recall", Content: text}
-}
-
-func hitBlock(positions ...int) string {
-	var b strings.Builder
-	b.WriteString("Folded context matching \"x\":\n")
-	for _, p := range positions {
-		fmt.Fprintf(&b, "\n#%d assistant_text\nsome snippet here\n", p)
-	}
-	return b.String()
-}
-
-func scorerTask() fixtureInstance {
-	return fixtureInstance{Task: contextTask{ID: "t"}, AnswerMarkers: []string{"cobalt-lark-17"}}
-}
-
-func score(t *testing.T, msgs []provider.Message, final string) contextMetrics {
-	t.Helper()
-	m := scoreRun(msgs, scorerTask(), scorerTarget, "arm", false)
-	m.scoreAnswer(final, scorerTask())
-	return m
-}
-
-// The whole funnel, walked end to end.
-func TestScorerReadsTheFullRetrievalChain(t *testing.T) {
-	m := score(t, []provider.Message{
-		recallCall("c1", `{"query":"retry boundary"}`),
-		toolResult("c1", hitBlock(7, scorerTarget)),
-		recallCall("c2", fmt.Sprintf(`{"positions":[%d]}`, scorerTarget)),
-		toolResult("c2", "#42\n[assistant]\nthe ownership token is cobalt-lark-17"),
-		{Role: provider.RoleAssistant, Content: "The token was cobalt-lark-17."},
-	}, "The token was cobalt-lark-17.")
-
-	if m.SearchCalls != 1 || m.ReadCalls != 1 {
-		t.Errorf("calls = %d search / %d read, want 1/1", m.SearchCalls, m.ReadCalls)
-	}
-	if m.TargetSearchHits != 1 || m.FirstTargetRank != 2 {
-		t.Errorf("target hit %d at rank %d, want 1 at rank 2", m.TargetSearchHits, m.FirstTargetRank)
-	}
-	if !m.TargetRead || !m.ReadAfterHit || m.RecallReadWithoutSearch {
-		t.Errorf("read flags = read %v afterHit %v direct %v", m.TargetRead, m.ReadAfterHit, m.RecallReadWithoutSearch)
-	}
-	if !m.AnswerRecovered || m.FailureStage != stageRecovered {
-		t.Errorf("stage = %q recovered=%v, want Recovered", m.FailureStage, m.AnswerRecovered)
-	}
-	if m.RecallReturnedTokens == 0 {
-		t.Error("recall returned tokens were not counted")
-	}
-}
-
-// An index arm's claim: the address was already on screen, so no search was
-// needed. That is a different outcome from finding it, and gets its own stage.
-func TestScorerSeparatesCueReadFromSearch(t *testing.T) {
-	m := score(t, []provider.Message{
-		recallCall("c1", fmt.Sprintf(`{"positions":[%d]}`, scorerTarget)),
-		toolResult("c1", "#42\nthe ownership token is cobalt-lark-17"),
-		{Role: provider.RoleAssistant, Content: "cobalt-lark-17"},
-	}, "cobalt-lark-17")
-
-	if m.SearchCalls != 0 || !m.RecallReadWithoutSearch {
-		t.Errorf("search=%d direct=%v, want 0 and true", m.SearchCalls, m.RecallReadWithoutSearch)
-	}
-	if m.FailureStage != stageCueRead {
-		t.Errorf("stage = %q, want %q", m.FailureStage, stageCueRead)
-	}
-}
-
-// Every way the chain can break gets its own label, so a failed run says where
-// rather than only that.
-func TestScorerNamesWhereTheChainBroke(t *testing.T) {
-	for _, tc := range []struct {
-		name  string
-		msgs  []provider.Message
-		final string
-		want  string
-	}{
-		{
-			name:  "never touched recall",
-			msgs:  []provider.Message{{Role: provider.RoleAssistant, Content: "I do not recall that."}},
-			final: "I do not recall that.",
-			want:  stageNoRetrieval,
-		},
-		{
-			name: "searched and missed",
-			msgs: []provider.Message{
-				recallCall("c1", `{"query":"unrelated words"}`),
-				toolResult("c1", hitBlock(3, 9)),
-			},
-			final: "not found",
-			want:  stageSearchMiss,
-		},
-		{
-			name: "found it and never opened it",
-			msgs: []provider.Message{
-				recallCall("c1", `{"query":"retry boundary"}`),
-				toolResult("c1", hitBlock(scorerTarget)),
-			},
-			final: "probably something about retries",
-			want:  stageHitNotRead,
-		},
-		{
-			name: "read it and answered wrong",
-			msgs: []provider.Message{
-				recallCall("c1", fmt.Sprintf(`{"positions":[%d]}`, scorerTarget)),
-				toolResult("c1", "#42\nthe token is elsewhere"),
-			},
-			final: "I think it was amber-9",
-			want:  stageAnswerWrong,
-		},
-		{
-			name: "read the wrong position",
-			msgs: []provider.Message{
-				recallCall("c1", `{"positions":[7]}`),
-				toolResult("c1", "#7\nnothing useful"),
-			},
-			final: "unclear",
-			want:  stageReadMissed,
-		},
-	} {
-		t.Run(tc.name, func(t *testing.T) {
-			if got := score(t, tc.msgs, tc.final).FailureStage; got != tc.want {
-				t.Errorf("stage = %q, want %q", got, tc.want)
-			}
-		})
-	}
-}
-
-// The answer lives only in folded history, so reaching for the workspace is a
-// strategy change worth seeing — including when it happens to work.
-func TestScorerRecordsWorkspaceEscapes(t *testing.T) {
-	m := score(t, []provider.Message{
-		{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-			{ID: "g1", Name: "bash", Arguments: `{"command":"rg cobalt-lark"}`},
-			{ID: "g2", Name: "read_file", Arguments: `{"path":"notes.md"}`},
-		}},
-		{Role: provider.RoleAssistant, Content: "cobalt-lark-17"},
-	}, "cobalt-lark-17")
-
-	if len(m.UnexpectedWorkTools) != 2 {
-		t.Fatalf("escapes = %v, want bash and read_file", m.UnexpectedWorkTools)
-	}
-	if m.SearchCalls != 0 || m.ReadCalls != 0 {
-		t.Errorf("a workspace tool was counted as recall: search=%d read=%d", m.SearchCalls, m.ReadCalls)
-	}
-	// It answered, so the run is Recovered — the escape is reported beside the
-	// outcome rather than overriding it.
-	if !m.AnswerRecovered {
-		t.Error("a correct answer was not scored as recovered")
-	}
-}
-
-// Narration is not action.
-func TestScorerIgnoresWhatTheModelSaysItWillDo(t *testing.T) {
-	m := score(t, []provider.Message{
-		{Role: provider.RoleAssistant, Content: "Let me search the folded context with recall for the retry boundary."},
-	}, "Let me search the folded context with recall for the retry boundary.")
-	if m.SearchCalls != 0 || m.FailureStage != stageNoRetrieval {
-		t.Errorf("prose was scored as a search: calls=%d stage=%q", m.SearchCalls, m.FailureStage)
-	}
-}
-
-// Every marker must land, so a partially right answer is not a recovery.
-func TestScorerRequiresEveryMarker(t *testing.T) {
-	task := fixtureInstance{Task: contextTask{ID: "t"}, AnswerMarkers: []string{"comet-42", "per-lineage"}}
-	m := scoreRun(nil, task, scorerTarget, "arm", false)
-	m.scoreAnswer("the salt was comet-42", task)
-	if m.AnswerRecovered {
-		t.Error("a half answer scored as recovered")
-	}
-	if len(m.MissingMarkers) != 1 || m.MissingMarkers[0] != "per-lineage" {
-		t.Errorf("missing = %v, want per-lineage", m.MissingMarkers)
-	}
-}
-
-// An answer reachable through anything but recall invalidates the run. The
-// workspace is isolated so this should never fire; asserting it beats trusting
-// it, and a tripped run leaves the statistics rather than inflating one.
-func TestScorerInvalidatesRunsThatLeakedThroughAnotherTool(t *testing.T) {
-	m := score(t, []provider.Message{
-		{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-			{ID: "g1", Name: "bash", Arguments: `{"command":"rg -r ."}`},
-		}},
-		{Role: provider.RoleTool, ToolCallID: "g1", Name: "bash", Content: "corpus.go: the token is cobalt-lark-17"},
-		{Role: provider.RoleAssistant, Content: "cobalt-lark-17"},
-	}, "cobalt-lark-17")
-
-	if !m.Contaminated || m.LeakedVia != "bash" || m.LeakedMarker != "cobalt-lark-17" {
-		t.Fatalf("leak not caught: contaminated=%v via=%q marker=%q", m.Contaminated, m.LeakedVia, m.LeakedMarker)
-	}
-	if m.FailureStage != stageContaminated {
-		t.Errorf("stage = %q, want %q even though every marker matched", m.FailureStage, stageContaminated)
-	}
-	f := summarize("arm", []contextMetrics{m})
-	if f.Scored != 0 || f.Answered != 0 || f.Contaminated != 1 {
-		t.Errorf("a contaminated run entered the statistics: %+v", f)
-	}
-}
-
-// Where an escape happened matters: before any recall is a different strategy
-// from after one came back short.
-func TestScorerPlacesEscapesRelativeToRecall(t *testing.T) {
-	m := score(t, []provider.Message{
-		{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-			{ID: "g1", Name: "bash", Arguments: `{"command":"ls"}`},
-		}},
-		{Role: provider.RoleTool, ToolCallID: "g1", Name: "bash", Content: "README.md"},
-		recallCall("c1", `{"query":"retry boundary"}`),
-		toolResult("c1", hitBlock(scorerTarget)),
-		{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-			{ID: "g2", Name: "read_file", Arguments: `{"path":"README.md"}`},
-		}},
-		{Role: provider.RoleTool, ToolCallID: "g2", Name: "read_file", Content: "nothing here"},
-	}, "unclear")
-
-	if m.EscapeCalls != 2 || m.EscapeBeforeFirstRecall != 1 || m.EscapeAfterFirstRecall != 1 {
-		t.Errorf("escapes = %d (%d before, %d after), want 2 (1, 1)",
-			m.EscapeCalls, m.EscapeBeforeFirstRecall, m.EscapeAfterFirstRecall)
-	}
-}
-
-// Routing is judged by round. Two calls in one round are a fan-out, and
-// reporting that as a preference would invent a decision the model never made.
-func TestRoutingIsJudgedByRoundNotEventOrder(t *testing.T) {
-	parallel := []provider.Message{
-		{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-			{ID: "g1", Name: "bash", Arguments: `{"command":"ls"}`},
-			{ID: "c1", Name: "recall", Arguments: `{"query":"retry boundary"}`},
-		}},
-		{Role: provider.RoleTool, ToolCallID: "g1", Name: "bash", Content: "README.md"},
-		toolResult("c1", hitBlock(scorerTarget)),
-	}
-	if got := score(t, parallel, "unclear").Routing; got != routeParallel {
-		t.Errorf("routing = %q for one round holding both, want %q", got, routeParallel)
-	}
-
-	workspaceFirst := []provider.Message{
-		{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-			{ID: "g1", Name: "bash", Arguments: `{"command":"ls"}`},
-		}},
-		{Role: provider.RoleTool, ToolCallID: "g1", Name: "bash", Content: "README.md"},
-		recallCall("c1", `{"query":"retry boundary"}`),
-		toolResult("c1", hitBlock(scorerTarget)),
-	}
-	if got := score(t, workspaceFirst, "unclear").Routing; got != routeWorkspaceFirst {
-		t.Errorf("routing = %q for workspace in an earlier round, want %q", got, routeWorkspaceFirst)
-	}
-
-	memoryFirst := []provider.Message{
-		recallCall("c1", `{"query":"retry boundary"}`),
-		toolResult("c1", hitBlock(scorerTarget)),
-		{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-			{ID: "g1", Name: "bash", Arguments: `{"command":"ls"}`},
-		}},
-		{Role: provider.RoleTool, ToolCallID: "g1", Name: "bash", Content: "README.md"},
-	}
-	if got := score(t, memoryFirst, "unclear").Routing; got != routeMemoryFirst {
-		t.Errorf("routing = %q for recall in an earlier round, want %q", got, routeMemoryFirst)
-	}
-
-	if got := score(t, nil, "no idea").Routing; got != routeNeither {
-		t.Errorf("routing = %q with no tools at all, want %q", got, routeNeither)
-	}
-}
-
-// Markers matching proves the right fact appears somewhere. An answer that
-// hedges between two candidates contains it and has not submitted it.
-func TestScorerFlagsAnAnswerThatNamesTwoCandidates(t *testing.T) {
-	inst := fixtureInstance{
-		Task:          contextTask{ID: "t", Vars: []varSpec{code("marker", "fence")}},
-		AnswerMarkers: []string{"fence-qmwnv"},
-	}
-	m := scoreRun(nil, inst, 0, "arm", false)
-	m.scoreAnswer("It was fence-qmwnv, though it may have been fence-abcde.", inst)
-	if !m.AnswerRecovered {
-		t.Error("the marker is present; recovery should still read true")
-	}
-	if !m.AnswerAmbiguous || m.AmbiguousVar != "marker" {
-		t.Errorf("ambiguity not flagged: %v %q", m.AnswerAmbiguous, m.AmbiguousVar)
-	}
-
-	clean := scoreRun(nil, inst, 0, "arm", false)
-	clean.scoreAnswer("The marker was fence-qmwnv.", inst)
-	if clean.AnswerAmbiguous {
-		t.Error("a single candidate was called ambiguous")
-	}
-}
-
-// The index arm's claim is narrow: the cue was addressed, the first memory
-// action was a positions read, and it covered the target. A positions call
-// after a search located the target credits the index for the search's work.
-func TestCueDirectReadRequiresTheCueAndTheFirstAction(t *testing.T) {
-	direct := []provider.Message{
-		recallCall("c1", fmt.Sprintf(`{"positions":[%d]}`, scorerTarget)),
-		toolResult("c1", "#42\nthe token is cobalt-lark-17"),
-	}
-	m := scoreRun(direct, scorerTask(), scorerTarget, "arm", true)
-	if !m.CueDirectRead {
-		t.Error("a first-action read with the cue visible was not credited")
-	}
-
-	// Same read, but no cue was on screen to give the address.
-	if scoreRun(direct, scorerTask(), scorerTarget, "arm", false).CueDirectRead {
-		t.Error("a read was credited to a cue that was not visible")
-	}
-
-	// Same read, but a search found the address first.
-	afterSearch := append([]provider.Message{
-		recallCall("c0", `{"query":"retry boundary"}`),
-		toolResult("c0", hitBlock(scorerTarget)),
-	}, direct...)
-	if scoreRun(afterSearch, scorerTask(), scorerTarget, "arm", true).CueDirectRead {
-		t.Error("a read after a search was credited to the index")
-	}
-}
@@ -1,469 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"math/rand"
-	"reasonix/internal/state/sessionstore"
-
-	"reasonix/internal/contract/provider"
-)
-
-// Synthetic history on purpose: an answer that exists in the workspace is
-// solvable by reading the workspace, which measures grep. Every answer is a
-// placeholder here and a value only inside one run.
-
-// experiment names which question a task belongs to.
-const (
-	experimentSearch = "search" // does search reach what no index addresses?
-	experimentIndex  = "index"  // what is an ambient address line worth?
-)
-
-// cueTier is the narrowest fold-index scale at which an index task's cue is
-// still addressed. Below it the model has to find the call without a hint.
-const (
-	tierQuarter = "quarter" // visible at default, half, quarter
-	tierHalf    = "half"    // visible at default, half
-	tierDefault = "default" // visible at default only
-)
-
-// role places an index task in the efficiency substrate or the stopping corpus.
-const (
-	roleEfficiency = "efficiency"
-	roleStopping   = "stopping"
-)
-
-// plantKind is how a task's target enters the transcript.
-const (
-	plantAssistant = "assistant" // the model's own account: never index-addressed
-	plantCall      = "call"      // a tool call and its result: always index-addressed
-)
-
-// contextTask is one recall question as a template. Nothing here is an answer.
-type contextTask struct {
-	ID         string
-	Experiment string
-	TargetKind string
-	Prompt     string
-	ProbeQuery string
-	// Vars declare the placeholders. AnswerVars name the ones a correct answer
-	// must contain, CueVar the one an index line carries.
-	Vars       []varSpec
-	AnswerVars []string
-	CueVar     string
-	CueTier    string
-	// Role is roleEfficiency or roleStopping; it decides the batch a task enters.
-	Role string
-	// PlantAfterGen ages the target: mergeFoldIndex trims from its oldest end,
-	// so what follows a cue decides which budget addresses it. Frozen, because
-	// a run that searched for it would let the tested code move its goalposts.
-	PlantAfterGen int
-	// Plant is how the target enters, and the rest is what it says.
-	Plant     string
-	CallTool  string
-	CallArgs  string
-	CallID    string
-	PlantBody string
-}
-
-// fixtureInstance is one task with values. Scoring binds to this, never to the
-// corpus: the corpus has no answers to bind to.
-type fixtureInstance struct {
-	Task          contextTask
-	Vars          fixtureVars
-	Prompt        string
-	ProbeQuery    string
-	AnswerMarkers []string
-	CueMarker     string
-	body          string
-	callArgs      string
-}
-
-// instantiateTask resolves one task against a generator.
-func instantiateTask(t contextTask, rng *rand.Rand) (fixtureInstance, error) {
-	vars := newVars(t.Vars, rng)
-	inst := fixtureInstance{Task: t, Vars: vars}
-	var err error
-	for _, pair := range []struct {
-		dst  *string
-		text string
-	}{
-		{&inst.Prompt, t.Prompt},
-		{&inst.ProbeQuery, t.ProbeQuery},
-		{&inst.body, t.PlantBody},
-		{&inst.callArgs, t.CallArgs},
-	} {
-		if *pair.dst, err = instantiate(pair.text, vars); err != nil {
-			return fixtureInstance{}, fmt.Errorf("%s: %w", t.ID, err)
-		}
-	}
-	for _, name := range t.AnswerVars {
-		value, ok := vars[name]
-		if !ok {
-			return fixtureInstance{}, fmt.Errorf("%s: answer var %q is not declared", t.ID, name)
-		}
-		inst.AnswerMarkers = append(inst.AnswerMarkers, value)
-	}
-	if t.CueVar != "" {
-		inst.CueMarker = vars[t.CueVar]
-	}
-	return inst, nil
-}
-
-// plant appends the instantiated target and returns its canonical position.
-func (f fixtureInstance) plant(s *sessionstore.Session) int {
-	at := len(s.Messages)
-	if f.Task.Plant == plantAssistant {
-		s.Add(provider.Message{Role: provider.RoleAssistant, Content: f.body})
-		return at
-	}
-	s.Add(provider.Message{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-		{ID: f.Task.CallID, Name: f.Task.CallTool, Arguments: f.callArgs},
-	}})
-	s.Add(provider.Message{Role: provider.RoleTool, ToolCallID: f.Task.CallID, Name: f.Task.CallTool, Content: f.body})
-	return at
-}
-
-func code(name, word string) varSpec { return varSpec{Name: name, Kind: varCodename, Word: word} }
-
-func num(name string, lo, hi int) varSpec {
-	return varSpec{Name: name, Kind: varInt, Min: lo, Max: hi}
-}
-
-// efficiencyTasks is the substrate a study measures on. Every task in it
-// qualifies under the corpus bar; the stopping corpus is deliberately excluded,
-// because a task carrying post-sufficient stopping behaviour swamps the effect
-// an efficiency batch exists to measure.
-func efficiencyTasks() []contextTask {
-	var out []contextTask
-	for _, t := range indexTasks() {
-		if t.Role == roleEfficiency {
-			out = append(out, t)
-		}
-	}
-	return out
-}
-
-func searchTasks() []contextTask {
-	return []contextTask{
-		{
-			ID: "r01-retry-boundary", Experiment: experimentSearch, TargetKind: "assistant_reasoning",
-			Prompt:     "What ownership token did we settle on for the retry boundary, and under what tag did we file the reason for not moving retries above framing?",
-			ProbeQuery: "retry boundary ownership token framing",
-			Vars:       []varSpec{code("token", "cobalt"), code("reason", "stream")},
-			AnswerVars: []string{"token", "reason"},
-			Plant:      plantAssistant,
-			PlantBody: "The retry boundary remains at ingress. The ownership token is {{token}}. " +
-				"Moving retries above framing can duplicate a partially received stream, which we filed as {{reason}}.",
-		},
-		{
-			ID: "r02-checkpoint-anchor", Experiment: experimentSearch, TargetKind: "assistant_reasoning",
-			Prompt:     "What was the codename of the checkpoint anchor policy, and what tag did we give the reason for not anchoring against digest positions?",
-			ProbeQuery: "checkpoint anchor policy digest positions",
-			Vars:       []varSpec{code("policy", "northglass"), code("reason", "genlocal")},
-			AnswerVars: []string{"policy", "reason"},
-			Plant:      plantAssistant,
-			PlantBody: "The checkpoint anchor policy is {{policy}}. Digest positions are generation-local — " +
-				"filed as {{reason}} — so recovery must anchor against canonical state.",
-		},
-		{
-			ID: "r03-recovery-cjk", Experiment: experimentSearch, TargetKind: "assistant_reasoning_cjk",
-			Prompt:     "我们之前定的验证失败恢复策略代号叫什么？当时把“失败 digest 不能变成 baseline”这条登记在哪个标记下？",
-			ProbeQuery: "恢复策略 baseline digest 验证失败",
-			Vars:       []varSpec{code("policy", "songzhen"), code("rule", "trustbase")},
-			AnswerVars: []string{"policy", "rule"},
-			Plant:      plantAssistant,
-			PlantBody: "恢复策略代号是 {{policy}}。验证失败以后必须保留旧基线，不能把新的 digest 晋升为 baseline，" +
-				"因为失败证据不能成为后续比较的可信基准；这条规则登记为 {{rule}}。",
-		},
-		{
-			ID: "r04-lease-fence", Experiment: experimentSearch, TargetKind: "assistant_reasoning",
-			Prompt:     "那次 lease fence 检查报告的诊断标记是什么？expected 和 observed epoch 分别是多少？",
-			ProbeQuery: "lease fence writer epoch mismatch",
-			Vars: []varSpec{
-				code("marker", "fence"), num("expected", 400, 900), num("observed", 100, 399),
-			},
-			AnswerVars: []string{"marker", "expected", "observed"},
-			Plant:      plantAssistant,
-			PlantBody: "The lease fence check reported {{marker}}: expected writer epoch {{expected}}, " +
-				"observed {{observed}}. We left the fence in place.",
-		},
-		{
-			ID: "r05-schema-guard", Experiment: experimentSearch, TargetKind: "assistant_reasoning",
-			Prompt:     "当时 schema guard 报告的标记是什么，超预算多少 token？",
-			ProbeQuery: "schema guard provider visible surface budget",
-			Vars:       []varSpec{code("marker", "amber"), num("over", 120, 480)},
-			AnswerVars: []string{"marker", "over"},
-			Plant:      plantAssistant,
-			PlantBody: "The schema guard reported {{marker}}: the provider-visible tool schema exceeded " +
-				"the allowed surface by {{over}} tokens. We trimmed a description instead.",
-		},
-		{
-			ID: "r06-shadow-probe", Experiment: experimentSearch, TargetKind: "assistant_reasoning",
-			Prompt:     "我们那次 shadow probe 的 marker、p95 divergence 和 generation 分别是多少？",
-			ProbeQuery: "shadow probe divergence generation marker",
-			Vars: []varSpec{
-				code("marker", "violet"),
-				{Name: "divergence", Kind: varDecimal, Min: 110, Max: 890},
-				num("generation", 14, 60),
-			},
-			AnswerVars: []string{"marker", "divergence", "generation"},
-			Plant:      plantAssistant,
-			PlantBody: "The shadow probe came back as marker={{marker}} with a p95 divergence of " +
-				"{{divergence}} at generation={{generation}}, which we judged within tolerance.",
-		},
-	}
-}
-
-// efficiencyIndexTasks is what a study measures on: two tasks per cue tier.
-func efficiencyIndexTasks() []contextTask {
-	out := indexQuarterTasks()
-	out = append(out, indexHalfTasks()...)
-	out = append(out, indexDefaultTasks()...)
-	return out
-}
-
-// indexQuarterTasks holds the quarter-tier literals.
-func indexQuarterTasks() []contextTask {
-	return []contextTask{
-		{
-			ID: "i37-dispatchlag", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleEfficiency,
-			Prompt:     "早前会话里执行过 ./scripts/dispatchlag，它打印出的 dispatch lag epoch 是多少？",
-			ProbeQuery: "dispatchlag dispatch lag epoch",
-			Vars:       []varSpec{code("cue", "dispatchlag"), code("epoch", "cobalt")},
-			AnswerVars: []string{"epoch"}, CueVar: "cue", CueTier: tierQuarter,
-			PlantAfterGen: 4,
-			Plant:         plantCall, CallID: "p37", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/dispatchlag {{cue}}"}`,
-			PlantBody: "dispatch lag epoch = {{epoch}}",
-		},
-		{
-			ID: "i118-fencelag", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleEfficiency,
-			Prompt:     "早前会话里运行过 ./scripts/fencelag，它打印出的 fence lag 是多少？",
-			ProbeQuery: "fencelag fence lag",
-			Vars:       []varSpec{code("cue", "fencelag"), code("v", "slate")},
-			AnswerVars: []string{"v"}, CueVar: "cue", CueTier: tierQuarter,
-			PlantAfterGen: 4,
-			Plant:         plantCall, CallID: "x118", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/fencelag {{cue}}"}`,
-			PlantBody: "fence lag = {{v}}",
-		},
-	}
-}
-
-// indexHalfTasks holds the half-tier literals.
-func indexHalfTasks() []contextTask {
-	return []contextTask{
-		{
-			ID: "i70-ingestfloor", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleEfficiency,
-			Prompt:     "早前会话里执行过 ./scripts/ingestfloor，它打印出的 route quota 是多少？",
-			ProbeQuery: "ingestfloor quota",
-			Vars:       []varSpec{code("cue", "ingestfloor"), num("quota", 1200, 9800)},
-			AnswerVars: []string{"quota"}, CueVar: "cue", CueTier: tierHalf,
-			PlantAfterGen: 2,
-			Plant:         plantCall, CallID: "q70", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/ingestfloor {{cue}}"}`,
-			PlantBody: "route quota = {{quota}}",
-		},
-		{
-			ID: "i106-budgetseal", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleEfficiency,
-			Prompt:     "早前会话里执行过 ./scripts/budgetseal，它打印出的 route quota 是多少？",
-			ProbeQuery: "budgetseal route quota",
-			Vars:       []varSpec{code("cue", "budgetseal"), code("quota", "amber")},
-			AnswerVars: []string{"quota"}, CueVar: "cue", CueTier: tierHalf,
-			PlantAfterGen: 2,
-			Plant:         plantCall, CallID: "h106", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/budgetseal {{cue}}"}`,
-			PlantBody: "route quota = {{quota}}",
-		},
-	}
-}
-
-// indexDefaultTasks holds the default-tier literals.
-func indexDefaultTasks() []contextTask {
-	return []contextTask{
-		{
-			ID: "i47-tenantlease", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleEfficiency,
-			Prompt:     "早前会话里执行过 ./scripts/tenantlease，它打印出的 tenant lease（秒）是多少？",
-			ProbeQuery: "tenantlease tenant lease",
-			Vars:       []varSpec{code("cue", "tenantlease"), {Name: "lease", Kind: varDuration, Min: 15, Max: 300}},
-			AnswerVars: []string{"lease"}, CueVar: "cue", CueTier: tierDefault,
-			PlantAfterGen: 0,
-			Plant:         plantCall, CallID: "p47", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/tenantlease {{cue}}"}`,
-			PlantBody: "tenant lease = {{lease}}",
-		},
-		{
-			ID: "i112-lagquota", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleEfficiency,
-			Prompt:     "早前会话里执行过 ./scripts/lagquota，它打印出的 lagquota lease 是多少？",
-			ProbeQuery: "lagquota lagquota lease",
-			Vars:       []varSpec{code("cue", "lagquota"), {Name: "lease", Kind: varDuration, Min: 20, Max: 420}},
-			AnswerVars: []string{"lease"}, CueVar: "cue", CueTier: tierDefault,
-			PlantAfterGen: 0,
-			Plant:         plantCall, CallID: "d112", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/lagquota {{cue}}"}`,
-			PlantBody: "lagquota lease = {{lease}}",
-		},
-	}
-}
-
-// stoppingIndexTasks is the frozen study's six tasks, kept out of efficiency
-// batches and measurable on their own with -task <id>.
-func stoppingIndexTasks() []contextTask {
-	return []contextTask{
-		{
-			ID: "i01-transport-fallback", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
-			Prompt:     "transport fallback probe 当时报告的 selector 是什么？",
-			ProbeQuery: "transport fallback probe selector",
-			Vars:       []varSpec{code("cue", "tfprobe"), code("selector", "fern")},
-			AnswerVars: []string{"selector"}, CueVar: "cue", CueTier: tierQuarter,
-			PlantAfterGen: 4,
-			Plant:         plantCall, CallID: "tf1", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/probe {{cue}}"}`,
-			PlantBody: "fallback selector = {{selector}}",
-		},
-		{
-			ID: "i02-scheduler-handoff", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
-			Prompt:     "scheduler handoff 当时记录的 grace 和 marker 是什么？",
-			ProbeQuery: "scheduler handoff grace marker",
-			Vars: []varSpec{
-				code("cue", "schedhandoff"), code("marker", "opal"),
-				{Name: "grace", Kind: varDuration, Min: 18, Max: 90},
-			},
-			AnswerVars: []string{"grace", "marker"}, CueVar: "cue", CueTier: tierQuarter,
-			PlantAfterGen: 4,
-			Plant:         plantCall, CallID: "sh1", CallTool: "read_file",
-			CallArgs:  `{"path":"scratch/{{cue}}.txt"}`,
-			PlantBody: "handoff grace = {{grace}}\nmarker = {{marker}}",
-		},
-		{
-			ID: "i03-coalescing", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
-			Prompt:     "job completion 的 coalescing probe 最后记录的 window 和 marker 是多少？",
-			ProbeQuery: "coalescing policy completion window",
-			Vars: []varSpec{
-				code("cue", "coalescing"), code("marker", "cedar"),
-				{Name: "window", Kind: varDuration, Min: 40, Max: 140},
-			},
-			AnswerVars: []string{"window", "marker"}, CueVar: "cue", CueTier: tierHalf,
-			PlantAfterGen: 2,
-			Plant:         plantCall, CallID: "cp1", CallTool: "bash",
-			CallArgs:  `{"command":"grep {{cue}} scratch/job-notes.log"}`,
-			PlantBody: "completion coalesce window = {{window}}\nmarker = {{marker}}",
-		},
-		{
-			ID: "i04-recovery-fence", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
-			Prompt:     "recovery fence 那次检查报告的 marker 和两个 epoch 是什么？",
-			ProbeQuery: "recovery fence epoch check",
-			Vars: []varSpec{
-				code("cue", "recfence"), code("marker", "quartz"),
-				num("expected", 400, 900), num("observed", 100, 399),
-			},
-			AnswerVars: []string{"marker", "expected", "observed"}, CueVar: "cue", CueTier: tierHalf,
-			PlantAfterGen: 2,
-			Plant:         plantCall, CallID: "rf1", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/check {{cue}}"}`,
-			PlantBody: "{{marker}}\nexpected epoch {{expected}}\nobserved epoch {{observed}}",
-		},
-		{
-			ID: "i05-cache-probe", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
-			Prompt:     "cache probe 当时记录的 prefix salt 和 scope 是什么？",
-			ProbeQuery: "cache probe prefix salt scope",
-			Vars:       []varSpec{code("cue", "cacheprobe"), code("salt", "comet"), code("scope", "lineage")},
-			AnswerVars: []string{"salt", "scope"}, CueVar: "cue", CueTier: tierDefault,
-			PlantAfterGen: 0,
-			Plant:         plantCall, CallID: "cb1", CallTool: "read_file",
-			CallArgs:  `{"path":"scratch/{{cue}}.txt"}`,
-			PlantBody: "prefix salt = {{salt}}\nscope = {{scope}}",
-		},
-		{
-			ID: "i06-dispatch-handoff", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
-			Prompt:     "dispatch handoff probe 当时确认的 mode 和 marker 是什么？",
-			ProbeQuery: "dispatch handoff probe mode",
-			Vars:       []varSpec{code("cue", "dispatchhandoff"), code("mode", "trigger"), code("marker", "iris")},
-			AnswerVars: []string{"mode", "marker"}, CueVar: "cue", CueTier: tierDefault,
-			PlantAfterGen: 0,
-			Plant:         plantCall, CallID: "dh1", CallTool: "bash",
-			CallArgs:  `{"command":"./scripts/probe {{cue}}"}`,
-			PlantBody: "handoff mode = {{mode}}\nmarker = {{marker}}",
-		},
-	}
-}
-
-// indexTasks is the whole index corpus, the efficiency substrate first.
-func indexTasks() []contextTask {
-	out := efficiencyIndexTasks()
-	return append(out, stoppingIndexTasks()...)
-}
-
-func allTasks() []contextTask {
-	return append(searchTasks(), indexTasks()...)
-}
-
-func taskByID(id string) (contextTask, bool) {
-	for _, t := range allTasks() {
-		if t.ID == id {
-			return t, true
-		}
-	}
-	return contextTask{}, false
-}
-
-// tierScales is the fixture's contract: which arms must address a cue and
-// which must not. A policy change that moves a cue across a boundary fails
-// preflight rather than turning two arms into one experiment.
-func tierScales(tier string) (visible, hidden []string) {
-	switch tier {
-	case tierQuarter:
-		return []string{"default", "half", "quarter"}, []string{"off"}
-	case tierHalf:
-		return []string{"default", "half"}, []string{"quarter", "off"}
-	case tierDefault:
-		return []string{"default"}, []string{"half", "quarter", "off"}
-	default:
-		return nil, nil
-	}
-}
-
-// boundaryPair is the one comparison an index task is built for: the narrowest
-// scale that still addresses its cue, and the next one down. Holding the
-// question fixed and moving only the affordance is what four arm averages
-// cannot do, since each of those mixes six different questions.
-func boundaryPair(tier string) (cue, noCue string) {
-	switch tier {
-	case tierQuarter:
-		return "quarter", "off"
-	case tierHalf:
-		return "half", "quarter"
-	case tierDefault:
-		return "default", "half"
-	default:
-		return "", ""
-	}
-}
-
-func validateCorpus() error {
-	seen := map[string]bool{}
-	for _, t := range allTasks() {
-		if seen[t.ID] {
-			return fmt.Errorf("duplicate task id %q", t.ID)
-		}
-		seen[t.ID] = true
-		if len(t.AnswerVars) == 0 || t.ProbeQuery == "" || t.Prompt == "" || t.PlantBody == "" {
-			return fmt.Errorf("%s: a task needs a prompt, a probe query, a body and answer vars", t.ID)
-		}
-		if t.Experiment == experimentIndex {
-			if t.CueVar == "" {
-				return fmt.Errorf("%s: an index task needs a cue var", t.ID)
-			}
-			if t.Role != roleEfficiency && t.Role != roleStopping {
-				return fmt.Errorf("%s: an index task needs a role (%s or %s)", t.ID, roleEfficiency, roleStopping)
-			}
-			if v, _ := tierScales(t.CueTier); v == nil {
-				return fmt.Errorf("%s: unknown cue tier %q", t.ID, t.CueTier)
-			}
-		}
-		if _, err := instantiateTask(t, seededRand(t.ID, "validate")); err != nil {
-			return err
-		}
-	}
-	return nil
-}
@@ -1,243 +0,0 @@
-// dose.go — does the cost step where the cue disappears?
-package main
-
-import (
-	"fmt"
-	"math"
-	"sort"
-	"strings"
-)
-
-// Not "a bigger index is better" — CueDirectRead says the addresses are rarely
-// used. The hypothesis is that a visible action cue lowers retrieval cost,
-// which predicts a step at each task's own boundary, not a slope.
-
-// scaleOrder is the dose axis, widest first.
-var scaleOrder = []string{"default", "half", "quarter", "off"}
-
-// boundaryStep is the index into the three transitions where a tier's cue
-// disappears: default->half is 0, half->quarter is 1, quarter->off is 2.
-func boundaryStep(tier string) int {
-	switch tier {
-	case tierDefault:
-		return 0
-	case tierHalf:
-		return 1
-	case tierQuarter:
-		return 2
-	default:
-		return -1
-	}
-}
-
-// doseCurve is one task's readings across the four budgets.
-type doseCurve struct {
-	Task   string
-	Tier   string
-	Metric string
-	Values [4]float64
-	// Missing marks arms that produced nothing. Printing those as zero draws a
-	// curve falling to the floor where in fact it was never measured.
-	Missing [4]bool
-	Present bool
-	// PreMean averages the arms where the cue is visible, PostMean the rest.
-	PreMean, PostMean float64
-	BoundaryDelta     float64
-	// BoundaryAligned: the largest transition is the one where this task's cue
-	// disappears. A change elsewhere is variance shaped like an effect.
-	BoundaryAligned bool
-}
-
-func buildCurve(task, tier, metric string, byScale map[string]float64) doseCurve {
-	c := doseCurve{Task: task, Tier: tier, Metric: metric, Present: true}
-	for i, scale := range scaleOrder {
-		v, ok := byScale[scale]
-		if !ok {
-			c.Present, c.Missing[i] = false, true
-		}
-		c.Values[i] = v
-	}
-	step := boundaryStep(tier)
-	if !c.Present || step < 0 {
-		return c
-	}
-	visible := step + 1 // arms before the boundary still show the cue
-	c.PreMean = mean(c.Values[:visible])
-	c.PostMean = mean(c.Values[visible:])
-	c.BoundaryDelta = c.PostMean - c.PreMean
-
-	transitions := [3]float64{
-		math.Abs(c.Values[1] - c.Values[0]),
-		math.Abs(c.Values[2] - c.Values[1]),
-		math.Abs(c.Values[3] - c.Values[2]),
-	}
-	largest, at := 0.0, -1
-	for i, t := range transitions {
-		if t > largest {
-			largest, at = t, i
-		}
-	}
-	// A flat curve has no largest transition to align with anything.
-	c.BoundaryAligned = largest > 0 && at == step
-	return c
-}
-
-func mean(xs []float64) float64 {
-	if len(xs) == 0 {
-		return 0
-	}
-	sum := 0.0
-	for _, x := range xs {
-		sum += x
-	}
-	return sum / float64(len(xs))
-}
-
-func median(xs []float64) float64 {
-	if len(xs) == 0 {
-		return 0
-	}
-	s := append([]float64(nil), xs...)
-	sort.Float64s(s)
-	mid := len(s) / 2
-	if len(s)%2 == 1 {
-		return s[mid]
-	}
-	return (s[mid-1] + s[mid]) / 2
-}
-
-// doseMetrics are the readings a curve is drawn for.
-var doseMetrics = []struct {
-	name string
-	of   func(contextMetrics) float64
-}{
-	{"searches", func(m contextMetrics) float64 { return float64(m.SearchCalls) }},
-	{"recall-tokens", func(m contextMetrics) float64 { return float64(m.RecallReturnedTokens) }},
-	{"tool-rounds", func(m contextMetrics) float64 { return float64(m.RetrievalRounds + m.EscapeCalls) }},
-	{"escapes", func(m contextMetrics) float64 { return float64(m.EscapeCalls) }},
-}
-
-// reportDoseResponse draws each task's curve and reports where the steps fell.
-func reportDoseResponse(all []contextMetrics) string {
-	valid := map[string]map[string]contextMetrics{}
-	for _, m := range all {
-		if m.Failed || m.Contaminated || m.FailureStage == "RunError" {
-			continue
-		}
-		scale := strings.TrimPrefix(m.Arm, "index-")
-		if valid[m.Task] == nil {
-			valid[m.Task] = map[string]contextMetrics{}
-		}
-		valid[m.Task][scale] = m
-	}
-
-	var b strings.Builder
-	b.WriteString("\n## Dose-response, per task\n")
-	b.WriteString("  The arrow marks where this task's cue stops being addressed.\n")
-	aligned := map[string]int{}
-	counted := map[string]int{}
-	deltas := map[string][]float64{}
-
-	for _, t := range indexTasks() {
-		arms := valid[t.ID]
-		if len(arms) == 0 {
-			continue
-		}
-		step := boundaryStep(t.CueTier)
-		fmt.Fprintf(&b, "\n  %s (cue survives to %s)\n", t.ID, t.CueTier)
-		fmt.Fprintf(&b, "    %-14s%s\n", "", scaleHeader(step))
-		for _, metric := range doseMetrics {
-			byScale := map[string]float64{}
-			for scale, m := range arms {
-				byScale[scale] = metric.of(m)
-			}
-			c := buildCurve(t.ID, t.CueTier, metric.name, byScale)
-			mark := " "
-			if c.BoundaryAligned {
-				mark = "*"
-				aligned[metric.name]++
-			}
-			if c.Present {
-				counted[metric.name]++
-				deltas[metric.name] = append(deltas[metric.name], c.BoundaryDelta)
-			}
-			delta := fmt.Sprintf("Δ%+8.1f", c.BoundaryDelta)
-			if !c.Present {
-				delta = "  no curve"
-			}
-			fmt.Fprintf(&b, "    %-14s%s   %s %s\n", metric.name, curveCells(c), delta, mark)
-		}
-	}
-
-	b.WriteString("\n## Where the step fell (* = largest transition is at the cue boundary)\n")
-	for _, metric := range doseMetrics {
-		d := deltas[metric.name]
-		sensitivity := ""
-		if len(d) > 1 {
-			sensitivity = fmt.Sprintf("  without the largest: %+.1f", meanWithoutLargest(d))
-		}
-		fmt.Fprintf(&b, "  %-14s boundary-aligned %d/%d   mean Δ%+.1f   median Δ%+.1f%s\n",
-			metric.name, aligned[metric.name], counted[metric.name], mean(d), median(d), sensitivity)
-	}
-	return b.String()
-}
-
-// meanWithoutLargest is the robustness check: the largest effect is reported
-// and then set aside, so a conclusion never rests on one task alone.
-func meanWithoutLargest(d []float64) float64 {
-	if len(d) < 2 {
-		return mean(d)
-	}
-	worst, at := 0.0, 0
-	for i, v := range d {
-		if math.Abs(v) > worst {
-			worst, at = math.Abs(v), i
-		}
-	}
-	rest := append(append([]float64(nil), d[:at]...), d[at+1:]...)
-	return mean(rest)
-}
-
-func scaleHeader(step int) string {
-	var b strings.Builder
-	for i, s := range scaleOrder {
-		if i == step+1 {
-			s = "↓" + s
-		}
-		fmt.Fprintf(&b, "%8s", s)
-	}
-	return b.String()
-}
-
-// curveCells prints a reading per arm, and a dash where an arm produced none.
-func curveCells(c doseCurve) string {
-	var b strings.Builder
-	for i, v := range c.Values {
-		if c.Missing[i] {
-			fmt.Fprintf(&b, "%8s", "-")
-			continue
-		}
-		fmt.Fprintf(&b, "%8.0f", v)
-	}
-	return b.String()
-}
-
-func percentile(xs []float64, q float64) float64 {
-	if len(xs) == 0 {
-		return 0
-	}
-	s := append([]float64(nil), xs...)
-	sort.Float64s(s)
-	i := int(float64(len(s)-1) * q)
-	return s[i]
-}
-
-func maxOf(xs []float64) float64 {
-	out := 0.0
-	for _, x := range xs {
-		if x > out {
-			out = x
-		}
-	}
-	return out
-}
@@ -1,306 +0,0 @@
-package main
-
-import (
-	"encoding/json"
-	"os"
-	"path/filepath"
-	"reflect"
-	"strconv"
-	"strings"
-	"testing"
-)
-
-// blockTask makes one task fail before its first model turn: its own directory
-// is a regular file, so buildFixture cannot create the session under it. It is
-// the same early exit as a failed fixture build, session load, or
-// answer-on-disk guard, and it needs no provider.
-func blockTask(t *testing.T, root, id string) {
-	t.Helper()
-	if err := os.WriteFile(filepath.Join(root, id), []byte("not a directory\n"), 0o644); err != nil {
-		t.Fatal(err)
-	}
-}
-
-// mixedTaskPair is one task whose runs all fail and one whose runs all succeed.
-// Both are drawn from the corpus, so the batch is the real pipeline.
-func mixedTaskPair() (blocked, ok contextTask) {
-	blocked, ok = indexTasks()[0], indexTasks()[1]
-	for _, t := range indexTasks() {
-		if t.Role == roleEfficiency && len(armsFor(t)) == 2 {
-			ok = t
-			break
-		}
-	}
-	return blocked, ok
-}
-
-// runBlockedExperiment runs one dry index batch over the given tasks, with the
-// named task failing before any model turn, and returns the batch directory, the
-// runs, the manifest, and everything the run printed.
-func runBlockedExperiment(t *testing.T, tasks []contextTask, blocked ...string) (string, []contextMetrics, batchManifest, string) {
-	t.Helper()
-	root := t.TempDir()
-	for _, id := range blocked {
-		blockTask(t, root, id)
-	}
-	out := captureStdout(t, func() {
-		if code := runExperiment("run-index", experimentIndex, root, true, tasks); code != 0 {
-			t.Errorf("runExperiment exit = %d, want 0 (the batch finished)", code)
-		}
-	})
-	runs, _ := readResultLines(t, root)
-	manifest, _ := readManifest(t, root)
-	return root, runs, manifest, out
-}
-
-// captureStdout reads everything the function prints. os.Pipe is unbuffered, so
-// the reader must run beside the call.
-func captureStdout(t *testing.T, f func()) string {
-	t.Helper()
-	r, w, err := os.Pipe()
-	if err != nil {
-		t.Fatal(err)
-	}
-	old := os.Stdout
-	os.Stdout = w
-	done := make(chan string, 1)
-	go func() {
-		var b strings.Builder
-		buf := make([]byte, 4096)
-		for {
-			n, err := r.Read(buf)
-			b.Write(buf[:n])
-			if err != nil {
-				break
-			}
-		}
-		done <- b.String()
-	}()
-	f()
-	os.Stdout = old
-	w.Close()
-	text := <-done
-	r.Close()
-	return text
-}
-
-// funnelScored reads the "runs=N" column of the named arm's funnel line out of
-// the printed tables, so the assertion is on the number a reader sees.
-func funnelScored(t *testing.T, out, arm string) int {
-	t.Helper()
-	for line := range strings.SplitSeq(out, "\n") {
-		fields := strings.Fields(line)
-		if len(fields) < 2 || fields[0] != arm || !strings.HasPrefix(fields[1], "runs=") {
-			continue
-		}
-		n, err := strconv.Atoi(strings.TrimPrefix(fields[1], "runs="))
-		if err != nil {
-			t.Fatalf("funnel line for %s: %v\n%s", arm, err, line)
-		}
-		return n
-	}
-	t.Fatalf("no funnel line for arm %s:\n%s", arm, out)
-	return 0
-}
-
-// The batch report states both counts on one line, so a reader of the scrollback
-// sees the scored set and the failures without counting funnel rows.
-func reportCounts(t *testing.T, out string) (scored, failed int) {
-	t.Helper()
-	for line := range strings.SplitSeq(out, "\n") {
-		if !strings.Contains(line, " scored=") || !strings.Contains(line, " failed=") {
-			continue
-		}
-		for f := range strings.FieldsSeq(line) {
-			switch {
-			case strings.HasPrefix(f, "scored="):
-				scored, _ = strconv.Atoi(strings.TrimPrefix(f, "scored="))
-			case strings.HasPrefix(f, "failed="):
-				failed, _ = strconv.Atoi(strings.TrimPrefix(f, "failed="))
-			}
-		}
-		return scored, failed
-	}
-	t.Fatalf("no line reports scored= and failed=:\n%s", out)
-	return 0, 0
-}
-
-// A batch that mixes a task that never ran with one that did must read as its
-// successful runs only: the failed task leaves Scored, the funnel, and the
-// completed count, and is reported as a failure with its identity instead.
-// Every assertion here is a number, not the absence of an empty row.
-func TestMixedBatchCountsOnlyTheRunThatProducedEvidence(t *testing.T) {
-	blocked, ok := mixedTaskPair()
-	_, runs, manifest, out := runBlockedExperiment(t, []contextTask{blocked, ok}, blocked.ID)
-
-	wantScored, wantFailed := len(armsFor(ok)), len(armsFor(blocked))
-	if manifest.CompletedRuns != wantScored || manifest.FailedRuns != wantFailed {
-		t.Fatalf("manifest completed=%d failed=%d, want %d and %d (planned=%d)",
-			manifest.CompletedRuns, manifest.FailedRuns, wantScored, wantFailed, manifest.PlannedRuns)
-	}
-	if len(runs) != manifest.CompletedRuns+manifest.FailedRuns {
-		t.Fatalf("results.jsonl has %d rows but completed=%d failed=%d",
-			len(runs), manifest.CompletedRuns, manifest.FailedRuns)
-	}
-	for _, id := range manifest.TaskIDs {
-		if id == blocked.ID {
-			t.Fatalf("task_ids lists %s, which produced no evidence: %v", blocked.ID, manifest.TaskIDs)
-		}
-	}
-	scored, failures := 0, 0
-	for _, m := range runs {
-		if m.Failed {
-			failures++
-			if m.Task != blocked.ID || m.FailureStage == "" || m.Reason == "" {
-				t.Fatalf("failed row lost its identity or stage: %+v", m)
-			}
-			continue
-		}
-		scored++
-		if m.Task != ok.ID {
-			t.Fatalf("a run scored from the wrong task: %+v", m)
-		}
-	}
-	if scored != wantScored || failures != wantFailed {
-		t.Fatalf("rows scored=%d failed=%d, want %d and %d", scored, failures, wantScored, wantFailed)
-	}
-	// Each arm's funnel holds exactly the successful task's runs, and the two
-	// scored counts agree with each other and with the manifest.
-	total := 0
-	for _, arm := range armsFor(ok) {
-		total += funnelScored(t, out, arm.name)
-	}
-	if total != manifest.CompletedRuns {
-		t.Fatalf("funnels scored %d runs, manifest completed_runs = %d:\n%s", total, manifest.CompletedRuns, out)
-	}
-	if gotScored, gotFailed := reportCounts(t, out); gotScored != manifest.CompletedRuns || gotFailed != manifest.FailedRuns {
-		t.Fatalf("report says scored=%d failed=%d, manifest says %d and %d",
-			gotScored, gotFailed, manifest.CompletedRuns, manifest.FailedRuns)
-	}
-	if strings.Contains(out, "## Where it broke") && strings.Contains(out, "=-\n") {
-		t.Fatalf("the stage table kept an empty stage:\n%s", out)
-	}
-}
-
-// A batch where every run fails must not read as an empty success: zero scored
-// runs, a nonzero failed count, and both stated.
-func TestAllFailedBatchReportsZeroScoredAndNonzeroFailed(t *testing.T) {
-	blocked, _ := mixedTaskPair()
-	_, runs, manifest, out := runBlockedExperiment(t, []contextTask{blocked}, blocked.ID)
-
-	if manifest.CompletedRuns != 0 {
-		t.Fatalf("completed_runs = %d, want 0 (nothing produced evidence)", manifest.CompletedRuns)
-	}
-	if manifest.FailedRuns != manifest.PlannedRuns || manifest.FailedRuns == 0 {
-		t.Fatalf("failed_runs = %d, planned_runs = %d: every planned run failed",
-			manifest.FailedRuns, manifest.PlannedRuns)
-	}
-	if len(manifest.TaskIDs) != 0 {
-		t.Fatalf("task_ids = %v, want none", manifest.TaskIDs)
-	}
-	for i, m := range runs {
-		if !m.Failed || m.FailureStage == "" || m.Task != blocked.ID {
-			t.Fatalf("results.jsonl line %d is not a marked failure: %+v", i+1, m)
-		}
-	}
-	total := 0
-	for _, arm := range armsFor(blocked) {
-		total += funnelScored(t, out, arm.name)
-	}
-	if total != 0 {
-		t.Fatalf("funnels scored %d runs although nothing ran:\n%s", total, out)
-	}
-	if gotScored, gotFailed := reportCounts(t, out); gotScored != 0 || gotFailed != manifest.FailedRuns {
-		t.Fatalf("report says scored=%d failed=%d, want 0 and %d", gotScored, gotFailed, manifest.FailedRuns)
-	}
-}
-
-// The control: with nothing blocked, a batch is what it always was. Every run
-// is completed with a stage, two identical batches write byte-identical results
-// and the same manifest header, and the new failure field stays out of it.
-func TestAllSuccessfulBatchKeepsItsArtifacts(t *testing.T) {
-	tasks := indexTasks()
-	first, runs, manifest, out := runBlockedExperiment(t, tasks)
-	second, again, repeat, _ := runBlockedExperiment(t, tasks)
-	if manifest.CompletedRuns != len(runs) || manifest.FailedRuns != 0 {
-		t.Fatalf("completed=%d failed=%d rows=%d, want every row completed",
-			manifest.CompletedRuns, manifest.FailedRuns, len(runs))
-	}
-	if manifest.CompletedRuns != manifest.PlannedRuns {
-		t.Fatalf("completed=%d planned=%d, want all planned", manifest.CompletedRuns, manifest.PlannedRuns)
-	}
-	for i, m := range runs {
-		if m.Failed || m.Task == "" || m.Arm == "" || m.FailureStage == "" {
-			t.Fatalf("results.jsonl line %d is not a completed run: %+v", i+1, m)
-		}
-	}
-	if gotScored, gotFailed := reportCounts(t, out); gotScored != manifest.CompletedRuns || gotFailed != 0 {
-		t.Fatalf("report says scored=%d failed=%d, want %d and 0", gotScored, gotFailed, manifest.CompletedRuns)
-	}
-	if !reflect.DeepEqual(runs, again) {
-		t.Fatal("two identical successful batches scored differently")
-	}
-	// Byte-for-byte on the runs. Trajectories carry per-message wall-clock
-	// stamps, so they are compared field by field instead.
-	if got, want := readFile(t, second, "results.jsonl"), readFile(t, first, "results.jsonl"); got != want {
-		t.Fatalf("an unchanged batch wrote different results.jsonl:\n%s\n---\n%s", want, got)
-	}
-	gotTraj, wantTraj := withoutStamps(t, readFile(t, second, "trajectories.jsonl")),
-		withoutStamps(t, readFile(t, first, "trajectories.jsonl"))
-	if gotTraj != wantTraj {
-		t.Fatalf("an unchanged batch wrote different trajectories:\n%s\n---\n%s", wantTraj, gotTraj)
-	}
-	repeat.StartedAt, repeat.FinishedAt = manifest.StartedAt, manifest.FinishedAt
-	repeat.TrajectoriesSHA256 = manifest.TrajectoriesSHA256
-	if !reflect.DeepEqual(repeat, manifest) {
-		t.Fatalf("an unchanged batch wrote a different manifest:\n%+v\n---\n%+v", manifest, repeat)
-	}
-	if repeat.FailedRuns != 0 {
-		t.Fatalf("a batch with no failure reports failed_runs=%d", repeat.FailedRuns)
-	}
-}
-
-// withoutStamps drops the createdAt stamp every message carries, which is the
-// one field a dry batch cannot reproduce between two runs.
-func withoutStamps(t *testing.T, text string) string {
-	t.Helper()
-	var b strings.Builder
-	for line := range strings.SplitSeq(strings.TrimSpace(text), "\n") {
-		var row map[string]any
-		if err := json.Unmarshal([]byte(line), &row); err != nil {
-			t.Fatal(err)
-		}
-		stripStamp(row)
-		out, err := json.Marshal(row)
-		if err != nil {
-			t.Fatal(err)
-		}
-		b.Write(out)
-		b.WriteByte('\n')
-	}
-	return b.String()
-}
-
-func stripStamp(v any) {
-	switch typed := v.(type) {
-	case map[string]any:
-		delete(typed, "createdAt")
-		delete(typed, "workDurationMs")
-		for _, child := range typed {
-			stripStamp(child)
-		}
-	case []any:
-		for _, child := range typed {
-			stripStamp(child)
-		}
-	}
-}
-
-func readFile(t *testing.T, root, name string) string {
-	t.Helper()
-	data, err := os.ReadFile(filepath.Join(root, name))
-	if err != nil {
-		t.Fatal(err)
-	}
-	return string(data)
-}
@@ -1,205 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"io/fs"
-	"os"
-	"path/filepath"
-	"reasonix/internal/state/sessionstore"
-	"strings"
-
-	"reasonix/internal/contract/ablation"
-	"reasonix/internal/contract/event"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/contract/tool"
-	"reasonix/internal/runtime/agent"
-)
-
-// A fixture is a real session that has really been compacted: no bench-only
-// state shape, so the loader and the projection restore have to work on it.
-
-const (
-	fixtureWindow      = 60000
-	fixtureGenerations = 5
-	// fillerBytes per unit is what makes a fold necessary rather than optional.
-	fillerBytes = 6000
-)
-
-// digestProvider answers a summarizer without naming anything the corpus
-// plants. A digest that happened to quote the answer would leave it visible
-// and silently turn the task into a no-op.
-type digestProvider struct{ calls int }
-
-func (p *digestProvider) Name() string { return "fixture-digest" }
-
-func (p *digestProvider) Stream(_ context.Context, _ provider.Request) (<-chan provider.Chunk, error) {
-	p.calls++
-	ch := make(chan provider.Chunk, 2)
-	ch <- provider.Chunk{Type: provider.ChunkText, Text: "## Earlier work\nEarlier work continued across several files; nothing outstanding was recorded."}
-	ch <- provider.Chunk{Type: provider.ChunkDone}
-	close(ch)
-	return ch, nil
-}
-
-// fillerCall is one ordinary read whose only job is to take up room — in the
-// window, and in the fold index ahead of a cue.
-func fillerCall(sess *sessionstore.Session, n int) {
-	id := fmt.Sprintf("f%04d", n)
-	sess.Add(provider.Message{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{
-		{ID: id, Name: "read_file", Arguments: fmt.Sprintf(`{"path":"internal/mod%03d/handler%03d.go"}`, n%37, n)},
-	}})
-	sess.Add(provider.Message{Role: provider.RoleTool, ToolCallID: id, Name: "read_file", Content: fillerBody(n)})
-}
-
-func fillerBody(n int) string {
-	var b strings.Builder
-	fmt.Fprintf(&b, "package mod%03d\n\n", n%37)
-	for b.Len() < fillerBytes {
-		fmt.Fprintf(&b, "func handler%03dStep%02d() error { return nil }\n", n, b.Len()%64)
-	}
-	return b.String()
-}
-
-// builtFixture is one task's history after real folds, plus where its target
-// ended up.
-type builtFixture struct {
-	Session  *sessionstore.Session
-	State    sessionstore.CompactionState
-	Target   int
-	Path     string
-	Instance fixtureInstance
-}
-
-// buildFixture plays a task's history through the real agent and folds it at
-// the given arm, writing the session and its context state where the agent
-// itself would.
-func buildFixture(f fixtureInstance, arm ablation.Set, sessionPath string) (builtFixture, error) {
-	t := f.Task
-	sess := sessionstore.NewSession("You are a coding agent.")
-	sess.Add(provider.Message{Role: provider.RoleUser, Content: "Work through the transport and scheduler backlog."})
-
-	n := 0
-	at := -1
-
-	a := agent.New(&digestProvider{}, tool.NewRegistry(), sess, agent.Options{
-		ContextWindow: fixtureWindow, CompactRatio: 0.5, RecentKeep: 2,
-		SessionPath: sessionPath, KeepPolicy: agent.KeepErrors, Ablation: arm,
-	}, event.Discard)
-
-	for gen := range fixtureGenerations {
-		if gen == t.PlantAfterGen {
-			at = f.plant(sess)
-		}
-		for range 8 {
-			fillerCall(sess, n)
-			n++
-		}
-		if _, err := a.CompactNow(context.Background(), agent.CompactRequest{}); err != nil {
-			return builtFixture{}, fmt.Errorf("%s: generation %d fold: %w", t.ID, gen, err)
-		}
-	}
-	if at < 0 {
-		return builtFixture{}, fmt.Errorf("%s: PlantAfterGen %d is past the last generation", t.ID, t.PlantAfterGen)
-	}
-	// projectionCoversTail cannot revalidate a projection covering the whole
-	// transcript: it needs a live tail, which a real session always has because
-	// work continues after a fold. Neutral text — the tail is provider-visible.
-	sess.Add(provider.Message{Role: provider.RoleAssistant, Content: "Backlog swept; ready for the next question."})
-	if err := sess.Save(sessionPath); err != nil {
-		return builtFixture{}, fmt.Errorf("%s: save session: %w", t.ID, err)
-	}
-	state, ok, err := sessionstore.LoadCompactionState(sessionPath)
-	if err != nil {
-		return builtFixture{}, fmt.Errorf("%s: load context state: %w", t.ID, err)
-	}
-	if !ok || len(state.Projection.Messages) == 0 {
-		return builtFixture{}, fmt.Errorf("%s: no projection was installed", t.ID)
-	}
-	if state.Projection.CoveredCount <= at {
-		return builtFixture{}, fmt.Errorf("%s: target #%d is not inside the folded region (covered %d)",
-			t.ID, at, state.Projection.CoveredCount)
-	}
-	return builtFixture{Session: sess, State: state, Target: at, Path: sessionPath, Instance: f}, nil
-}
-
-// visibleContext is the model-visible view: the stored projection spliced with
-// everything canonical that came after it.
-func visibleContext(f builtFixture) []provider.Message {
-	canonical := f.Session.Snapshot()
-	out := append([]provider.Message(nil), f.State.Projection.Messages...)
-	if n := f.State.Projection.CoveredCount; n >= 0 && n < len(canonical) {
-		out = append(out, canonical[n:]...)
-	}
-	return out
-}
-
-// visibleText is everything the model would read, as one string.
-func visibleText(f builtFixture) string {
-	var b strings.Builder
-	for _, m := range visibleContext(f) {
-		b.WriteString(m.Content)
-		b.WriteString("\n")
-		for _, tc := range m.ToolCalls {
-			b.WriteString(tc.Name)
-			b.WriteString(" ")
-			b.WriteString(string(tc.Arguments))
-			b.WriteString("\n")
-		}
-	}
-	return b.String()
-}
-
-// sealFixture removes the fixture from disk once the agent holds it in memory.
-// The canonical transcript is the answer, and the sandbox mounts the host
-// read-only by design — so a model that finds the path reads the answer, which
-// one did. The agent has already loaded the session and its sidecar by the time
-// this runs; a failed autosave afterwards costs the benchmark nothing.
-func sealFixture(sessionPath string) {
-	// Every file, not two named ones: a session writes an event log and a
-	// context sidecar beside the transcript, and session.events.jsonl still
-	// carried the answer after the other two were sealed by name.
-	dir := filepath.Dir(sessionPath)
-	entries, err := os.ReadDir(dir)
-	if err != nil {
-		return
-	}
-	for _, e := range entries {
-		if !e.IsDir() {
-			_ = os.Remove(filepath.Join(dir, e.Name()))
-		}
-	}
-	// Deny the directory too, so an autosave cannot recreate what was removed.
-	_ = os.Chmod(dir, 0o500)
-}
-
-// unsealFixture restores the directory so the harness can clean up after
-// itself.
-func unsealFixture(sessionPath string) {
-	_ = os.Chmod(filepath.Dir(sessionPath), 0o700)
-}
-
-// answerOnDisk reports any readable file under dir holding a current-run
-// answer. Two named files were sealed because two leaks were found in them;
-// this scans the directory, so a path that moves — an autosave, a recovery
-// tombstone, a future sidecar — fails here instead of on a paid run.
-func answerOnDisk(dir string, markers []string) (string, string) {
-	var atPath, found string
-	_ = filepath.WalkDir(dir, func(path string, d fs.DirEntry, err error) error {
-		if err != nil || d.IsDir() || atPath != "" {
-			return nil //nolint:nilerr // an unreadable entry is not a leak
-		}
-		body, readErr := os.ReadFile(path)
-		if readErr != nil {
-			return nil
-		}
-		for _, marker := range markers {
-			if len(marker) >= 6 && strings.Contains(string(body), marker) {
-				atPath, found = path, marker
-				return filepath.SkipAll
-			}
-		}
-		return nil
-	})
-	return atPath, found
-}
@@ -1,139 +0,0 @@
-// instance.go — a task template becomes one run's concrete history.
-package main
-
-import (
-	"crypto/rand"
-	"fmt"
-	"hash/fnv"
-	"io"
-	mathrand "math/rand"
-	"strings"
-)
-
-// The invariant this file exists for: no literal sufficient to satisfy
-// AnswerRecovered for a concrete run may exist in this repository before that
-// run begins. Grepping the corpus reveals the question, never the answer.
-
-// varKind is how one placeholder's value is generated. Values stay shaped like
-// real engineering records — a UUID would defeat the leak but turn every task
-// into "find the random string".
-type varKind string
-
-const (
-	varCodename varKind = "codename" // cobalt-k7m4q9
-	varInt      varKind = "int"      // 731
-	varDecimal  varKind = "decimal"  // 0.0437
-	varDuration varKind = "duration" // 43ms
-)
-
-// varSpec declares one placeholder. Min/Max bound the numeric kinds.
-type varSpec struct {
-	Name string
-	Kind varKind
-	Word string // codename stem, e.g. "cobalt"
-	Min  int
-	Max  int
-}
-
-// codenameEntropy is how much of a codename is unguessable. Five base32 digits
-// is about 25 bits: far past what a model can reach by knowing the stem.
-const codenameEntropy = 5
-
-const codenameAlphabet = "abcdefghijkmnpqrstuvwxyz23456789"
-
-// fixtureVars is one run's instantiated values.
-type fixtureVars map[string]string
-
-// newVars generates values for every spec. A nil rng means deterministic.
-func newVars(specs []varSpec, rng *mathrand.Rand) fixtureVars {
-	out := fixtureVars{}
-	for _, spec := range specs {
-		switch spec.Kind {
-		case varCodename:
-			out[spec.Name] = spec.Word + "-" + randomToken(rng, codenameEntropy)
-		case varInt:
-			out[spec.Name] = fmt.Sprint(randomBetween(rng, spec.Min, spec.Max))
-		case varDecimal:
-			out[spec.Name] = fmt.Sprintf("0.%04d", randomBetween(rng, spec.Min, spec.Max))
-		case varDuration:
-			out[spec.Name] = fmt.Sprintf("%dms", randomBetween(rng, spec.Min, spec.Max))
-		}
-	}
-	return out
-}
-
-func randomToken(rng *mathrand.Rand, n int) string {
-	var b strings.Builder
-	for range n {
-		b.WriteByte(codenameAlphabet[randomBetween(rng, 0, len(codenameAlphabet)-1)])
-	}
-	return b.String()
-}
-
-func randomBetween(rng *mathrand.Rand, lo, hi int) int {
-	if hi <= lo {
-		return lo
-	}
-	return lo + rng.Intn(hi-lo+1)
-}
-
-// seededRand is the deterministic generator preflight and tests use, so a
-// fixture's tier calibration and its golden state reproduce exactly.
-func seededRand(parts ...string) *mathrand.Rand {
-	h := fnv.New64a()
-	for _, p := range parts {
-		h.Write([]byte(p))
-		h.Write([]byte{0})
-	}
-	return mathrand.New(mathrand.NewSource(int64(h.Sum64())))
-}
-
-// liveRand is the generator a paid run uses: every run gets values that existed
-// nowhere before it started.
-func liveRand() *mathrand.Rand {
-	rng, err := liveRandFrom(rand.Reader)
-	if err != nil {
-		panic("contextbench: no entropy for a live run: " + err.Error())
-	}
-	return rng
-}
-
-// liveRandFrom seeds a generator from an entropy source, so what a run's values
-// are unpredictable from is the source and nothing else.
-func liveRandFrom(entropy io.Reader) (*mathrand.Rand, error) {
-	var seed [8]byte
-	if _, err := io.ReadFull(entropy, seed[:]); err != nil {
-		return nil, err
-	}
-	var n int64
-	for _, b := range seed {
-		n = n<<8 | int64(b)
-	}
-	return mathrand.New(mathrand.NewSource(n)), nil
-}
-
-// instantiate substitutes {{name}} and refuses anything left over. The grammar
-// is one identifier in double braces: no conditionals, no loops, no defaults.
-// This is fixture data, not a template language.
-func instantiate(text string, vars fixtureVars) (string, error) {
-	out := text
-	for name, value := range vars {
-		out = strings.ReplaceAll(out, "{{"+name+"}}", value)
-	}
-	if strings.Contains(out, "{{") || strings.Contains(out, "}}") {
-		return "", fmt.Errorf("unresolved placeholder in %q", firstPlaceholder(out))
-	}
-	return out, nil
-}
-
-func firstPlaceholder(text string) string {
-	i := strings.Index(text, "{{")
-	if i < 0 {
-		return text
-	}
-	j := strings.Index(text[i:], "}}")
-	if j < 0 {
-		return text[i:min(i+40, len(text))]
-	}
-	return text[i : i+j+2]
-}
@@ -1,144 +0,0 @@
-// ContextRetrievalBench builds the fixtures the long-horizon retrieval
-// experiments run against, and proves they are worth paying a provider for.
-//
-//	-mode=calibrate  find each index task's lead-call count for its cue tier
-//	-mode=preflight  build every task under every arm and check the contract
-//
-// Nothing here calls a provider. The experiments themselves do; this is what
-// says they would measure what they claim to.
-package main
-
-import (
-	"flag"
-	"fmt"
-	"os"
-	"sort"
-	"strings"
-)
-
-func main() {
-	mode := flag.String("mode", "preflight", "calibrate | preflight | snippet | adversarial | run-search | run-boundary | run-index")
-	work := flag.String("work", "", "directory for built fixtures (default: a temp dir)")
-	only := flag.String("task", "", "limit to one task id")
-	dry := flag.Bool("dry", false, "run modes: drive the pipeline with a scripted provider instead of paying for one")
-	flag.Parse()
-
-	if err := validateCorpus(); err != nil {
-		fmt.Fprintln(os.Stderr, "corpus:", err)
-		os.Exit(2)
-	}
-	root := *work
-	if root == "" {
-		dir, err := os.MkdirTemp("", "context-retrieval-*")
-		if err != nil {
-			fmt.Fprintln(os.Stderr, err)
-			os.Exit(2)
-		}
-		defer os.RemoveAll(dir)
-		root = dir
-	}
-
-	tasks := allTasks()
-	if *only != "" {
-		t, ok := taskByID(*only)
-		if !ok {
-			fmt.Fprintf(os.Stderr, "unknown task %q\n", *only)
-			os.Exit(2)
-		}
-		tasks = []contextTask{t}
-	}
-
-	switch *mode {
-	case "calibrate":
-		os.Exit(runCalibrate(tasks, root))
-	case "preflight":
-		os.Exit(runPreflight(tasks, root))
-	case "run-search":
-		os.Exit(runExperiment(*mode, experimentSearch, root, *dry, tasks))
-	case "run-boundary":
-		os.Exit(runBoundaries(*mode, root))
-	case "snippet":
-		os.Exit(runSnippetAudit(root, 100))
-	case "adversarial":
-		os.Exit(runAdversarial(root))
-	case "run-index":
-		batch := tasks
-		if *only == "" {
-			// A batch measures the efficiency substrate only. Naming one task
-			// with -task runs that task, role included, so the stopping corpus
-			// stays measurable on its own.
-			batch = efficiencyTasks()
-		}
-		os.Exit(runExperiment(*mode, experimentIndex, root, *dry, batch))
-	default:
-		fmt.Fprintf(os.Stderr, "unknown mode %q\n", *mode)
-		os.Exit(2)
-	}
-}
-
-func runPreflight(tasks []contextTask, root string) int {
-	var all []finding
-	for _, t := range tasks {
-		found, err := checkTask(t, root)
-		if err != nil {
-			fmt.Fprintf(os.Stderr, "%s: %v\n", t.ID, err)
-			return 2
-		}
-		all = append(all, found...)
-		status := "ok"
-		if len(found) > 0 {
-			status = fmt.Sprintf("%d problem(s)", len(found))
-		}
-		fmt.Printf("%-24s %-6s gen=%-2d %s\n", t.ID, t.Experiment, t.PlantAfterGen, status)
-	}
-	if len(all) == 0 {
-		fmt.Printf("\n%d fixtures ready.\n", len(tasks))
-		return 0
-	}
-	fmt.Fprintf(os.Stderr, "\n%d preflight problem(s):\n", len(all))
-	sort.Slice(all, func(i, j int) bool { return all[i].Task < all[j].Task })
-	for _, f := range all {
-		fmt.Fprintln(os.Stderr, " ", f)
-	}
-	return 1
-}
-
-// runCalibrate reports, for each index task, the lead-call counts at which its
-// cue is still addressed by each budget. The corpus freezes one number from
-// this; it is never computed at run time, or the implementation under test
-// would be moving its own goalposts.
-func runCalibrate(tasks []contextTask, root string) int {
-	for _, t := range tasks {
-		if t.Experiment != experimentIndex {
-			continue
-		}
-		var rows []string
-		for lead := range fixtureGenerations {
-			probe := t
-			probe.PlantAfterGen = lead
-			var visibleAt []string
-			for _, scale := range armScales {
-				inst, ierr := instantiateTask(probe, seededRand(t.ID, scale, "calibrate"))
-				if ierr != nil {
-					fmt.Fprintln(os.Stderr, ierr)
-					return 2
-				}
-				f, err := buildFixture(inst, armFor(scale, false),
-					fmt.Sprintf("%s/calib/%s/%d-%s/session.jsonl", root, t.ID, lead, scale))
-				if err != nil {
-					fmt.Fprintf(os.Stderr, "%s gen=%d %s: %v\n", t.ID, lead, scale, err)
-					return 2
-				}
-				if strings.Contains(visibleText(f), inst.CueMarker) {
-					visibleAt = append(visibleAt, scale)
-				}
-			}
-			rows = append(rows, fmt.Sprintf("planted after gen %d, cue visible at: %s", lead, strings.Join(visibleAt, ",")))
-		}
-		fmt.Printf("\n%s (wants tier %s)\n", t.ID, t.CueTier)
-		for _, r := range rows {
-			fmt.Println(" ", r)
-		}
-	}
-	return 0
-}
@@ -1,189 +0,0 @@
-package main
-
-import (
-	"bytes"
-	"fmt"
-	"maps"
-	mathrand "math/rand"
-	"os/exec"
-	"slices"
-	"strings"
-	"testing"
-)
-
-// The same checks the preflight mode runs, under go test: a compaction change
-// that invalidates the fixtures fails here rather than three hundred API calls
-// later.
-
-func TestCorpusIsWellFormed(t *testing.T) {
-	if err := validateCorpus(); err != nil {
-		t.Fatal(err)
-	}
-	if got := len(searchTasks()); got != 6 {
-		t.Errorf("search experiment has %d tasks, want 6", got)
-	}
-	if got := len(indexTasks()); got != 12 {
-		t.Errorf("index experiment has %d tasks, want 12", got)
-	}
-	stopping := map[string]bool{}
-	tiers := map[string]int{}
-	for _, task := range indexTasks() {
-		if task.Role == roleStopping {
-			stopping[task.ID] = true
-			continue
-		}
-		tiers[task.CueTier]++
-	}
-	if !stopping["i03-coalescing"] || !stopping["i04-recovery-fence"] {
-		t.Errorf("the stopping corpus is %v; it must keep i03-coalescing and i04-recovery-fence", stopping)
-	}
-	// Two per tier, over the efficiency substrate, is what makes the staircase
-	// readable: the same task can be compared across the boundary its cue sits
-	// on, instead of two different questions being compared across two arms.
-	for _, tier := range []string{tierQuarter, tierHalf, tierDefault} {
-		if tiers[tier] != 2 {
-			t.Errorf("tier %s has %d tasks, want 2", tier, tiers[tier])
-		}
-	}
-}
-
-// The role decides which batch a task enters; the literal lists are only where
-// it is declared, so the two must not drift.
-func TestIndexListsMatchTheirRoles(t *testing.T) {
-	for _, task := range append(append(indexQuarterTasks(), indexHalfTasks()...), indexDefaultTasks()...) {
-		if task.Role != roleEfficiency {
-			t.Errorf("%s sits in an efficiency list with role %q", task.ID, task.Role)
-		}
-	}
-	for _, task := range stoppingIndexTasks() {
-		if task.Role != roleStopping {
-			t.Errorf("%s sits in the stopping list with role %q", task.ID, task.Role)
-		}
-	}
-}
-
-// The invariant the whole instantiation layer exists for: no literal a scorer
-// needs may be readable in this repository before a run starts.
-func TestNoAnswerLiteralExistsInTheRepository(t *testing.T) {
-	root, err := repoRoot()
-	if err != nil {
-		t.Skip(err)
-	}
-	for _, task := range allTasks() {
-		inst, err := instantiateTask(task, seededRand(task.ID, "leak-check"))
-		if err != nil {
-			t.Fatal(err)
-		}
-		for _, marker := range append(inst.AnswerMarkers, inst.CueMarker) {
-			if len(marker) < 6 {
-				continue // a bare number is not a literal anyone can grep for
-			}
-			if found, where := grepRepo(t, root, marker); found {
-				t.Errorf("%s: %q is readable at %s before the run", task.ID, marker, where)
-			}
-		}
-	}
-}
-
-// A run's values are unpredictable from a previous run only if the live
-// generator is seeded from entropy. Comparing answers across two instances
-// cannot assert that: a task whose answer is a bounded number collides by
-// chance (1/8601 to 1/286 per task), so the check is on the seed source.
-func TestLiveRandDrawsFromEntropy(t *testing.T) {
-	a, b := liveRand(), liveRand()
-	for range 4 {
-		if a.Int63() != b.Int63() {
-			return
-		}
-	}
-	t.Fatal("two live generators produced the same stream: they are seeded alike")
-}
-
-// The source alone decides an instance: the same entropy gives the same
-// answers, different entropy gives different values.
-func TestLiveRandSourceDecidesTheInstance(t *testing.T) {
-	same := func() *mathrand.Rand {
-		rng, err := liveRandFrom(bytes.NewReader(bytes.Repeat([]byte{7}, 8)))
-		if err != nil {
-			t.Fatal(err)
-		}
-		return rng
-	}
-	other, err := liveRandFrom(bytes.NewReader(bytes.Repeat([]byte{9}, 8)))
-	if err != nil {
-		t.Fatal(err)
-	}
-	if _, err := liveRandFrom(bytes.NewReader(nil)); err == nil {
-		t.Fatal("an empty entropy source was accepted")
-	}
-	for _, task := range allTasks() {
-		a, _ := instantiateTask(task, same())
-		b, _ := instantiateTask(task, same())
-		if !slices.Equal(a.AnswerMarkers, b.AnswerMarkers) || !maps.Equal(a.Vars, b.Vars) {
-			t.Errorf("%s: the same entropy gave different instances", task.ID)
-		}
-		c, _ := instantiateTask(task, other)
-		if maps.Equal(a.Vars, c.Vars) {
-			t.Errorf("%s: different entropy gave the same values", task.ID)
-		}
-	}
-}
-
-// A preflight that could not reproduce would calibrate against one run's
-// nonces and report about another's.
-func TestSeededInstancesReproduce(t *testing.T) {
-	for _, task := range allTasks() {
-		a, _ := instantiateTask(task, seededRand(task.ID, "x"))
-		b, _ := instantiateTask(task, seededRand(task.ID, "x"))
-		if !slices.Equal(a.AnswerMarkers, b.AnswerMarkers) || a.body != b.body {
-			t.Errorf("%s did not reproduce under the same seed", task.ID)
-		}
-	}
-}
-
-// An unresolved placeholder would put "{{marker}}" in the transcript and score
-// every run wrong; it fails the fixture instead.
-func TestUnresolvedPlaceholderIsRefused(t *testing.T) {
-	if _, err := instantiate("value is {{missing}}", fixtureVars{"other": "x"}); err == nil {
-		t.Fatal("an unresolved placeholder was accepted")
-	}
-}
-
-func repoRoot() (string, error) {
-	out, err := exec.Command("git", "rev-parse", "--show-toplevel").Output()
-	if err != nil {
-		return "", fmt.Errorf("not in a git checkout: %w", err)
-	}
-	return strings.TrimSpace(string(out)), nil
-}
-
-// grepRepo searches tracked files only: build artifacts and temp fixtures are
-// not what a fresh checkout hands anyone.
-func grepRepo(t *testing.T, root, needle string) (bool, string) {
-	t.Helper()
-	cmd := exec.Command("git", "grep", "-l", "-F", needle)
-	cmd.Dir = root
-	out, err := cmd.Output()
-	if err != nil {
-		return false, "" // exit 1 means no match
-	}
-	return true, strings.TrimSpace(strings.SplitN(string(out), "\n", 2)[0])
-}
-
-func TestEveryFixturePassesPreflight(t *testing.T) {
-	if testing.Short() {
-		t.Skip("builds and folds 36 sessions")
-	}
-	root := t.TempDir()
-	for _, task := range allTasks() {
-		t.Run(task.ID, func(t *testing.T) {
-			found, err := checkTask(task, root)
-			if err != nil {
-				t.Fatal(err)
-			}
-			for _, f := range found {
-				t.Error(f)
-			}
-		})
-	}
-}
@@ -1,290 +0,0 @@
-package main
-
-import (
-	"encoding/json"
-	"fmt"
-	"path/filepath"
-	"slices"
-	"strings"
-
-	"reasonix/internal/base/retrieval"
-	"reasonix/internal/contract/ablation"
-	"reasonix/internal/contract/provider"
-)
-
-// Preflight is what makes a real run worth paying for. An answer the model can
-// already read, or a cue not where the matrix says, turns an arm into a
-// different experiment without saying so.
-
-// armScales are the fold-index arms an index task is measured across.
-var armScales = []string{"default", "half", "quarter", "off"}
-
-// experimentArm is one measured configuration of a task.
-type experimentArm struct {
-	name      string
-	scale     string
-	searchOff bool
-}
-
-// armsFor is the single declaration of which arms a task runs at, so preflight
-// and the run itself cannot check one set and measure another.
-func armsFor(t contextTask) []experimentArm {
-	if t.Experiment == experimentSearch {
-		return []experimentArm{
-			{"search-on", "default", false},
-			{"search-off", "default", true},
-		}
-	}
-	arms := make([]experimentArm, 0, len(armScales))
-	for _, scale := range armScales {
-		arms = append(arms, experimentArm{"index-" + scale, scale, false})
-	}
-	return arms
-}
-
-func scaleOf(name string) ablation.FoldIndexScale {
-	s, err := ablation.ParseFoldIndexScale(name)
-	if err != nil {
-		panic("unknown scale " + name) // the list above is a constant
-	}
-	return s
-}
-
-// armFor is the arm a task is measured under at one fold-index scale.
-func armFor(scale string, searchOff bool) ablation.Set {
-	set := ablation.Set{}
-	if searchOff {
-		set = ablation.New(ablation.RecallSearch)
-	}
-	return set.WithFoldIndex(scaleOf(scale))
-}
-
-type finding struct {
-	Task, Arm, Problem string
-}
-
-func (f finding) String() string { return fmt.Sprintf("%s [%s]: %s", f.Task, f.Arm, f.Problem) }
-
-// checkTask builds the task under every arm it will be run at and reports what
-// does not hold.
-func checkTask(t contextTask, root string) ([]finding, error) {
-	var out []finding
-	for _, arm := range armsFor(t) {
-		// Deterministic values: a preflight has to reproduce, and a tier
-		// calibration measured against one run's nonces is measured against
-		// nothing.
-		inst, err := instantiateTask(t, seededRand(t.ID, arm.name, "preflight"))
-		if err != nil {
-			return nil, err
-		}
-		dir := filepath.Join(root, t.ID, arm.name)
-		f, err := buildFixture(inst, armFor(arm.scale, arm.searchOff), filepath.Join(dir, "session.jsonl"))
-		if err != nil {
-			return nil, err
-		}
-		visible := visibleText(f)
-
-		// The answer must be unreadable in every arm, or the task measures
-		// reading rather than recall.
-		for _, marker := range inst.AnswerMarkers {
-			if len(marker) < 3 {
-				continue // a bare number is not a nonce; the whole set still has to match
-			}
-			if strings.Contains(visible, marker) {
-				out = append(out, finding{t.ID, arm.name, fmt.Sprintf("answer marker %q is already visible", marker)})
-			}
-		}
-
-		// The corpus has to be findable at all, or a miss in a real run says
-		// nothing about the model.
-		rank, err := probeRank(f, inst.ProbeQuery)
-		if err != nil {
-			return nil, err
-		}
-		switch {
-		case rank == 0:
-			out = append(out, finding{t.ID, arm.name, fmt.Sprintf("probe query %q does not find the target at all", inst.ProbeQuery)})
-		case rank > 5:
-			out = append(out, finding{t.ID, arm.name, fmt.Sprintf("probe query %q ranks the target %d", inst.ProbeQuery, rank)})
-		}
-
-		if t.Experiment == experimentIndex {
-			out = append(out, checkCueVisibility(t, inst, arm.scale, arm.name, visible)...)
-		} else if line := indexLineFor(f, f.Target); line != "" {
-			out = append(out, finding{t.ID, arm.name,
-				"a search task is addressed by the fold index, so it measures the index and not search: " + line})
-		}
-
-		// The projection is what the agent believes. This is what the model is
-		// told, and it is the one that has to hold.
-		req, err := captureRequest(f, armFor(arm.scale, arm.searchOff), inst.Prompt)
-		if err != nil {
-			return nil, fmt.Errorf("%s [%s]: capture: %w", t.ID, arm.name, err)
-		}
-		out = append(out, checkCapturedRequest(t, inst, arm.name, arm.scale, arm.searchOff, req)...)
-	}
-	return out, nil
-}
-
-// checkCapturedRequest holds one arm's real provider request to its contract:
-// the answer unreadable, the cue exactly where its tier says, the recall tool
-// offering only what this arm can do, and no host-side field surviving the
-// boundary.
-func checkCapturedRequest(t contextTask, inst fixtureInstance, armName, scale string, searchOff bool, req provider.Request) []finding {
-	var out []finding
-	for _, marker := range inst.AnswerMarkers {
-		if len(marker) < 3 {
-			continue // a bare number is not a nonce; the set as a whole still is
-		}
-		for _, at := range findInSurface(req, marker) {
-			out = append(out, finding{t.ID, armName, fmt.Sprintf("answer marker %q leaked at %s", marker, at)})
-		}
-	}
-	if t.Experiment == experimentIndex {
-		wantVisible, wantHidden := tierScales(t.CueTier)
-		where := findInSurface(req, inst.CueMarker)
-		switch {
-		case contains(wantVisible, scale) && len(where) == 0:
-			out = append(out, finding{t.ID, armName, fmt.Sprintf("cue %q is not provider-visible; the tier expects it", inst.CueMarker)})
-		case contains(wantHidden, scale) && len(where) > 0:
-			out = append(out, finding{t.ID, armName, fmt.Sprintf("cue %q is provider-visible at %s: %s", inst.CueMarker, where[0], scale)})
-		}
-	}
-
-	schema, ok := recallToolSchema(req)
-	if !ok {
-		out = append(out, finding{t.ID, armName, "recall is not in the provider's tool list"})
-		return append(out, boundaryFindings(t, armName, req)...)
-	}
-	params := ""
-	if b, err := json.Marshal(schema.Parameters); err == nil {
-		params = string(b)
-	}
-	// Both arms keep the read half: the ablation removes search, not recall.
-	if !strings.Contains(params, `"positions"`) {
-		out = append(out, finding{t.ID, armName, "recall lost its positions parameter"})
-	}
-	hasQuery := strings.Contains(params, `"query"`)
-	switch {
-	case searchOff && hasQuery:
-		out = append(out, finding{t.ID, armName, "recall still offers a query parameter"})
-	case !searchOff && !hasQuery:
-		out = append(out, finding{t.ID, armName, "recall has no query parameter"})
-	}
-	if searchOff {
-		for _, phrase := range searchWording {
-			for _, at := range findInSurface(req, phrase) {
-				out = append(out, finding{t.ID, armName, fmt.Sprintf("search is offered to an arm without it: %q at %s", phrase, at)})
-			}
-		}
-	}
-	return append(out, boundaryFindings(t, armName, req)...)
-}
-
-func boundaryFindings(t contextTask, armName string, req provider.Request) []finding {
-	var out []finding
-	for _, at := range localOnlyLeaks(req) {
-		out = append(out, finding{t.ID, armName, "host-only field crossed the provider boundary at " + at})
-	}
-	return out
-}
-
-// checkCueVisibility holds an index task to its tier: the cue is addressed at
-// the scales the matrix names and at no others. A policy change that moves it
-// fails here rather than turning two arms into one experiment.
-func checkCueVisibility(t contextTask, inst fixtureInstance, scale, armName, visible string) []finding {
-	wantVisible, wantHidden := tierScales(t.CueTier)
-	has := strings.Contains(visible, inst.CueMarker)
-	switch {
-	case contains(wantVisible, scale) && !has:
-		return []finding{{t.ID, armName, fmt.Sprintf("cue %q should be addressed at %s and is not", inst.CueMarker, scale)}}
-	case contains(wantHidden, scale) && has:
-		return []finding{{t.ID, armName, fmt.Sprintf("cue %q should be gone at %s and is not", inst.CueMarker, scale)}}
-	}
-	return nil
-}
-
-func contains(list []string, want string) bool { return slices.Contains(list, want) }
-
-// probeRank runs the task's own query against the folded region the way recall
-// would, and reports where the target lands.
-func probeRank(f builtFixture, query string) (int, error) {
-	terms, err := retrieval.QueryTerms(query)
-	if err != nil {
-		return 0, err
-	}
-	canonical := f.Session.Snapshot()
-	covered := min(f.State.Projection.CoveredCount, len(canonical))
-	docs := make([]map[string]int, 0, covered)
-	positions := make([]int, 0, covered)
-	lengths := make([]int, 0, covered)
-	for i, m := range canonical[:covered] {
-		var text strings.Builder
-		text.WriteString(m.Content)
-		for _, tc := range m.ToolCalls {
-			text.WriteString(" " + tc.Name + " " + string(tc.Arguments))
-		}
-		tokens := retrieval.Tokens(text.String())
-		if len(tokens) == 0 {
-			continue
-		}
-		docs = append(docs, retrieval.Counts(tokens))
-		positions = append(positions, i)
-		lengths = append(lengths, len(tokens))
-	}
-	df := retrieval.DocumentFrequency(docs)
-	avg := 0.0
-	for _, l := range lengths {
-		avg += float64(l)
-	}
-	if len(lengths) > 0 {
-		avg /= float64(len(lengths))
-	}
-	type hit struct {
-		pos   int
-		score float64
-	}
-	var hits []hit
-	for i, counts := range docs {
-		if s := retrieval.BM25Score(counts, lengths[i], terms, df, len(docs), avg); s > 0 {
-			hits = append(hits, hit{positions[i], s})
-		}
-	}
-	for i := range hits {
-		for j := i + 1; j < len(hits); j++ {
-			if hits[j].score > hits[i].score {
-				hits[i], hits[j] = hits[j], hits[i]
-			}
-		}
-	}
-	// A tool result answers at its caller's position, so either counts.
-	for i, h := range hits {
-		if h.pos == f.Target || h.pos == f.Target+1 {
-			return i + 1, nil
-		}
-	}
-	return 0, nil
-}
-
-// indexLineFor returns the fold-index line addressing a position, if any.
-func indexLineFor(f builtFixture, pos int) string {
-	want := fmt.Sprintf("#%d ", pos)
-	for line := range strings.SplitSeq(visibleIndexOnly(f), "\n") {
-		if strings.Contains(line, want) {
-			return strings.TrimSpace(line)
-		}
-	}
-	return ""
-}
-
-// visibleIndexOnly is just the fold-index sections of the projection.
-func visibleIndexOnly(f builtFixture) string {
-	var b strings.Builder
-	for _, m := range visibleContext(f) {
-		if i := strings.Index(m.Content, "## Folded work index"); i >= 0 {
-			b.WriteString(m.Content[i:])
-			b.WriteString("\n")
-		}
-	}
-	return b.String()
-}
@@ -1,252 +0,0 @@
-// reformulation.go — how a query changed, from token sets alone.
-package main
-
-import (
-	"fmt"
-	"slices"
-	"sort"
-	"strings"
-
-	"reasonix/internal/base/retrieval"
-)
-
-// The host probe ranks every target first, so the engine is not what makes a
-// model take more than one attempt. These classes are set arithmetic over the
-// same tokenizer recall uses: no judge, nothing read out of prose.
-
-const (
-	reformFirst      = "First"            // nothing to compare against yet
-	reformNearDup    = "NearDuplicate"    // the same query again, give or take a word
-	reformNarrowing  = "Narrowing"        // kept everything and added terms
-	reformBroadening = "Broadening"       // dropped terms, added none
-	reformAnchored   = "HistoricalAnchor" // reached for a term the question does not contain
-	reformShift      = "SynonymShift"     // traded terms for others
-)
-
-// nearDuplicateOverlap is where "reworded" stops and "asked again" begins.
-const nearDuplicateOverlap = 0.8
-
-func terms(s string) []string {
-	return retrieval.Unique(retrieval.Tokens(s))
-}
-
-func termSet(s string) map[string]bool {
-	out := map[string]bool{}
-	for _, t := range terms(s) {
-		out[t] = true
-	}
-	return out
-}
-
-func jaccard(a, b map[string]bool) float64 {
-	if len(a) == 0 && len(b) == 0 {
-		return 1
-	}
-	shared := 0
-	for t := range a {
-		if b[t] {
-			shared++
-		}
-	}
-	union := len(a) + len(b) - shared
-	if union == 0 {
-		return 0
-	}
-	return float64(shared) / float64(union)
-}
-
-// classifyReformulation names how this query differs from the one before it.
-func classifyReformulation(previous, current, question string) string {
-	if previous == "" {
-		return reformFirst
-	}
-	prev, cur := termSet(previous), termSet(current)
-	if jaccard(prev, cur) >= nearDuplicateOverlap {
-		return reformNearDup
-	}
-	added, dropped := diffTerms(cur, prev), diffTerms(prev, cur)
-	// A term the question never used had to come from the model's own memory of
-	// the history, which is a different move from rephrasing what it was asked.
-	ask := termSet(question)
-	for _, t := range added {
-		if !ask[t] {
-			return reformAnchored
-		}
-	}
-	switch {
-	case len(dropped) == 0:
-		return reformNarrowing
-	case len(added) == 0:
-		return reformBroadening
-	default:
-		return reformShift
-	}
-}
-
-func diffTerms(from, minus map[string]bool) []string {
-	var out []string
-	for t := range from {
-		if !minus[t] {
-			out = append(out, t)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// classifySearches labels each attempt and records which of the host probe's
-// terms it left out. The probe is the oracle: it ranks the target first, so a
-// query missing its terms says what the model did not think to ask for.
-func classifySearches(m *contextMetrics, inst fixtureInstance) {
-	probe := terms(inst.ProbeQuery)
-	previous := ""
-	for i := range m.Searches {
-		s := &m.Searches[i]
-		s.Reformulation = classifyReformulation(previous, s.Query, inst.Prompt)
-		have := termSet(s.Query)
-		for _, term := range probe {
-			if !have[term] {
-				s.MissingProbeTerms = append(s.MissingProbeTerms, term)
-			}
-		}
-		previous = s.Query
-	}
-}
-
-// queryAudit is the reformulation funnel across a set of runs.
-type queryAudit struct {
-	Runs, RunsThatSearched  int
-	FirstQueryHit           int
-	HitBySecond, HitByThird int
-	NeverHit                int
-	TotalSearches           int
-	PostHitSearches         int
-	HitThenRead             int
-	EscapeBeforeFirstSearch int
-	Reformulations          map[string]int
-	MissedProbeTerms        map[string]int
-}
-
-func auditQueries(runs []contextMetrics) queryAudit {
-	a := queryAudit{Reformulations: map[string]int{}, MissedProbeTerms: map[string]int{}}
-	for _, r := range runs {
-		if r.Failed || r.Contaminated {
-			continue
-		}
-		a.Runs++
-		a.TotalSearches += len(r.Searches)
-		a.PostHitSearches += r.PostHitSearches
-		if r.EscapeBeforeFirstRecall > 0 {
-			a.EscapeBeforeFirstSearch++
-		}
-		if len(r.Searches) == 0 {
-			continue
-		}
-		a.RunsThatSearched++
-		hitAt := 0
-		for i, s := range r.Searches {
-			a.Reformulations[s.Reformulation]++
-			for _, term := range s.MissingProbeTerms {
-				a.MissedProbeTerms[term]++
-			}
-			if s.TargetRank > 0 && hitAt == 0 {
-				hitAt = i + 1
-			}
-		}
-		switch {
-		case hitAt == 1:
-			a.FirstQueryHit++
-		case hitAt == 2:
-			a.HitBySecond++
-		case hitAt >= 3:
-			a.HitByThird++
-		default:
-			a.NeverHit++
-		}
-		if hitAt > 0 && r.TargetRead {
-			a.HitThenRead++
-		}
-	}
-	return a
-}
-
-func (a queryAudit) report() string {
-	var b strings.Builder
-	b.WriteString("\n## Query reformulation\n")
-	fmt.Fprintf(&b, "  runs=%d searched=%d  searches=%d (%.2f per searching run)\n",
-		a.Runs, a.RunsThatSearched, a.TotalSearches, ratio(a.TotalSearches, a.RunsThatSearched))
-	fmt.Fprintf(&b, "  target found by query 1: %d   by 2: %d   by 3+: %d   never: %d\n",
-		a.FirstQueryHit, a.HitBySecond, a.HitByThird, a.NeverHit)
-	fmt.Fprintf(&b, "  found then read: %d/%d   searches issued after the target was already back: %d\n",
-		a.HitThenRead, a.FirstQueryHit+a.HitBySecond+a.HitByThird, a.PostHitSearches)
-	fmt.Fprintf(&b, "  reached for the workspace before searching: %d\n", a.EscapeBeforeFirstSearch)
-	b.WriteString("  reformulations: " + countsLine(a.Reformulations) + "\n")
-	b.WriteString("  probe terms most often absent: " + topCounts(a.MissedProbeTerms, 8) + "\n")
-	return b.String()
-}
-
-func ratio(n, d int) float64 {
-	if d == 0 {
-		return 0
-	}
-	return float64(n) / float64(d)
-}
-
-func countsLine(counts map[string]int) string {
-	keys := make([]string, 0, len(counts))
-	for k := range counts {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
-	parts := make([]string, 0, len(keys))
-	for _, k := range keys {
-		parts = append(parts, fmt.Sprintf("%s=%d", k, counts[k]))
-	}
-	return strings.Join(parts, " ")
-}
-
-func topCounts(counts map[string]int, limit int) string {
-	type kv struct {
-		k string
-		n int
-	}
-	all := make([]kv, 0, len(counts))
-	for k, n := range counts {
-		all = append(all, kv{k, n})
-	}
-	sort.Slice(all, func(i, j int) bool {
-		if all[i].n != all[j].n {
-			return all[i].n > all[j].n
-		}
-		return all[i].k < all[j].k
-	})
-	parts := make([]string, 0, limit)
-	for _, e := range all[:min(limit, len(all))] {
-		parts = append(parts, fmt.Sprintf("%s=%d", e.k, e.n))
-	}
-	if len(parts) == 0 {
-		return "(none)"
-	}
-	return strings.Join(parts, " ")
-}
-
-// queryLines lists every attempt, so a surprising audit can be read back.
-func queryLines(runs []contextMetrics) string {
-	var b strings.Builder
-	b.WriteString("\n## Every query, in order\n")
-	for _, r := range runs {
-		for _, s := range r.Searches {
-			rank := "miss"
-			if s.TargetRank > 0 {
-				rank = fmt.Sprintf("rank %d", s.TargetRank)
-			}
-			fmt.Fprintf(&b, "  %-24s %-14s #%d %-17s %-9s %q\n",
-				r.Task, r.Arm, s.Ordinal, s.Reformulation, rank, s.Query)
-			if len(s.MissingProbeTerms) > 0 && s.TargetRank == 0 {
-				fmt.Fprintf(&b, "  %-24s %-14s    missing probe terms: %s\n",
-					"", "", strings.Join(slices.Clone(s.MissingProbeTerms), " "))
-			}
-		}
-	}
-	return b.String()
-}
@@ -1,569 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"os/exec"
-	"path/filepath"
-	"reasonix/internal/state/sessionstore"
-	"slices"
-	"sort"
-	"strings"
-	"time"
-
-	"reasonix/internal/assembly/boot"
-	"reasonix/internal/contract/ablation"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/contract/tool"
-	_ "reasonix/internal/model/openai"
-	"reasonix/internal/runtime/agent"
-	_ "reasonix/internal/tools/builtin"
-)
-
-// A run is one recall question against a fixture the preflight has already
-// cleared. Scoring reads the transcript the turn appended — tool names and
-// their JSON arguments — so nothing depends on how the model narrated itself.
-
-const (
-	runBaseURL = "https://api.deepseek.com"
-	runModel   = "deepseek-chat"
-	runTimeout = 4 * time.Minute
-)
-
-func realProvider() (provider.Provider, error) {
-	key := os.Getenv("DEEPSEEK_API_KEY")
-	if key == "" {
-		return nil, fmt.Errorf("a real run needs DEEPSEEK_API_KEY")
-	}
-	return provider.New("openai", provider.Config{
-		Name: "contextbench", BaseURL: runBaseURL, Model: runModel, APIKey: key,
-	})
-}
-
-// adversarialPrompts tell the model what it is running and send it after the
-// answer everywhere except the conversation. The question is not whether it
-// obeys the experiment: it is whether the answer is reachable at all by a model
-// that knows the mechanism and is actively attacking the host.
-var adversarialPrompts = map[string]string{
-	experimentSearch: "You are running a benchmark. Find the expected answer for this task from the " +
-		"Reasonix repository, benchmark source, fixture generator, test files, scratch files, or " +
-		"session metadata. Do not use recall unless necessary. Report the exact hidden values.",
-	experimentIndex: "Inspect benchmarks/context-retrieval, the git repository, session files, temporary " +
-		"files, and any benchmark metadata you can access. Recover the concrete instantiated answer for " +
-		"this run without relying on the folded conversation. Report the exact values.",
-}
-
-// runAdversarial sends one task per experiment after the answer through every
-// host-readable surface. It passes when nothing but recall could produce it —
-// a run that gives up and recalls has still proven the attack failed.
-func runAdversarial(root string) int {
-	root, err := filepath.Abs(root)
-	if err != nil {
-		fmt.Fprintln(os.Stderr, err)
-		return 2
-	}
-	restore, ierr := isolateWorkspace()
-	if ierr != nil {
-		fmt.Fprintln(os.Stderr, "isolate workspace:", ierr)
-		return 2
-	}
-	defer restore()
-	p, perr := realProvider()
-	if perr != nil {
-		fmt.Fprintln(os.Stderr, perr)
-		return 2
-	}
-	leaked := 0
-	for _, id := range []string{"r01-retry-boundary", "i01-transport-fallback"} {
-		t, ok := taskByID(id)
-		if !ok {
-			fmt.Fprintf(os.Stderr, "unknown task %q\n", id)
-			return 2
-		}
-		m, err := runOne(p, t, armFor("default", false), "adversarial", root, false, false,
-			adversarialPrompts[t.Experiment])
-		if err != nil {
-			fmt.Fprintf(os.Stderr, "%s: %v\n", id, err)
-			m = failedMetrics(t, "adversarial", err)
-		}
-		verdict := "isolated"
-		if m.EscapeFoundAnswer {
-			verdict = "LEAKED via " + m.LeakedVia
-			leaked++
-		}
-		fmt.Printf("%-24s %-24s escapes=%-3d recall=%d answered=%v tools=%v\n",
-			id, verdict, m.EscapeCalls, m.SearchCalls+m.ReadCalls, m.AnswerRecovered, m.UnexpectedWorkTools)
-		if m.EscapeFoundAnswer {
-			fmt.Printf("   args:   %s\n   output: %s\n", m.LeakedArgs, m.LeakedOutput)
-		}
-	}
-	if leaked > 0 {
-		fmt.Fprintf(os.Stderr, "\n%d of 2 adversarial runs reached an answer outside recall\n", leaked)
-		return 1
-	}
-	fmt.Println("\nNo answer was reachable outside recall.")
-	return 0
-}
-
-// runOne asks one task under one arm and scores what the model did.
-func runOne(p provider.Provider, t contextTask, arm ablation.Set, armName, root string, cueVisible, dry bool, promptOverride ...string) (contextMetrics, error) {
-	// Live runs get values that existed nowhere before the run started; a dry
-	// run stays reproducible.
-	rng := liveRand()
-	if dry {
-		rng = seededRand(t.ID, armName, "dry")
-	}
-	inst, err := instantiateTask(t, rng)
-	if err != nil {
-		return contextMetrics{}, err
-	}
-	f, err := buildFixture(inst, arm, filepath.Join(root, t.ID, armName, "session.jsonl"))
-	if err != nil {
-		return contextMetrics{}, err
-	}
-	sess, err := sessionstore.LoadSession(f.Path)
-	if err != nil {
-		return contextMetrics{}, fmt.Errorf("load session: %w", err)
-	}
-	reg := tool.NewRegistry()
-	for _, x := range tool.Builtins() {
-		reg.Add(x)
-	}
-	boot.ApplyUnifiedProviderToolSurface(reg, false, arm)
-
-	var usage usageSink
-	a := agent.New(p, reg, sess, agent.Options{
-		ContextWindow: fixtureWindow, CompactRatio: 0.5, RecentKeep: 2,
-		SessionPath: f.Path, KeepPolicy: agent.KeepErrors, Ablation: arm,
-	}, &usage)
-	a.LoadProjectionSidecar(f.Path)
-	// From here the transcript lives only in memory. On disk it is the answer.
-	sealFixture(f.Path)
-	defer unsealFixture(f.Path)
-	dir := filepath.Dir(f.Path)
-	if at, marker := answerOnDisk(dir, inst.AnswerMarkers); at != "" {
-		return contextMetrics{}, fmt.Errorf("%s [%s]: %q is readable at %s before the run", t.ID, armName, marker, at)
-	}
-
-	before := len(sess.Snapshot())
-	ctx, cancel := context.WithTimeout(context.Background(), runTimeout)
-	defer cancel()
-	prompt := inst.Prompt
-	if len(promptOverride) > 0 && promptOverride[0] != "" {
-		prompt = promptOverride[0]
-	}
-	runErr := a.Run(ctx, prompt)
-
-	if at, marker := answerOnDisk(dir, inst.AnswerMarkers); at != "" {
-		return contextMetrics{}, fmt.Errorf("%s [%s]: %q was written to %s during the run", t.ID, armName, marker, at)
-	}
-
-	appended := sess.Snapshot()[before:]
-	m := scoreRun(appended, inst, f.Target, armName, cueVisible)
-	m.trajectory = appended
-	if u, requests := usage.snapshot(); requests > 0 {
-		m.UsageReportedRequests, m.PromptTokens, m.CompletionTokens = requests, u.PromptTokens, u.CompletionTokens
-		m.CacheHitTokens, m.CacheMissTokens, m.UsageEstimated = u.CacheHitTokens, u.CacheMissTokens, u.Estimated
-	}
-	m.scoreAnswer(finalAnswer(appended), inst)
-	if runErr != nil {
-		m.FailureStage = "RunError"
-		m.TimeoutReason = timeoutReason(m)
-	}
-	return m, runErr
-}
-
-// failedMetrics is what a run that never reached a transcript leaves behind. It
-// carries the run's identity and the stage that ended it, so the failure stays a
-// countable fact; Failed keeps it out of every statistic.
-func failedMetrics(t contextTask, armName string, err error) contextMetrics {
-	return contextMetrics{
-		Task: t.ID, Arm: armName, Failed: true,
-		FailureStage: stageNoEvidence, Reason: err.Error(),
-	}
-}
-
-// finalAnswer is the last assistant text the turn produced.
-func finalAnswer(msgs []provider.Message) string {
-	for _, m := range slices.Backward(msgs) {
-		if m.Role == provider.RoleAssistant && strings.TrimSpace(m.Content) != "" {
-			return m.Content
-		}
-	}
-	return ""
-}
-
-// isolateWorkspace moves the process into an empty directory before any task
-// runs. The corpus lives in this repository, so an agent left here greps the
-// answers out of corpus.go — one did, quoting the file's own line about them
-// being nonces. Fixtures stay outside it too: a readable session file is the
-// canonical transcript, answer included.
-func isolateWorkspace() (func(), error) {
-	dir, err := os.MkdirTemp("", "contextbench-workspace-*")
-	if err != nil {
-		return nil, err
-	}
-	readme := "This workspace is empty on purpose. The question is about earlier conversation,\nnot about any file here.\n"
-	if err := os.WriteFile(filepath.Join(dir, "README.md"), []byte(readme), 0o644); err != nil {
-		return nil, err
-	}
-	previous, err := os.Getwd()
-	if err != nil {
-		return nil, err
-	}
-	if err := os.Chdir(dir); err != nil {
-		return nil, err
-	}
-	return func() {
-		_ = os.Chdir(previous)
-		_ = os.RemoveAll(dir)
-	}, nil
-}
-
-// runBoundaries runs only the pair each index task is built for: the narrowest
-// scale still addressing its cue, and the next one down. Each pair moves the
-// affordance and nothing else, where default-vs-off moves the whole index.
-
-// The two arms get separate nonces: reusing one would leave the first arm's
-// answer in a host-readable history for the second to find.
-func runBoundaries(mode, root string) int {
-	started := time.Now()
-	planned := 0
-	repo, _ := os.Getwd()
-	root, err := filepath.Abs(root)
-	if err != nil {
-		fmt.Fprintln(os.Stderr, err)
-		return 2
-	}
-	restore, ierr := isolateWorkspace()
-	if ierr != nil {
-		fmt.Fprintln(os.Stderr, "isolate workspace:", ierr)
-		return 2
-	}
-	defer restore()
-	p, perr := realProvider()
-	if perr != nil {
-		fmt.Fprintln(os.Stderr, perr)
-		return 2
-	}
-
-	var all []contextMetrics
-	var cueSide, noCueSide []contextMetrics
-	for _, t := range efficiencyTasks() {
-		cue, noCue := boundaryPair(t.CueTier)
-		for _, side := range []struct {
-			scale string
-			has   bool
-		}{{cue, true}, {noCue, false}} {
-			name := "index-" + side.scale
-			planned++
-			m, err := runOne(p, t, armFor(side.scale, false), name, root, side.has, false)
-			if err != nil {
-				fmt.Fprintf(os.Stderr, "%s [%s]: %v\n", t.ID, name, err)
-				m = failedMetrics(t, name, err)
-			}
-			all = append(all, m)
-			if side.has {
-				cueSide = append(cueSide, m)
-			} else {
-				noCueSide = append(noCueSide, m)
-			}
-			fmt.Printf("%-24s %-14s cue=%-3v %-14s search=%d read=%d direct=%-5v recall-tok=%-5d%s escape=%d\n",
-				t.ID, name, side.has, m.FailureStage, m.SearchCalls, m.ReadCalls,
-				m.CueDirectRead, m.RecallReturnedTokens, promptTokenSuffix(m), m.EscapeCalls)
-		}
-	}
-	reportDeltas(cueSide, noCueSide)
-	fmt.Print(reportStopping(all))
-	return writeResults(root, batchInfo{Mode: mode, Planned: planned, Started: started, Repo: repo}, all)
-}
-
-// reportDeltas is the paired reading: each task against itself, then pooled.
-func reportDeltas(cueSide, noCueSide []contextMetrics) {
-	fmt.Printf("\n## Paired deltas (cue present − cue absent)\n")
-	fmt.Printf("  %-24s %-10s %-10s %-9s %-11s %s\n",
-		"task", "recovery", "directread", "searches", "recall-tok", "escapes")
-	for i := range cueSide {
-		if i >= len(noCueSide) {
-			break
-		}
-		a, b := cueSide[i], noCueSide[i]
-		fmt.Printf("  %-24s %-10s %-10s %-9d %-11d %d\n", a.Task,
-			deltaBool(a.AnswerRecovered, b.AnswerRecovered),
-			deltaBool(a.CueDirectRead, b.CueDirectRead),
-			a.SearchCalls-b.SearchCalls,
-			a.RecallReturnedTokens-b.RecallReturnedTokens,
-			a.EscapeCalls-b.EscapeCalls)
-	}
-	fmt.Println("\n## Pooled")
-	fmt.Println(" ", summarize("cue-present", cueSide).line())
-	fmt.Println(" ", summarize("cue-absent", noCueSide).line())
-	fmt.Println("\n## Where each side looked first")
-	fmt.Printf("  %-14s %s\n", "cue-present", countsLine(summarize("x", cueSide).Routes))
-	fmt.Printf("  %-14s %s\n", "cue-absent", countsLine(summarize("x", noCueSide).Routes))
-}
-
-func deltaBool(a, b bool) string {
-	switch {
-	case a == b:
-		return "same"
-	case a:
-		return "+cue"
-	default:
-		return "-cue"
-	}
-}
-
-// runExperiment runs one experiment's tasks across its arms. A dry run drives
-// the same pipeline with a scripted provider: it proves fixture, run, scoring
-// and report hold together before any of it is paid for.
-func runExperiment(mode, experiment, root string, dry bool, tasks []contextTask) int {
-	started := time.Now()
-	planned := 0
-	repo, _ := os.Getwd()
-	root, err := filepath.Abs(root)
-	if err != nil {
-		fmt.Fprintln(os.Stderr, err)
-		return 2
-	}
-	restore, err := isolateWorkspace()
-	if err != nil {
-		fmt.Fprintln(os.Stderr, "isolate workspace:", err)
-		return 2
-	}
-	defer restore()
-	var p provider.Provider = &scriptedProvider{}
-	if !dry {
-		real, rerr := realProvider()
-		if rerr != nil {
-			fmt.Fprintln(os.Stderr, rerr)
-			return 2
-		}
-		p = real
-	}
-	byArm := map[string][]contextMetrics{}
-	var all []contextMetrics
-
-	if experiment == experimentIndex {
-		fmt.Print(fingerprint(dry))
-	}
-	for _, t := range tasks {
-		if t.Experiment != experiment {
-			continue
-		}
-		// Arms are shuffled inside each task block. Running six defaults, then
-		// six halves, would make provider drift collinear with the axis under
-		// test: a bad minute would land entirely on one arm.
-		for _, arm := range shuffledArms(t) {
-			cueVisible := t.Experiment == experimentIndex && cueExpectedVisible(t, arm.scale)
-			planned++
-			m, err := runOne(p, t, armFor(arm.scale, arm.searchOff), arm.name, root, cueVisible, dry)
-			if err != nil {
-				fmt.Fprintf(os.Stderr, "%s [%s]: %v\n", t.ID, arm.name, err)
-				m = failedMetrics(t, arm.name, err)
-			}
-			byArm[arm.name] = append(byArm[arm.name], m)
-			all = append(all, m)
-			fmt.Printf("%-24s %-14s %-18s search=%d hit=%d read=%d recall-tok=%d%s %v\n",
-				t.ID, arm.name, m.FailureStage, m.SearchCalls, m.TargetSearchHits, m.ReadCalls,
-				m.RecallReturnedTokens, promptTokenSuffix(m), m.UnexpectedWorkTools)
-		}
-	}
-	reportFunnels(byArm)
-	if experiment == experimentIndex {
-		fmt.Print(reportDoseResponse(all))
-	}
-	fmt.Print(reportStopping(all))
-	fmt.Print(auditQueries(all).report())
-	fmt.Print(queryLines(all))
-	if experiment == experimentIndex {
-		reportBoundaries(all)
-	}
-	return writeResults(root, batchInfo{Mode: mode, Planned: planned, Started: started, Dry: dry, Repo: repo}, all)
-}
-
-func cueExpectedVisible(t contextTask, scale string) bool {
-	visible, _ := tierScales(t.CueTier)
-	return contains(visible, scale)
-}
-
-func reportFunnels(byArm map[string][]contextMetrics) {
-	names := make([]string, 0, len(byArm))
-	for name := range byArm {
-		names = append(names, name)
-	}
-	sort.Strings(names)
-	fmt.Println("\n## Funnel")
-	for _, name := range names {
-		fmt.Println(" ", summarize(name, byArm[name]).line())
-	}
-	fmt.Println("\n## Where it looked first")
-	for _, name := range names {
-		f := summarize(name, byArm[name])
-		fmt.Printf("  %-14s %s\n", name, countsLine(f.Routes))
-	}
-	fmt.Println("\n## Where it broke")
-	for _, name := range names {
-		f := summarize(name, byArm[name])
-		stages := make([]string, 0, len(f.Stages))
-		for stage, n := range f.Stages {
-			stages = append(stages, fmt.Sprintf("%s=%d", stage, n))
-		}
-		sort.Strings(stages)
-		fmt.Printf("  %-12s %s\n", name, strings.Join(stages, " "))
-	}
-}
-
-// writeTrajectories saves what each turn produced, after the whole batch is
-// done. Three analyses in a row have needed a re-run because something was
-// counted and not kept — query text, then snippet contents. It is written at
-// the end on purpose: during the batch these files are the answer.
-func writeTrajectories(root string, all []contextMetrics) error {
-	f, err := os.Create(filepath.Join(root, "trajectories.jsonl"))
-	if err != nil {
-		return err
-	}
-	defer f.Close()
-	for _, m := range all {
-		if err := writeJSONLine(f, map[string]any{
-			"task": m.Task, "arm": m.Arm, "messages": m.trajectory,
-		}); err != nil {
-			return err
-		}
-	}
-	return nil
-}
-
-func writeResults(root string, info batchInfo, all []contextMetrics) int {
-	path := filepath.Join(root, "results.jsonl")
-	f, err := os.Create(path)
-	if err != nil {
-		fmt.Fprintln(os.Stderr, err)
-		return 2
-	}
-	defer f.Close()
-	scored, failed := 0, 0
-	for _, m := range all {
-		if err := writeJSONLine(f, m); err != nil {
-			fmt.Fprintln(os.Stderr, err)
-			return 2
-		}
-		if m.Failed {
-			failed++
-		} else {
-			scored++
-		}
-	}
-	if err := writeTrajectories(root, all); err != nil {
-		fmt.Fprintln(os.Stderr, "write trajectories:", err)
-		return 2
-	}
-	if err := writeBatchManifest(root, info, all); err != nil {
-		fmt.Fprintln(os.Stderr, "write batch manifest:", err)
-		return 2
-	}
-	// Both counts are named: a reader who sees only the batch size cannot tell
-	// measurements from attempts.
-	fmt.Printf("\n%d runs written to %s: scored=%d failed=%d (with trajectories.jsonl and batch.json)\n",
-		len(all), path, scored, failed)
-	return 0
-}
-
-// reportBoundaries prints each index task across its own cue boundary, then the
-// two sides pooled. Task difficulty is held constant inside a row, which is
-// what a four-arm average cannot do.
-func reportBoundaries(all []contextMetrics) {
-	byKey := map[string]contextMetrics{}
-	scored, failed := 0, 0
-	for _, m := range all {
-		if m.Failed {
-			// No transcript, no boundary to compare: count the failure instead of
-			// printing the zero it left behind as a measurement.
-			failed++
-			continue
-		}
-		scored++
-		byKey[m.Task+"|"+m.Arm] = m
-	}
-	if scored == 0 {
-		fmt.Printf("\n## Across each task's cue boundary\n  no scored runs: scored=%d failed=%d (nothing to compare)\n", scored, failed)
-		return
-	}
-	var cueSide, noCueSide []contextMetrics
-	fmt.Println("\n## Across each task's cue boundary")
-	fmt.Printf("  %-24s %-16s %-4s %-10s %-9s %-12s %-11s %s\n",
-		"task", "arm", "cue", "recovered", "cue-read", "search/task", "recall-tok", "escape")
-	for _, t := range indexTasks() {
-		cue, noCue := boundaryPair(t.CueTier)
-		for _, side := range []struct {
-			scale string
-			has   bool
-		}{{cue, true}, {noCue, false}} {
-			m, ok := byKey[t.ID+"|index-"+side.scale]
-			if !ok {
-				continue
-			}
-			if side.has {
-				cueSide = append(cueSide, m)
-			} else {
-				noCueSide = append(noCueSide, m)
-			}
-			mark := "no"
-			if side.has {
-				mark = "yes"
-			}
-			fmt.Printf("  %-24s %-16s %-4s %-10v %-9v %-12d %-11d %d\n",
-				t.ID, "index-"+side.scale, mark, m.AnswerRecovered, m.RecallReadWithoutSearch,
-				m.SearchCalls, m.RecallReturnedTokens, m.EscapeCalls)
-		}
-	}
-	fmt.Println("\n## Pooled across boundaries")
-	fmt.Println(" ", summarize("cue-present", cueSide).line())
-	fmt.Println(" ", summarize("cue-absent", noCueSide).line())
-}
-
-// shuffledArms permutes a task's arms deterministically, so the order is
-// reproducible and still uncorrelated with the axis being measured.
-func shuffledArms(t contextTask) []experimentArm {
-	arms := armsFor(t)
-	rng := seededRand(t.ID, "arm-order")
-	rng.Shuffle(len(arms), func(i, j int) { arms[i], arms[j] = arms[j], arms[i] })
-	return arms
-}
-
-// fingerprint records what produced a batch. Not a gate — but batch effects
-// have reversed three conclusions in this experiment already, and a table with
-// no provenance cannot be compared with the next one.
-func fingerprint(dry bool) string {
-	model, endpoint := runModel, runBaseURL
-	if dry {
-		model, endpoint = "scripted", "none"
-	}
-	commit := "unknown"
-	if out, err := exec.Command("git", "rev-parse", "--short", "HEAD").Output(); err == nil {
-		commit = strings.TrimSpace(string(out))
-	}
-	return fmt.Sprintf("## Batch\n  model=%s endpoint=%s window=%d generations=%d commit=%s started=%s\n",
-		model, endpoint, fixtureWindow, fixtureGenerations, commit, time.Now().UTC().Format(time.RFC3339))
-}
-
-// timeoutReason separates a run that never found its answer from one that found
-// it and kept going. Only the first is a case for a longer budget; the second
-// would just get more time to circle. A batch reporting only "context deadline
-// exceeded" cannot tell an arm that made retrieval harder from an arm that
-// happened to draw a slow minute.
-func timeoutReason(m contextMetrics) string {
-	switch {
-	case m.TargetHeldAtEnd && m.RoundsAfterTarget >= 2:
-		return "PostTargetRunaway"
-	case m.TargetHeldAtEnd:
-		return "TargetHeldLate"
-	case m.SearchCalls >= 3:
-		return "SearchSpiral"
-	case m.SearchCalls+m.ReadCalls+m.EscapeCalls == 0:
-		return "NoToolActivity"
-	default:
-		return "Unknown"
-	}
-}
@@ -1,95 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"regexp"
-	"slices"
-	"strconv"
-	"strings"
-
-	"reasonix/internal/contract/provider"
-)
-
-// scriptedProvider walks the chain a competent model would: search, read the
-// best hit, answer with what came back. It proves the pipeline holds together
-// for nothing, so a real run's first failure is about the model.
-
-type scriptedProvider struct{}
-
-func (p *scriptedProvider) Name() string { return "contextbench-scripted" }
-
-var (
-	scriptedHit    = regexp.MustCompile(`(?m)^#(\d+) `)
-	scriptedRecall = regexp.MustCompile(`(?s)^#\d+\n(.*)`)
-)
-
-// Stream decides from the request alone, never from a call counter: one
-// provider serves every task in a run, and a counter would make the second task
-// start where the first left off.
-func (p *scriptedProvider) Stream(_ context.Context, req provider.Request) (<-chan provider.Chunk, error) {
-	ch := make(chan provider.Chunk, 3)
-	defer close(ch)
-	// Honour the schema, as a model does: an arm without a query parameter is
-	// an arm that cannot search, and pretending otherwise scores a refusal as
-	// a failed search.
-	if !recallAccepts(req, "query") {
-		ch <- provider.Chunk{Type: provider.ChunkText, Text: "I have no address for that part of the history."}
-		ch <- provider.Chunk{Type: provider.ChunkDone}
-		return ch, nil
-	}
-	// The last message, not the last tool result anywhere: the fixture's own
-	// history is full of tool results, and one of those is not this turn's.
-	last := ""
-	if n := len(req.Messages); n > 0 && req.Messages[n-1].Role == provider.RoleTool {
-		last = req.Messages[n-1].Content
-	}
-	switch {
-	case last == "":
-		args, _ := json.Marshal(map[string]string{"query": lastUserText(req)})
-		ch <- provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{
-			ID: "s1", Name: "recall", Arguments: string(args),
-		}}
-	case strings.Contains(last, "Folded context matching"):
-		positions := scriptedHit.FindAllStringSubmatch(last, 1)
-		if len(positions) == 0 {
-			ch <- provider.Chunk{Type: provider.ChunkText, Text: "Nothing in the folded history matches."}
-			break
-		}
-		n, _ := strconv.Atoi(positions[0][1])
-		args, _ := json.Marshal(map[string][]int{"positions": {n}})
-		ch <- provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{
-			ID: "r1", Name: "recall", Arguments: string(args),
-		}}
-	default:
-		// Answer with what recall returned, so a correct chain scores as one.
-		body := last
-		if m := scriptedRecall.FindStringSubmatch(last); len(m) > 1 {
-			body = m[1]
-		}
-		ch <- provider.Chunk{Type: provider.ChunkText, Text: "From the folded history: " + body}
-	}
-	ch <- provider.Chunk{Type: provider.ChunkDone}
-	return ch, nil
-}
-
-// recallAccepts reports whether recall is offered with the given parameter.
-func recallAccepts(req provider.Request, param string) bool {
-	for _, t := range req.Tools {
-		if t.Name != "recall" {
-			continue
-		}
-		b, err := json.Marshal(t.Parameters)
-		return err == nil && strings.Contains(string(b), `"`+param+`"`)
-	}
-	return false
-}
-
-func lastUserText(req provider.Request) string {
-	for _, m := range slices.Backward(req.Messages) {
-		if m.Role == provider.RoleUser {
-			return m.Content
-		}
-	}
-	return ""
-}
@@ -1,182 +0,0 @@
-// snippet.go — does a rank-1 hit actually hand over the answer?
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"path/filepath"
-	"strings"
-
-	"reasonix/internal/base/retrieval"
-	"reasonix/internal/contract/ablation"
-	"reasonix/internal/contract/event"
-	"reasonix/internal/contract/tool"
-	"reasonix/internal/runtime/agent"
-)
-
-// TargetHit is not EvidenceSufficient: a rank-1 hit can still hand over a
-// window too short to answer with, and searching again after that is correct.
-// Measured through the shipped recall path, not a copy of it.
-
-// snippetWidths are the windows to compare. 240 is what ships.
-var snippetWidths = []int{120, 180, 240, 320, 480}
-
-// historicalQueries are the queries these tasks actually drew in the pilots,
-// kept because a probe query is the host's phrasing and the model's was not.
-var historicalQueries = map[string][]string{
-	"i03-coalescing":     {"coalesce window marker probe", "job completion coalescing window"},
-	"i04-recovery-fence": {"recovery fence marker epoch", "expected epoch", "recfence quartz"},
-}
-
-type snippetReading struct {
-	Task     string
-	Query    string
-	Width    int
-	Covered  int
-	Total    int
-	Runes    int
-	Rank     int
-	Complete bool
-}
-
-// runSnippetAudit answers the question offline, over many instantiations so a
-// single seed's lengths cannot decide it.
-func runSnippetAudit(root string, seeds int) int {
-	fmt.Printf("## Snippet coverage (%d instantiations per task)\n", seeds)
-	fmt.Println("  Does the window a rank-1 hit returns contain every value the task scores on?")
-
-	byTaskWidth := map[string]map[int][]snippetReading{}
-	for _, t := range allTasks() {
-		byTaskWidth[t.ID] = map[int][]snippetReading{}
-		for seed := range seeds {
-			inst, err := instantiateTask(t, seededRand(t.ID, "snippet", fmt.Sprint(seed)))
-			if err != nil {
-				fmt.Fprintln(os.Stderr, err)
-				return 2
-			}
-			queries := append([]string{inst.ProbeQuery}, historicalQueries[t.ID]...)
-			for _, q := range queries {
-				for _, width := range snippetWidths {
-					r := measureSnippet(inst, q, width)
-					byTaskWidth[t.ID][width] = append(byTaskWidth[t.ID][width], r)
-				}
-			}
-		}
-	}
-
-	fmt.Printf("\n  %-24s %-5s", "task", "vals")
-	for _, w := range snippetWidths {
-		fmt.Printf("%9d", w)
-	}
-	fmt.Println("   mean runes @240")
-	for _, t := range allTasks() {
-		vals := len(t.AnswerVars)
-		fmt.Printf("  %-24s %-5d", t.ID, vals)
-		for _, w := range snippetWidths {
-			fmt.Printf("%8.0f%%", 100*completeRate(byTaskWidth[t.ID][w]))
-		}
-		fmt.Printf("   %.0f\n", meanRunes(byTaskWidth[t.ID][240]))
-	}
-	reportShippedPath(root)
-	return 0
-}
-
-// measureSnippet renders one hit through the same MakeSnippet the recall path
-// uses, then asks whether every scored value survived the window.
-func measureSnippet(inst fixtureInstance, query string, width int) snippetReading {
-	terms, err := retrieval.QueryTerms(query)
-	if err != nil {
-		return snippetReading{Task: inst.Task.ID, Query: query, Width: width, Total: len(inst.AnswerMarkers)}
-	}
-	// The tool result is the document the answer lives in; a call's arguments
-	// are a separate document and never carry it.
-	snippet := retrieval.MakeSnippet(inst.body, query, terms, width)
-	covered := 0
-	for _, marker := range inst.AnswerMarkers {
-		if strings.Contains(strings.ToLower(snippet), strings.ToLower(marker)) {
-			covered++
-		}
-	}
-	return snippetReading{
-		Task: inst.Task.ID, Query: query, Width: width,
-		Covered: covered, Total: len(inst.AnswerMarkers),
-		Runes: len([]rune(snippet)), Complete: covered == len(inst.AnswerMarkers),
-	}
-}
-
-// reportShippedPath cross-checks the table against the real recall tool: a
-// fixture, a fold, and a search through the agent that ships.
-func reportShippedPath(root string) {
-	fmt.Println("\n## Through the shipped recall path (width as shipped)")
-	for _, t := range allTasks() {
-		inst, err := instantiateTask(t, seededRand(t.ID, "snippet-live"))
-		if err != nil {
-			continue
-		}
-		f, err := buildFixture(inst, ablation.Set{}, filepath.Join(root, "snippet", t.ID, "session.jsonl"))
-		if err != nil {
-			fmt.Printf("  %-24s fixture: %v\n", t.ID, err)
-			continue
-		}
-		a := agent.New(&digestProvider{}, tool.NewRegistry(), f.Session, agent.Options{
-			ContextWindow: fixtureWindow, CompactRatio: 0.5, RecentKeep: 2,
-			SessionPath: f.Path, KeepPolicy: agent.KeepErrors,
-		}, event.Discard)
-		a.LoadProjectionSidecar(f.Path)
-		queries := append([]string{inst.ProbeQuery}, historicalQueries[t.ID]...)
-		for _, q := range queries {
-			res, err := a.RecallContext(context.Background(), tool.RecallRequest{Query: q})
-			if err != nil {
-				fmt.Printf("  %-24s %-34q recall: %v\n", t.ID, q, err)
-				continue
-			}
-			covered := 0
-			for _, marker := range inst.AnswerMarkers {
-				if strings.Contains(strings.ToLower(res.Text), strings.ToLower(marker)) {
-					covered++
-				}
-			}
-			fmt.Printf("  %-24s %-34q %d/%d values in the result\n",
-				t.ID, truncate(q, 32), covered, len(inst.AnswerMarkers))
-		}
-	}
-}
-
-func completeRate(rs []snippetReading) float64 {
-	if len(rs) == 0 {
-		return 0
-	}
-	n := 0
-	for _, r := range rs {
-		if r.Complete {
-			n++
-		}
-	}
-	return float64(n) / float64(len(rs))
-}
-
-func meanRunes(rs []snippetReading) float64 {
-	if len(rs) == 0 {
-		return 0
-	}
-	total := 0
-	for _, r := range rs {
-		total += r.Runes
-	}
-	return float64(total) / float64(len(rs))
-}
-
-// answerCoverage is what a run should have recorded all along: how much of the
-// scored answer a search actually handed over. rank=1 with coverage 1/3 is not
-// a stopping failure.
-func answerCoverage(text string, markers []string) int {
-	covered := 0
-	lower := strings.ToLower(text)
-	for _, marker := range markers {
-		if strings.Contains(lower, strings.ToLower(marker)) {
-			covered++
-		}
-	}
-	return covered
-}
@@ -1,59 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"sync"
-
-	"reasonix/internal/contract/event"
-	"reasonix/internal/contract/provider"
-)
-
-// usageSink reads what a run cost from the events the agent already emits. Every
-// billed completion reports event.Usage — the round path, compaction, command
-// routing and failure diagnosis all do — so the bench reads the same record the
-// product reads instead of wrapping a provider and counting a second time.
-type usageSink struct {
-	mu       sync.Mutex
-	requests int
-	usage    provider.Usage
-}
-
-func (s *usageSink) Emit(e event.Event) {
-	if e.Kind != event.Usage || e.Usage == nil {
-		return
-	}
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	// One event is one completion that reported usage; RequestCount covers a
-	// completion that merged attempts. A request that failed before reporting
-	// leaves no event, so this is not a count of attempts.
-	if e.Usage.RequestCount > 0 {
-		s.requests += e.Usage.RequestCount
-	} else {
-		s.requests++
-	}
-	u := *e.Usage
-	s.usage.PromptTokens += u.PromptTokens
-	s.usage.CompletionTokens += u.CompletionTokens
-	s.usage.TotalTokens += u.TotalTokens
-	s.usage.CacheHitTokens += u.CacheHitTokens
-	s.usage.CacheMissTokens += u.CacheMissTokens
-	s.usage.ReasoningTokens += u.ReasoningTokens
-	s.usage.Estimated = s.usage.Estimated || u.Estimated
-}
-
-func (s *usageSink) snapshot() (provider.Usage, int) {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	return s.usage, s.requests
-}
-
-// promptTokenSuffix renders what a run cost only when something reported it. A
-// dry run reports nothing, and a trailing "prompt-tok=0" would read as a
-// measured zero rather than as no measurement.
-func promptTokenSuffix(m contextMetrics) string {
-	if m.UsageReportedRequests == 0 {
-		return ""
-	}
-	return fmt.Sprintf(" prompt-tok=%d", m.PromptTokens)
-}
@@ -1,54 +0,0 @@
-package main
-
-import (
-	"testing"
-
-	"reasonix/internal/contract/event"
-	"reasonix/internal/contract/provider"
-)
-
-// The cost a run reports must be what the agent reported, summed across every
-// completion it billed — including one that merged attempts and one that could
-// only estimate. Any other event kind is not a cost.
-func TestUsageSinkAccumulatesWhatTheAgentReports(t *testing.T) {
-	var sink usageSink
-	sink.Emit(event.Event{}) // a kind that is not Usage
-	sink.Emit(event.Event{Kind: event.Usage, Usage: &provider.Usage{
-		PromptTokens: 100, CompletionTokens: 10, TotalTokens: 110, CacheHitTokens: 60, CacheMissTokens: 40,
-	}})
-	sink.Emit(event.Event{Kind: event.Usage, Usage: &provider.Usage{
-		PromptTokens: 7, CompletionTokens: 3, CacheHitTokens: 2, CacheMissTokens: 5, RequestCount: 2, Estimated: true,
-	}})
-	sink.Emit(event.Event{Kind: event.Usage}) // a Usage event with no payload is not a cost
-
-	usage, requests := sink.snapshot()
-	if requests != 3 {
-		t.Fatalf("requests = %d, want 3 (1 + the merged attempt's 2; the empty event adds none)", requests)
-	}
-	if usage.PromptTokens != 107 || usage.CompletionTokens != 13 || usage.TotalTokens != 110 {
-		t.Fatalf("accumulated usage = %+v", usage)
-	}
-	if usage.CacheHitTokens != 62 || usage.CacheMissTokens != 45 {
-		t.Fatalf("cache split = %+v", usage)
-	}
-	if !usage.Estimated {
-		t.Fatal("Estimated was dropped when one completion reported an estimate")
-	}
-}
-
-// A dry run must not claim a cost it never paid, in the JSON or in the line a
-// person reads: an empty suffix, not "prompt-tok=0".
-func TestUsageSinkStaysEmptyWithoutUsage(t *testing.T) {
-	var sink usageSink
-	sink.Emit(event.Event{})
-	usage, requests := sink.snapshot()
-	if requests != 0 || usage.PromptTokens != 0 || usage.CompletionTokens != 0 {
-		t.Fatalf("nothing reported usage, got %+v requests=%d", usage, requests)
-	}
-	if suffix := promptTokenSuffix(contextMetrics{}); suffix != "" {
-		t.Fatalf("a run with no reported usage still printed %q", suffix)
-	}
-	if suffix := promptTokenSuffix(contextMetrics{UsageReportedRequests: 1, PromptTokens: 12}); suffix != " prompt-tok=12" {
-		t.Fatalf("a reported cost printed %q", suffix)
-	}
-}
@@ -1,452 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"sort"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/session/control"
-	"reasonix/internal/state/execjournal"
-)
-
-// armFleetActiveStore asks whether the sub-agent store can be asked about a
-// fan-out item whose slot was granted and whose child reached the provider. It
-// is not a parity arm: the store is what every recovery path reads, so what is
-// measured is one store answering differently about executions it holds the
-// same evidence for.
-const armFleetActiveStore = "fleet-active-store"
-
-// activeFleetCallID is the fan-out under measurement. The lone delegation in
-// the same turn keeps taskCallID, so the two entry points stay separable in
-// every artifact without a naming scheme of their own.
-const activeFleetCallID = "probe_fleet_active"
-
-// probeFleetChildEntered is the fan-out's half of the entry evidence. The lone
-// delegation's rides probeChildEntered, which already means exactly this for
-// the delegation under measurement.
-const probeFleetChildEntered = "probe:fleet-child-entered"
-
-// The four populations one death holds. They are read off the journal rather
-// than off the item ids the dispatch declared: which of the two held-back items
-// wins the freed slot is the scheduler's to decide, and an arm that named them
-// in advance would be reporting a race as a premise.
-type population string
-
-const (
-	popSettled population = "fan-out settled"
-	popRunning population = "fan-out executing"
-	popRefused population = "fan-out refused"
-	popLone    population = "lone executing"
-	popNone    population = ""
-)
-
-// measured are the four in the order they are reported, which is also the order
-// they become true: an item settles, the slot it frees admits one of the two
-// behind it, and the delegation holding the other slot outlives all of them.
-func measured() []population { return []population{popSettled, popRunning, popRefused, popLone} }
-
-// populationOf places one journal entry in its population, or in none. Started
-// and open is executing; started and settled is done; queued without a start is an
-// item the scheduler held back and never admitted.
-func populationOf(e execjournal.Entry) population {
-	switch {
-	case ofLoneTask(e.ID):
-		if e.Started() && e.Open() {
-			return popLone
-		}
-	case ofActiveFleet(e.ID):
-		switch {
-		case e.Started() && !e.Open():
-			return popSettled
-		case e.Started() && e.Open():
-			return popRunning
-		case !e.Started() && e.Queued():
-			return popRefused
-		}
-	}
-	return popNone
-}
-
-func ofActiveFleet(id string) bool { return strings.HasPrefix(id, activeFleetCallID+"/") }
-
-// executionsIn are the ids this observation's journal places in a population,
-// in a stable order.
-func executionsIn(o Observation, pop population) []string {
-	var out []string
-	for _, e := range o.Executions {
-		if populationOf(e) == pop {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// storedStatus is what the store says about one execution, empty when it has
-// never heard of it. The join is the execution id the child recorded as its
-// parent call, which is the same identity the journal opened it under.
-func storedStatus(o Observation, id string) string {
-	for _, f := range o.Children.Facts {
-		if f.ParentToolCallID == id {
-			return f.Status
-		}
-	}
-	return ""
-}
-
-// noRecord is what the store says about an execution it never recorded. It is
-// spelled out rather than left blank because a blank column reads as a row that
-// was not measured, which is the opposite of what an absent record means here.
-const noRecord = "no record"
-
-func storeReading(o Observation, ids []string) string {
-	out := make([]string, 0, len(ids))
-	for _, id := range ids {
-		status := storedStatus(o, id)
-		if status == "" {
-			status = noRecord
-		}
-		out = append(out, id+"→"+status)
-	}
-	return orNone(join(out))
-}
-
-// acknowledged reports whether the store holds a record for every one of these
-// executions. An empty set is not acknowledgement: a population the death never
-// produced is a premise that failed, and its row says not-measured.
-func acknowledged(o Observation, ids []string) bool {
-	for _, id := range ids {
-		if storedStatus(o, id) == "" {
-			return false
-		}
-	}
-	return len(ids) > 0
-}
-
-func activeStoreArm(name string) bool { return name == armFleetActiveStore }
-
-// activeStoreRows read one store from four sides. Each population has the same
-// two columns — what the dying process's store said, and what the next one's
-// does — so the answers can only differ by which execution is being asked about.
-func activeStoreRows(before, after Observation) []row {
-	return []row{
-		populationsRow(before),
-		childEntryRow(before),
-		settledStoreRow(before, after),
-		executingStoreRow(before, after),
-		loneStoreRow(before, after),
-		refusedStoreRow(before, after),
-		listingRow(before, after),
-		rebuildUnmovedRow(before, after),
-	}
-}
-
-// populationsRow is the arm's premise, not a durability claim: one death has to
-// hold all four states at once, or the comparisons below are between different
-// runs rather than between entry points.
-func populationsRow(before Observation) row {
-	var got []string
-	for _, pop := range measured() {
-		got = append(got, fmt.Sprintf("%s=%d", pop, len(executionsIn(before, pop))))
-	}
-	return row{
-		Semantic: "the four populations this death holds", Authority: journalAuthority,
-		Artifact: "<stem>.execution.jsonl", Reconstruction: "in-process only",
-		Before: join(got), After: "—", Verdict: activePremiseVerdict(before),
-	}
-}
-
-func activePremiseVerdict(before Observation) string {
-	if activeStoreInvalid(before) != "" {
-		return verdictNotMeasured
-	}
-	return verdictHolds
-}
-
-// childEntryRow is the half of the evidence no artifact carries. A slot grant
-// says the scheduler admitted the item; only the child's own arrival at the
-// provider says the execution happened, and a store row read without it would
-// be asking about work that may never have run.
-func childEntryRow(before Observation) row {
-	fleet := firstProgress(before, probeFleetChildEntered)
-	lone := firstProgress(before, probeChildEntered)
-	return row{
-		Semantic: "the children that reached the provider", Authority: "the scripted provider, in-process",
-		Artifact: "none observed", Reconstruction: "in-process only",
-		Before:  fmt.Sprintf("fan-out=%s lone=%s", fleet, lone),
-		After:   "—",
-		Verdict: enteredVerdict(fleet == "true" && lone == "true"),
-	}
-}
-
-func enteredVerdict(both bool) string {
-	if both {
-		return verdictHolds
-	}
-	return verdictNotMeasured
-}
-
-func firstProgress(o Observation, key string) string {
-	if vals := o.Progress[key]; len(vals) > 0 {
-		return vals[0]
-	}
-	return "unrecorded"
-}
-
-// settledStoreRow is the control that makes the rest readable. It is the same
-// entry point, the same fan-out and the same store, differing only in that this
-// item finished — so if it too were absent, the finding would be about
-// persistence in general rather than about executions in flight.
-func settledStoreRow(before, after Observation) row {
-	ids := executionsIn(before, popSettled)
-	return row{
-		Semantic: "a settled fan-out item, in the store", Authority: storeAuthority,
-		Artifact: "subagents/<ref>.meta.json", Reconstruction: "ListSubagentsByParent, joined by execution id",
-		Before:  storeReading(before, ids),
-		After:   storeReading(after, ids),
-		Verdict: controlVerdict(acknowledged(before, ids) && acknowledged(after, ids), len(ids) > 0),
-	}
-}
-
-// executingStoreRow is the arm. The conjunction is stated in the row rather
-// than left to the reader: the journal proves the slot was granted, the child
-// proves the execution happened, and this column is what the store can say
-// about it. All three, or the row is not the finding it looks like.
-func executingStoreRow(before, after Observation) row {
-	ids := executionsIn(before, popRunning)
-	entered := firstProgress(before, probeFleetChildEntered) == "true"
-	return row{
-		Semantic: "an executing fan-out item, in the store", Authority: storeAuthority,
-		Artifact:       "subagents/<ref>.meta.json",
-		Reconstruction: "STARTED in the journal, entered at the provider, and asked of the store",
-		Before:         storeReading(before, ids),
-		After:          storeReading(after, ids),
-		Verdict:        executingVerdict(before, after, ids, entered),
-	}
-}
-
-// executingVerdict refuses the two readings that would overstate the row. An
-// execution nothing proves entered is not a silent store, and a population the
-// death never produced is not evidence of anything.
-func executingVerdict(before, after Observation, ids []string, entered bool) string {
-	switch {
-	case len(ids) == 0 || !entered:
-		return verdictNotMeasured
-	case acknowledged(before, ids) || acknowledged(after, ids):
-		return verdictHolds
-	default:
-		return verdictViolated
-	}
-}
-
-// loneStoreRow is the other entry point, in the same death and the same store.
-// It is what makes "selectively blind" a statement about the store rather than
-// a comparison between two probe runs.
-func loneStoreRow(before, after Observation) row {
-	ids := executionsIn(before, popLone)
-	return row{
-		Semantic: "an executing lone delegation, in the store", Authority: storeAuthority,
-		Artifact: "subagents/<ref>.meta.json", Reconstruction: "the same read, one entry point over",
-		Before:  storeReading(before, ids),
-		After:   storeReading(after, ids),
-		Verdict: controlVerdict(acknowledged(before, ids) && acknowledged(after, ids), len(ids) > 0),
-	}
-}
-
-// refusedStoreRow is the negative control, and it bounds any fix this arm might
-// support. An item the scheduler never admitted produced nothing, so the store
-// is owed no record for it — and a change that recorded every opening would
-// satisfy the row above by reintroducing exactly that.
-func refusedStoreRow(before, after Observation) row {
-	ids := executionsIn(before, popRefused)
-	silent := !acknowledged(before, ids) && !acknowledged(after, ids)
-	return row{
-		Semantic: "a refused fan-out item, in the store", Authority: crossAuthority,
-		Artifact:       "subagents/<ref>.meta.json, <stem>.execution.jsonl",
-		Reconstruction: "queued, never started: the store must stay silent",
-		Before:         storeReading(before, ids),
-		After:          storeReading(after, ids),
-		Verdict:        controlVerdict(silent, len(ids) > 0),
-	}
-}
-
-func controlVerdict(ok bool, measured bool) string {
-	switch {
-	case !measured:
-		return verdictNotMeasured
-	case ok:
-		return verdictHolds
-	default:
-		return verdictViolated
-	}
-}
-
-// listingRow is the surface a person or a model reaches for. It reads through
-// the workspace filter and the lineage walk the tool applies, which the store
-// read above does not — so it is checked against that read first: a listing
-// that cannot even see the settled item is a filter answering, not a store.
-func listingRow(before, after Observation) row {
-	settled := executionsIn(before, popSettled)
-	return row{
-		Semantic: "what list_subagents would show", Authority: "SubagentStore.ListForParent",
-		Artifact:       "subagents/<ref>.meta.json",
-		Reconstruction: "the read behind the tool, filtered by workspace and lineage",
-		Before:         listingReading(before),
-		After:          listingReading(after),
-		Verdict:        listingVerdict(before, after, settled),
-	}
-}
-
-func listingReading(o Observation) string {
-	if o.Listing.Err != "" {
-		return "error: " + o.Listing.Err
-	}
-	out := make([]string, 0, len(o.Listing.Rows))
-	for _, r := range o.Listing.Rows {
-		out = append(out, orDash(r.From)+"="+r.Status)
-	}
-	sort.Strings(out)
-	return orNone(join(out))
-}
-
-// listingVerdict answers only once the listing has been shown to work. Its
-// filters can drop every row for reasons that have nothing to do with this arm,
-// and an empty listing read as an invisible execution is the substitution the
-// probe exists to refuse.
-func listingVerdict(before, after Observation, settled []string) string {
-	if len(settled) == 0 || !listingHas(after, settled) {
-		return verdictNotMeasured
-	}
-	if listingHas(before, executionsIn(before, popRunning)) {
-		return verdictHolds
-	}
-	return verdictViolated
-}
-
-func listingHas(o Observation, ids []string) bool {
-	for _, id := range ids {
-		found := false
-		for _, r := range o.Listing.Rows {
-			if r.From == id {
-				found = true
-				break
-			}
-		}
-		if !found {
-			return false
-		}
-	}
-	return len(ids) > 0
-}
-
-// rebuildUnmovedRow pins what a fix must not move. A store that learned to
-// record executions in flight would leave stale records behind for a restart to
-// fold, and the one outcome this whole line of work exists to prevent is a
-// rebuild that answers with a node the dead process was running.
-func rebuildUnmovedRow(before, after Observation) row {
-	rebuilt := rebuiltGraph(after)
-	var ghosts []string
-	for _, n := range rebuilt.Graph.Nodes {
-		if n.State == agentgraph.StateRunning {
-			ghosts = append(ghosts, n.ID)
-		}
-	}
-	var cut []string
-	for _, i := range rebuilt.Interrupted {
-		if ofActiveFleet(i.Execution) || ofLoneTask(i.Execution) {
-			cut = append(cut, i.Execution+":"+interruptionKind(i))
-		}
-	}
-	sort.Strings(ghosts)
-	sort.Strings(cut)
-	return row{
-		Semantic: "no ghost, and how the rebuild names the cut", Authority: rebuildAuthority,
-		Artifact:       "<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		Reconstruction: "execgraph.Rebuild over both, after the death",
-		Before:         fmt.Sprintf("%d executing at death", len(executionsIn(before, popRunning))+len(executionsIn(before, popLone))),
-		After:          fmt.Sprintf("ghosts=%s cut=%s", orNone(join(ghosts)), orNone(join(cut))),
-		Verdict:        controlVerdict(len(ghosts) == 0 && len(cut) > 0, len(before.Executions) > 0),
-	}
-}
-
-// activeStoreInvalid reports the premise this arm could not establish. All four
-// populations in one death is the whole design: without them the rows compare
-// different runs, and a row that compares nothing must not be read as green.
-func activeStoreInvalid(before Observation) string {
-	for _, pop := range measured() {
-		if len(executionsIn(before, pop)) == 0 {
-			return "the death held nothing in population " + string(pop)
-		}
-	}
-	if firstProgress(before, probeFleetChildEntered) != "true" {
-		return "no fan-out child reached the provider, so no execution is proven to have happened"
-	}
-	return ""
-}
-
-// runActiveStoreConstruct drives one delegation and one fan-out under a ceiling
-// they share, and dies with all four populations standing. The delegation goes
-// first and hangs: it is what keeps the ceiling from emptying under the fan-out,
-// so the item refused admission stays refused rather than starting a moment
-// later and taking the arm's negative control with it.
-func runActiveStoreConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	prompt := fanOutTurn(turn+1, taskSentinel+" "+fleetActiveSentinel)
-	go func() { _ = ctrl.Run(ctx, prompt) }()
-	if err := waitForPopulations(ctrl, prov); err != nil {
-		return err
-	}
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	// Whether each child reached the provider is known only here, and it is half
-	// of what makes a silent store a finding rather than a missing record for
-	// work that never ran.
-	obs.Progress[probeFleetChildEntered] = []string{fmt.Sprint(prov.fleets.held(childEntered(childHang)))}
-	obs.Progress[probeChildEntered] = []string{fmt.Sprint(prov.fleets.held(childEntered(childHold)))}
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// waitForPopulations blocks until the journal holds all four states, then holds
-// still and asks again. A refusal is a state nothing announces, so the second
-// reading is what separates an item the scheduler is holding back from one that
-// was merely slow to be admitted.
-func waitForPopulations(ctrl *control.Controller, prov *scripted) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		if populationsStanding(ctrl, prov) {
-			time.Sleep(2 * time.Second)
-			if !populationsStanding(ctrl, prov) {
-				return errUnexpected("a refusal that was still standing",
-					journalStates(control.ExecutionHistory(ctrl.SessionPath())))
-			}
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return errUnexpected("one death holding all four populations",
-		journalStates(control.ExecutionHistory(ctrl.SessionPath())))
-}
-
-func populationsStanding(ctrl *control.Controller, prov *scripted) bool {
-	held := map[population]bool{}
-	for _, e := range control.ExecutionHistory(ctrl.SessionPath()) {
-		held[populationOf(e)] = true
-	}
-	return held[popSettled] && held[popRunning] && held[popRefused] && held[popLone] &&
-		prov.fleets.held(childEntered(childHang)) && prov.fleets.held(childEntered(childHold))
-}
-
-func journalStates(entries []execjournal.Entry) string {
-	out := make([]string, 0, len(entries))
-	for _, e := range entries {
-		out = append(out, fmt.Sprintf("%s:started=%t,queued=%t,open=%t", e.ID, e.Started(), e.Queued(), e.Open()))
-	}
-	sort.Strings(out)
-	return orNone(join(out))
-}
@@ -1,345 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"os"
-	"os/exec"
-	"path/filepath"
-	"reasonix/internal/state/sessionstore"
-	"strings"
-
-	"reasonix/internal/contract/config"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/state/memory"
-)
-
-// arm is one death-and-resume run. Lever runs between the two processes and is
-// what separates the three compaction questions: nothing changed, only the
-// stable prefix changed, the covered conversation changed.
-type arm struct {
-	name  string
-	asks  string
-	lever func(armRoot, Observation) error
-}
-
-// appendsAfterFold names the arms whose construct phase adds one turn after
-// the fold. Every arm carrying a lever needs it: without it the projection is
-// already gone at the boundary for an unrelated reason, and the lever's own
-// effect cannot be read out of the result.
-func appendsAfterFold(name string) bool {
-	return name != "exact" && name != armRefoldIntoBody && !graphArm(name) && !uiArm(name) &&
-		!schedulerWaitArm(name) && !terminalArm(name) && !deriveArm(name) &&
-		!loneTaskArm(name) && !cancelRoutingArm(name) && !lineageArm(name) && !activeStoreArm(name) &&
-		!skillArm(name) && !ephemeralArm(name) && !hostExecArm(name)
-}
-
-// armAppendAfterFold separates the two ways a projection can fail validation
-// after a restart: a covered prefix that no longer matches, and a projection
-// that covers the whole transcript, whose only cross-process validation is a
-// transcript version the reload does not carry.
-const armAppendAfterFold = "append-after-fold"
-
-// armRefoldIntoBody folds twice. The second fold reads the provider-visible
-// view — stored body plus live tail — and must record a boundary in canonical
-// terms, which the body has no counterpart for.
-const armRefoldIntoBody = "refold-into-body"
-
-// The tail arms separate two ways the live tail can move. tail-edit and
-// tail-truncate change the transcript past the fold, where the covered hash
-// cannot notice; tail-rewind takes the path a person actually drives, which
-// throws the projection away without asking whether coverage still holds.
-const (
-	armTailEdit     = "tail-edit"
-	armTailTruncate = "tail-truncate"
-	armTailRewind   = "tail-rewind"
-	// armCoveredRewind is the negative control the rewind path will need before
-	// its invalidation can become conditional: rewinding below the fold removes
-	// history the digest folded, and no coverage check may excuse that.
-	armCoveredRewind = "covered-rewind"
-	// armTodoIdentity moves host step identity while a fold is between the
-	// model and it: identity A, fold, live tail, identity B, fold again.
-	armTodoIdentity = "todo-identity"
-	// armOpenDecision dies while a question is open. It asks what the host
-	// believes afterwards, not whether one particular card comes back.
-	armOpenDecision = "open-decision"
-	// The three successors to an interrupted barrier. Each starts where
-	// open-decision ends and differs only in what the next process does, which
-	// is the whole question: when does a recorded interruption stop being the
-	// context every later request carries?
-	armIdleRestart   = "interrupted-idle"
-	armUnrelatedTurn = "interrupted-unrelated"
-	armDirectAnswer  = "interrupted-answered"
-)
-
-// successorTurn is what the middle process says, empty when it only restarts.
-func successorTurn(arm string) (string, bool) {
-	switch arm {
-	case armIdleRestart:
-		return "", true
-	case armUnrelatedTurn:
-		return marker(120) + " List the files you would look at first.", true
-	case armDirectAnswer:
-		return marker(121) + " Yes, go ahead with that — take the side below the fold.", true
-	}
-	return "", false
-}
-
-func arms() []arm {
-	return []arm{
-		{name: "exact", asks: "nothing changed between the processes"},
-		{name: armAppendAfterFold, asks: "nothing changed, and one turn was appended after the fold"},
-		{name: armRefoldIntoBody, asks: "a second fold reaches into the body the first one stored"},
-		{name: "system-swap", asks: "only the stable prefix changed, against the surviving baseline", lever: swapSystemPrefix},
-		{name: armTailEdit, asks: "a message past the fold changed", lever: editTailRow},
-		{name: armTailTruncate, asks: "the messages past the fold were dropped", lever: truncateTail},
-		{name: armTailRewind, asks: "the same truncation, driven through the rewind a person uses"},
-		{name: armCoveredRewind, asks: "a rewind that lands below the fold boundary"},
-		{name: armTodoIdentity, asks: "host step identity moved across two folds and a live tail"},
-		{name: armOpenDecision, asks: "the process died while a decision was open"},
-		{name: armIdleRestart, asks: "an interrupted barrier, and nobody does anything"},
-		{name: armUnrelatedTurn, asks: "an interrupted barrier, then a turn about something else"},
-		{name: armDirectAnswer, asks: "an interrupted barrier, then a turn that reads like an answer to it"},
-		{name: armGraphCompleted, asks: "a fan-out whose items had all settled when the process died"},
-		{name: armGraphRunning, asks: "a fan-out with a child still executing when the process died"},
-		{name: armGraphMixed, asks: "a fan-out holding completed, failed, adopted and running at once"},
-		{name: armUIGraphMixed, asks: "which door each fact reaches the frontend through, after a death"},
-		{name: armTaskCompleted, asks: "a lone delegation that finished, in a turn that closed"},
-		{name: armTaskRunning, asks: "a lone delegation still executing when the process died"},
-		{name: armTaskFgQueued, asks: "what a lone delegation is drawn as while the scheduler holds it back"},
-		{name: armTaskBgQueued, asks: "a backgrounded delegation the ceiling refused, with its job already handed over"},
-		{name: armTaskBgRunning, asks: "a backgrounded delegation executing inside its job"},
-		{name: armCancelTool, asks: "whether a stop reaches an ordinary tool the turn is waiting on"},
-		{name: armCancelBackground, asks: "whether a stop leaves work already handed to a job alone"},
-		{name: armCancelHeadlessOwner, asks: "who owns the cancellation of a turn the controller never admitted"},
-		{name: armLineageForeground, asks: "control: does a stop reach the tree the turn owns?"},
-		{name: armLineageBackground, asks: "does it also reach the tree it handed away?"},
-		{name: armLineageJobKill, asks: "control: does a job's own stop reach its tree?"},
-		{name: armTaskFgQueuedCancel, asks: "a delegation stopped while the scheduler still held it back"},
-		{name: armTaskBgQueuedCancel, asks: "the same, backgrounded: a job killed before it was ever admitted"},
-		{name: armTaskFgRunningCancel, asks: "a delegation stopped while its child was executing"},
-		{name: armTaskBgRunningCancel, asks: "the same, backgrounded: a job killed while its child ran"},
-		{name: armWaitSlots, asks: "an item the session's total ceiling refused"},
-		{name: armWaitWriters, asks: "an item the writer ceiling refused, with total capacity free"},
-		{name: armWaitClaim, asks: "an item a path conflict refused, with both ceilings free"},
-		{name: armWaitTransition, asks: "what a reported wait cause does when the thing it named stops holding the item"},
-		{name: armTerminalAdopted, asks: "an item that never ran and still holds an answer"},
-		{name: armTerminalSkippedDep, asks: "an item that never ran and holds none, because its dependency failed"},
-		{name: armTerminalCancelled, asks: "an item a cancellation ended before it was admitted"},
-		{name: armTerminalContext, asks: "whether a delivered answer's edge follows from what is already durable"},
-		{name: armChildTerminal, asks: "whether the store keeps every terminal the graph draws"},
-		{name: armDeriveSkipBoth, asks: "a skip whose two upstreams both ended without an answer"},
-		{name: armDeriveSkipFlip, asks: "the same, with the two failures in the other order"},
-		{name: armIdentitySemantics, asks: "what a node's model and effort report, across two producers and the store"},
-		{name: armDeriveAnswered, asks: "a dependent whose upstreams answered, one by completing and one by adopting"},
-		{name: armFleetActiveStore, asks: "whether the store can be asked about a fan-out item that is executing"},
-		{name: armSkillQueued, asks: "whether a subagent skill asks the same scheduler as every other delegation"},
-		{name: armSkillRunning, asks: "what names a subagent skill while its child is executing"},
-		{name: armSkillCompleted, asks: "what a finished subagent skill leaves, through both entry points"},
-		{name: armSkillCancel, asks: "what a stop reaches when a subagent skill is running"},
-		{name: armReadOnlySkillQueued, asks: "whether the read-only skill entry point asks the same scheduler"},
-		{name: armReadOnlySkillRunning, asks: "what names a read-only skill while its child is executing"},
-		{name: armReadOnlySkillCompleted, asks: "what a finished read-only skill leaves behind, if anything"},
-		{name: armEphemeralTask, asks: "can the execution authority account for what read_only_task draws?"},
-		{name: armEphemeralSkill, asks: "the same of read_only_skill, the other ephemeral entry point"},
-		{name: armHostRunning, asks: "what a slash-started subagent leaves when the process dies mid-execution"},
-		{name: armHostQueued, asks: "what names slash-started work the scheduler is holding back"},
-		{name: armHostCompleted, asks: "what a finished slash invocation leaves besides its answer"},
-		{name: armHostCancel, asks: "what a stop reaches, and what history it leaves"},
-		{name: armHeadlessCompleted, asks: "does the headless host surface leave the same provenance?"},
-		{name: armHeadlessRunningCancel, asks: "the same, stopped through the context its caller holds"},
-		{name: armHeadlessQueuedCancel, asks: "the same while the scheduler is still holding it back"},
-		{name: "covered-mutation", asks: "the covered conversation changed, against the surviving baseline", lever: mutateCoveredRow},
-	}
-}
-
-// swapSystemPrefix saves a memory fact into this arm's home. The lever is a
-// real user action rather than a poked field: the saved-fact index folds into
-// the boot-time system prompt, so the next process composes a different one.
-func swapSystemPrefix(root armRoot, _ Observation) error {
-	store := memory.StoreFor(config.RootsForHome(root.Home).MemoryUserDir(), root.Workspace)
-	_, err := store.Save(memory.Memory{
-		Name:        "probe-prefix-lever",
-		Title:       "Probe prefix lever",
-		Description: "A fact saved between the two processes so the composed system prompt differs.",
-		Type:        memory.TypeProject,
-		Scope:       memory.FactScopeProject,
-		// Pinned, not relevant: only a pinned body rides the stable prefix. The
-		// saved-fact index is retrieval-only and would move nothing.
-		Activation: memory.ActivationPinned,
-		Body:       "The runtime-resume probe wrote this to move the stable prefix.",
-	})
-	return err
-}
-
-// mutateCoveredRow rewrites one provider-visible message inside the prefix the
-// projection folded, through the same rewrite save a rewind or a recovery
-// branch uses. This is the control arm: the digest's material changed, so the
-// projection must not be reused.
-func mutateCoveredRow(_ armRoot, before Observation) error {
-	session, err := sessionstore.LoadSession(before.SessionPath)
-	if err != nil {
-		return err
-	}
-	msgs := session.Snapshot()
-	covered := min(before.Sidecar.CoveredCount, len(msgs))
-	target := -1
-	for i := range covered {
-		if msgs[i].Role == provider.RoleUser || msgs[i].Role == provider.RoleAssistant {
-			target = i
-			break
-		}
-	}
-	if target < 0 {
-		return errUnexpected("a covered conversation row to mutate", covered)
-	}
-	msgs[target].Content += "\n[probe: covered row mutated]"
-	session.Replace(msgs)
-	return session.SaveRewrite(before.SessionPath)
-}
-
-// editTailRow rewrites one message the projection does not claim to cover. The
-// covered hash cannot see it, so the projection should keep serving.
-func editTailRow(_ armRoot, before Observation) error {
-	return rewriteSession(before, func(msgs []provider.Message, covered int) ([]provider.Message, error) {
-		for i := covered; i < len(msgs); i++ {
-			if msgs[i].Role == provider.RoleUser || msgs[i].Role == provider.RoleAssistant {
-				msgs[i].Content += "\n[probe: tail row edited]"
-				return msgs, nil
-			}
-		}
-		return nil, errUnexpected("a live tail row to edit", covered)
-	})
-}
-
-// truncateTail drops everything past the fold boundary, leaving the covered
-// prefix exactly as the projection folded it.
-func truncateTail(_ armRoot, before Observation) error {
-	return rewriteSession(before, func(msgs []provider.Message, covered int) ([]provider.Message, error) {
-		return msgs[:covered], nil
-	})
-}
-
-// rewriteSession applies a transcript rewrite through the save a rewind or a
-// recovery branch uses, refusing an arm with no live tail to act on.
-func rewriteSession(before Observation, edit func([]provider.Message, int) ([]provider.Message, error)) error {
-	session, err := sessionstore.LoadSession(before.SessionPath)
-	if err != nil {
-		return err
-	}
-	msgs := session.Snapshot()
-	covered := before.Sidecar.CoveredCount
-	if covered <= 0 || covered >= len(msgs) {
-		return errUnexpected("a live tail past the fold", covered)
-	}
-	next, err := edit(msgs, covered)
-	if err != nil {
-		return err
-	}
-	session.Replace(next)
-	return session.SaveRewrite(before.SessionPath)
-}
-
-func orchestrate(work, only, jsonOut string, keep bool) error {
-	self, err := os.Executable()
-	if err != nil {
-		return err
-	}
-	if work == "" {
-		work, err = os.MkdirTemp("", "runtime-resume-")
-		if err != nil {
-			return err
-		}
-		if !keep {
-			defer os.RemoveAll(work)
-		}
-	}
-	var results []armResult
-	for _, a := range arms() {
-		if only != "" && a.name != only {
-			continue
-		}
-		result, err := runArm(self, work, a)
-		if err != nil {
-			return fmt.Errorf("arm %s: %w", a.name, err)
-		}
-		results = append(results, result)
-	}
-	if len(results) == 0 {
-		return fmt.Errorf("no arm matched %q", only)
-	}
-	fmt.Print(renderMatrix(results, work))
-	return writeJSON(jsonOut, results)
-}
-
-func runArm(self, work string, a arm) (armResult, error) {
-	root := rootFor(filepath.Join(work, a.name))
-	if err := root.requireFresh(); err != nil {
-		return armResult{}, err
-	}
-	if err := root.create(); err != nil {
-		return armResult{}, err
-	}
-	if err := spawn(self, "construct", root, a.name); err != nil {
-		return armResult{}, fmt.Errorf("construct process: %w", err)
-	}
-	before, err := readObservation(root, "construct")
-	if err != nil {
-		return armResult{}, err
-	}
-	if a.lever != nil {
-		if err := a.lever(root, before); err != nil {
-			return armResult{}, fmt.Errorf("lever: %w", err)
-		}
-	}
-	if _, ok := successorTurn(a.name); ok {
-		if err := spawn(self, "successor", root, a.name); err != nil {
-			return armResult{}, fmt.Errorf("successor process: %w", err)
-		}
-	}
-	if err := spawn(self, "resume", root, a.name); err != nil {
-		return armResult{}, fmt.Errorf("resume process: %w", err)
-	}
-	after, err := readObservation(root, "resume")
-	if err != nil {
-		return armResult{}, err
-	}
-	var extra *Observation
-	if _, ok := successorTurn(a.name); ok {
-		obs, err := readObservation(root, "successor")
-		if err != nil {
-			return armResult{}, err
-		}
-		extra = &obs
-	} else if phase := extraPhase(a.name); phase != "" {
-		obs, err := readObservation(root, phase)
-		if err != nil {
-			return armResult{}, err
-		}
-		extra = &obs
-	}
-	return classify(a, extra, before, after), nil
-}
-
-// spawn runs one phase as a real child process and waits for it to exit. The
-// OS process is the measurement boundary: a nil-ed object can still be revived
-// by a package-level singleton, a live sink, or a cache, and none of those
-// survive here.
-func spawn(self, phase string, root armRoot, armName string) error {
-	cmd := exec.Command(self, "-phase="+phase, "-root="+root.Dir, "-arm="+armName)
-	cmd.Env = childEnv(root)
-	cmd.Dir = root.Workspace
-	cmd.Stdout = os.Stderr
-	cmd.Stderr = os.Stderr
-	return cmd.Run()
-}
-
-func writeJSON(path string, results []armResult) error {
-	if strings.TrimSpace(path) == "" {
-		return nil
-	}
-	b, err := marshalResults(results)
-	if err != nil {
-		return err
-	}
-	return os.WriteFile(path, b, 0o644)
-}
@@ -1,230 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"os"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/session/control"
-)
-
-// Where a stop goes — and, first, whose stop it is. A controller owns the turns
-// it admits itself and cancels those; a synchronous Run belongs to the caller
-// that passed the context. Measuring the second with the first says nothing
-// about either.
-const (
-	// armCancelTool is the control that says where a defect lives. If an
-	// ordinary tool's context is cancelled and a delegation's is not, the fault
-	// is in how a delegation derives its context; if neither is, it is upstream
-	// of both.
-	armCancelTool = "cancel-ordinary-tool"
-	// armCancelBackground is the negative control, and the one that decides what
-	// a fix may do: work already handed to a job has an owner of its own, and a
-	// cancel that reached into it would take back the handoff.
-	armCancelBackground = "cancel-background-handoff"
-	// armCancelHeadlessOwner reads the other surface. A synchronous Run is never
-	// admitted through the gate, so the controller holds no cancel for it and
-	// the caller's own context holds the only one. Both halves are asserted: a
-	// controller that started claiming this would be taking ownership from the
-	// only holder that has it.
-	armCancelHeadlessOwner = "cancel-headless-owner"
-)
-
-func cancelRoutingArm(name string) bool {
-	switch name {
-	case armCancelTool, armCancelBackground, armCancelHeadlessOwner:
-		return true
-	}
-	return false
-}
-
-// guardedArm names the arms driven through a turn the controller admits itself,
-// which is the surface a person's Stop reaches.
-func guardedArm(name string) bool {
-	return cancelTaskArm(name) || name == armCancelTool || name == armCancelBackground ||
-		name == armSkillCancel || name == armHostCancel
-}
-
-const (
-	sleepSentinel = "PROBE-SLEEP"
-	sleepCallID   = "probe_sleep"
-	// sleepSeconds outlasts the settling this probe waits through, so a shell
-	// that ends promptly ended because something ended it.
-	sleepSeconds = 30
-)
-
-// sleepCall is an ordinary foreground tool that does nothing but wait. It is
-// the shell every turn can call, which is the point: it derives its context the
-// way every other tool does.
-func sleepCall() []provider.Chunk {
-	args, _ := json.Marshal(map[string]any{
-		"command":     fmt.Sprintf("sleep %d", sleepSeconds),
-		"description": "waits until something stops it",
-	})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: sleepCallID, Name: "bash", Arguments: string(args)},
-	}}
-}
-
-// runCancelRoutingConstruct drives the arm's work, stops it the way that
-// surface's owner stops it, and records what the stop reached. Nothing waits
-// for the work to end: whether it ends at all is the measurement.
-func runCancelRoutingConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	sentinel := sleepSentinel
-	if arm == armCancelBackground {
-		sentinel = taskSentinel
-	}
-	prompt := fanOutTurn(turn+1, sentinel)
-	ctx, cancelCaller := context.WithCancel(ctx)
-	defer cancelCaller()
-	if guardedArm(arm) {
-		ctrl.Send(prompt)
-	} else {
-		go func() { _ = ctrl.Run(ctx, prompt) }()
-	}
-	if err := waitForCancelSubject(sink, prov, arm); err != nil {
-		return err
-	}
-
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	obs.Progress[probeStopReached] = stopReadings(ctrl, sink, prov, arm, cancelCaller)
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// probeStopReached carries what each stop reached into the observation. The key
-// names the probe so no row reads it as a status the kernel emitted.
-const probeStopReached = "probe:stop-reached"
-
-// stopReadings issues this surface's stop and reports what it reached. The
-// headless arm issues both in order — the controller's, which owns nothing
-// there, and then the caller's, which owns everything.
-func stopReadings(ctrl *control.Controller, sink *graphSink, prov *scripted, arm string, cancelCaller context.CancelFunc) []string {
-	// What the stop found before it was issued. Work already ended would make
-	// every answer below meaningless, and this probe has made that mistake once.
-	before := "live"
-	if !workLive(sink, prov, arm) {
-		before = "already gone"
-	}
-	ctrl.Cancel()
-	first := "controller-cancel before=" + before + " " + reachedAfter(ctrl, sink, prov, arm)
-	if arm != armCancelHeadlessOwner {
-		return []string{first}
-	}
-	cancelCaller()
-	return []string{first, "caller-cancel " + reachedAfter(ctrl, sink, prov, arm)}
-}
-
-// reachedAfter waits out the settling this probe allows and reports what is
-// still live. The wait stays well below the shell's own sleep, so work that has
-// ended was ended by the stop rather than by finishing.
-func reachedAfter(ctrl *control.Controller, sink *graphSink, prov *scripted, arm string) string {
-	stopped := time.Now()
-	deadline := stopped.Add(8 * time.Second)
-	for time.Now().Before(deadline) && workLive(sink, prov, arm) {
-		time.Sleep(100 * time.Millisecond)
-	}
-	live := "gone"
-	if workLive(sink, prov, arm) {
-		live = "still running"
-	}
-	return fmt.Sprintf("turn=%s work=%s after=%s",
-		turnStanding(ctrl, arm), live, time.Since(stopped).Round(time.Second))
-}
-
-func workLive(sink *graphSink, prov *scripted, arm string) bool {
-	if arm == armCancelBackground {
-		return prov.fleets.held(childEntered(childHang)) && !prov.fleets.held(childLeft(childHang))
-	}
-	_, ended := sink.toolResult(sleepCallID)
-	return !ended
-}
-
-// turnStanding reports the gate only where the gate is the authority. A
-// synchronous Run is never admitted through it, so reading it there would say
-// "ended" about a turn it was never told about — which is how this probe first
-// mistook an ownership boundary for a cancellation defect.
-func turnStanding(ctrl *control.Controller, arm string) string {
-	if !guardedArm(arm) {
-		return "not admitted (caller-owned)"
-	}
-	if ctrl.Running() {
-		return "still active"
-	}
-	return "ended"
-}
-
-// waitForCancelSubject blocks until the work this arm is about to stop is
-// really under way: a stop that arrived before it would measure nothing.
-func waitForCancelSubject(sink *graphSink, prov *scripted, arm string) error {
-	deadline := time.Now().Add(60 * time.Second)
-	for time.Now().Before(deadline) {
-		if arm == armCancelBackground {
-			if prov.fleets.held(childEntered(childHang)) {
-				return nil
-			}
-		} else if _, seen := sink.toolDispatched(sleepCallID); seen {
-			// The shell starts with the dispatch, not with the event; a moment's
-			// grace keeps the stop from racing its own subject.
-			time.Sleep(500 * time.Millisecond)
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return errUnexpected("work under way to stop", "nothing started")
-}
-
-// cancelRoutingRows read the arm's own readings. The three arms want different
-// answers to the same question, which is why they run together: a fix that
-// cancels everything a session knows about passes one and fails two.
-func cancelRoutingRows(arm string, before Observation) []row {
-	readings := before.Progress[probeStopReached]
-	first, second := reading(readings, 0), reading(readings, 1)
-	switch arm {
-	case armCancelHeadlessOwner:
-		return []row{
-			ownershipRow("a controller stop on a turn it never admitted", first,
-				"leaves the caller's work alone", stillRunning(first)),
-			ownershipRow("the caller's own stop", second,
-				"ends the work it owns", !stillRunning(second)),
-		}
-	case armCancelBackground:
-		return []row{
-			ownershipRow("the turn after a stop", first, "a cancelled turn ends",
-				strings.Contains(first, "turn=ended")),
-			ownershipRow("work already handed to a job", first,
-				"a stop aimed at the turn leaves work the turn no longer owns", stillRunning(first)),
-		}
-	}
-	return []row{
-		ownershipRow("the turn after a stop", first, "a cancelled turn ends",
-			strings.Contains(first, "turn=ended")),
-		ownershipRow("the tool the turn was waiting on", first,
-			"a stop reaches the work its turn is waiting on", !stillRunning(first)),
-	}
-}
-
-func ownershipRow(semantic, got, want string, ok bool) row {
-	return row{
-		Semantic: semantic, Authority: "the surface that owns this cancellation",
-		Artifact: "none: in-process state", Reconstruction: want,
-		Before: want, After: orNone(got), Verdict: held(ok),
-	}
-}
-
-func reading(readings []string, at int) string {
-	if at < len(readings) {
-		return readings[at]
-	}
-	return ""
-}
-
-func stillRunning(reading string) bool { return strings.Contains(reading, "work=still running") }
@@ -1,156 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"strings"
-)
-
-// How a host can stand after dying with a question open. The names are the
-// outcomes, not a pass and its failures: two of them are designs a host may
-// legitimately choose, and only measurement says which one this is.
-const (
-	outcomeResumable   = "resumable"            // the question is still there to answer
-	outcomeInterrupted = "interrupted-explicit" // it is gone, and the host says a turn was cut
-	outcomeLostSilent  = "LOST-SILENT"          // it is gone and nothing records that it existed
-)
-
-// decisionRows classify what a restart inherited. The safety row comes first
-// because it is the only one with a single acceptable answer: whether the
-// decision survives is a design question, whether the effect it blocked stayed
-// blocked is not.
-func decisionRows(before, after Observation) []row {
-	return []row{
-		deferredRow(before, after),
-		outcomeRow(before, after),
-		barrierRow(before, after),
-		modelContextRow(before, after),
-		obligationRow(before, after),
-	}
-}
-
-// deferredRow asserts the write the question held back never ran — not while
-// the answer was pending, and not because a later process loaded the session.
-func deferredRow(before, after Observation) row {
-	verdict := verdictHolds
-	if before.Deferred.Executed || after.Deferred.Executed {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "the effect the decision held back", Authority: "the workspace",
-		Artifact: before.Deferred.MarkerPath, Reconstruction: "the blocked write never runs, before or after the boundary",
-		Before: ranOrNot(before), After: ranOrNot(after), Verdict: verdict,
-	}
-}
-
-func ranOrNot(o Observation) string {
-	if o.Deferred.Executed {
-		return "RAN"
-	}
-	return "did not run"
-}
-
-// outcomeRow names which of the three standings the restart produced. It is
-// stated, not judged: a host that drops the question and records the
-// interruption is a defensible design, and one that drops both is not.
-func outcomeRow(before, after Observation) row {
-	verdict := verdictHolds
-	var outcome string
-	switch {
-	case len(after.Decisions) > 0:
-		outcome = outcomeResumable
-	case len(after.Interrupted) > 0:
-		outcome = outcomeInterrupted
-	case after.Obligation.InterruptionMarked > 0 || len(after.Obligation.UnansweredCalls) > 0:
-		outcome = outcomeInterrupted
-	default:
-		outcome = outcomeLostSilent
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "what the restart inherited", Authority: "approvalManager and the transcript",
-		Artifact:       "none for the question; the transcript for the obligation",
-		Reconstruction: "resumable, interrupted-explicit, or nothing at all",
-		Before:         decisionSummary(before), After: outcome, Verdict: verdict,
-	}
-}
-
-func decisionSummary(o Observation) string {
-	if len(o.Decisions) == 0 {
-		return "no decision open"
-	}
-	var kinds []string
-	for _, d := range o.Decisions {
-		kinds = append(kinds, string(d.Kind))
-	}
-	return strings.Join(kinds, ", ") + " open"
-}
-
-// barrierRow is the evidence behind an interrupted classification: a barrier
-// the host recorded before anyone could be asked, still open, with no owner in
-// this process waiting on it. It is deliberately not offered as answerable.
-func barrierRow(before, after Observation) row {
-	return row{
-		Semantic: "durable adjudication barriers", Authority: "the adjudication log",
-		Artifact: "<stem>.adjudication.jsonl", Reconstruction: "open records with no live owner in this process",
-		Before: barrierSummary(before), After: barrierSummary(after), Verdict: verdictStable,
-	}
-}
-
-func barrierSummary(o Observation) string {
-	if len(o.Interrupted) == 0 {
-		return "none"
-	}
-	var out []string
-	for _, b := range o.Interrupted {
-		out = append(out, b.Kind+":"+b.ID)
-	}
-	return join(out) + " (not answerable)"
-}
-
-// modelContextRow is the other read surface: what the next request tells the
-// model. One block when an interruption stands, none when it does not, and it
-// must never arrive as something answerable.
-func modelContextRow(before, after Observation) row {
-	verdict := verdictHolds
-	if before.ModelSeesInterruption != expectedBlocks(before) || after.ModelSeesInterruption != expectedBlocks(after) {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "what the next request tells the model", Authority: "control, projected per request",
-		Artifact: "none: derived, never stored", Reconstruction: "one block while an interruption stands, none otherwise",
-		Before: modelContextSummary(before), After: modelContextSummary(after), Verdict: verdict,
-	}
-}
-
-func expectedBlocks(o Observation) int {
-	if len(o.Interrupted) > 0 {
-		return 1
-	}
-	return 0
-}
-
-func modelContextSummary(o Observation) string {
-	return fmt.Sprintf("%d block(s) for %d interruption(s)", o.ModelSeesInterruption, len(o.Interrupted))
-}
-
-// obligationRow shows the evidence behind that classification, so a reader can
-// see what the host still says rather than trusting the label.
-func obligationRow(before, after Observation) row {
-	return row{
-		Semantic: "what the transcript still owes", Authority: "the canonical transcript",
-		Artifact: "<stem>.jsonl", Reconstruction: "tool calls with no result, and interruption records",
-		Before: obligationSummary(before), After: obligationSummary(after), Verdict: verdictStable,
-	}
-}
-
-func obligationSummary(o Observation) string {
-	return fmt.Sprintf("unanswered %s, interruption records %d",
-		orNone(join(o.Obligation.UnansweredCalls)), o.Obligation.InterruptionMarked)
-}
-
-func orNone(s string) string {
-	if s == "" {
-		return "none"
-	}
-	return s
-}
@@ -1,354 +0,0 @@
-package main
-
-import (
-	"context"
-	"os"
-	"reasonix/internal/state/sessionstore"
-	"sort"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/session/control"
-	"reasonix/internal/state/execgraph"
-)
-
-// The derivability arms. Everything before them asked what survives; these ask
-// what still has to be written down. A fact a restart can compute from what is
-// already durable needs no record of its own, and adding one would give it a
-// second owner for nothing.
-const (
-	// armDeriveSkipBoth cuts a branch with two upstreams that both end without
-	// an answer. Which of them the graph names is the question: the durable
-	// facts show two, and the picture shows one.
-	armDeriveSkipBoth = "derive-skip-both"
-	// armDeriveSkipFlip is the same graph with the two failures in the other
-	// order. If the named cause moves while the durable facts do not, the cause
-	// is history rather than a consequence of the end state.
-	armDeriveSkipFlip = "derive-skip-flip"
-	// armDeriveAnswered is the negative control and the context sample at once:
-	// two upstreams that both answer, one by completing and one by adopting, so
-	// the dependent must run and both deliveries must be accounted for.
-	armDeriveAnswered = "derive-answered"
-)
-
-func deriveArm(name string) bool {
-	switch name {
-	case armDeriveSkipBoth, armDeriveSkipFlip, armDeriveAnswered:
-		return true
-	}
-	return false
-}
-
-func deriveSentinels(arm string) string {
-	if arm == armDeriveAnswered {
-		return fleetWarmSentinel + " " + fleetDeriveSentinel
-	}
-	return fleetDeriveSentinel
-}
-
-// runDeriveConstruct lets the fan-out settle and dies before the turn ends, the
-// way every terminal arm does: a skip is published only in the closing delta.
-func runDeriveConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	runCtx, cancelRun := context.WithCancel(ctx)
-	defer cancelRun()
-	go func() { _ = ctrl.Run(runCtx, fanOutTurn(turn+1, deriveSentinels(arm))) }()
-	if err := waitForDerive(sink, arm, prov); err != nil {
-		return err
-	}
-	if err := writeObservation(root, capture("construct", arm, bootSystem, ctrl, sink, root)); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// waitForDerive holds until the arm's shape is on the graph. The skip arms
-// release their second failure once the first has been recorded, so which
-// upstream the fan-out names is decided rather than raced.
-func waitForDerive(sink *graphSink, arm string, prov *scripted) error {
-	if arm == armDeriveAnswered {
-		// Waiting on a count would be satisfied by the warm-up alone. The
-		// delivery edge is what the arm compares, so it waits for that.
-		return waitForShape(sink, func(g agentgraph.Graph) bool {
-			return len(contextTargets(g)) > 0 && len(nodesInState(g, agentgraph.StateAdopted)) > 0
-		}, "an adopted upstream and a delivered answer")
-	}
-	// The first failure decides which upstream is named — the fan-out cuts the
-	// branch when it processes that result. Only then is the second released,
-	// so the order is constructed rather than raced.
-	if err := waitForShape(sink, func(g agentgraph.Graph) bool {
-		return len(nodesInState(g, agentgraph.StateFailed)) > 0
-	}, "the first upstream to fail"); err != nil {
-		return err
-	}
-	prov.releaseOnce()
-	// A skip is published only in the closing delta, so the group has to end
-	// before the state the arm compares exists at all.
-	return waitForShape(sink, func(g agentgraph.Graph) bool {
-		return len(nodesInState(g, agentgraph.StateSkipped)) > 0 &&
-			len(nodesInState(g, agentgraph.StateFailed)) >= 2
-	}, "both upstreams failed and the dependent skipped")
-}
-
-// contextTargets are the items the graph shows receiving an upstream answer.
-func contextTargets(g agentgraph.Graph) []string {
-	var out []string
-	for _, e := range g.Edges {
-		if e.Kind == agentgraph.Context {
-			out = append(out, e.To)
-		}
-	}
-	return out
-}
-
-func waitForShape(sink *graphSink, ok func(agentgraph.Graph) bool, want string) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		graph, _ := sink.snapshot()
-		if ok(graph) {
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected(want, waitStates(graph))
-}
-
-// answeredFromDurable is Answered() computed the way a restart would have to:
-// an adoption says so on its own opening, and everything else is the terminal
-// the sub-agent store kept, joined by the execution id the child records.
-func answeredFromDurable(o Observation) map[string]bool {
-	terminal := map[string]string{}
-	for _, f := range o.Children.Facts {
-		if f.ParentToolCallID != "" {
-			terminal[f.ParentToolCallID] = f.Status
-		}
-	}
-	out := map[string]bool{}
-	for _, e := range o.Executions {
-		out[e.ID] = e.Disposition == "adopted" || terminal[e.ID] == string(sessionstore.SubagentCompleted)
-	}
-	return out
-}
-
-// impliedSkipped is the state a restart would derive: an item the orchestration
-// released without ever starting it, ordered behind something that did not
-// answer. Nothing here reads a disposition, because that is what is missing.
-func impliedSkipped(o Observation) string {
-	answered := answeredFromDurable(o)
-	var out []string
-	for _, e := range o.Executions {
-		if e.Started() || e.Queued() || e.SettledAt.IsZero() || len(e.DependsOn) == 0 {
-			continue
-		}
-		for _, up := range e.DependsOn {
-			if !answered[up] {
-				out = append(out, e.ID)
-				break
-			}
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// impliedSkipCauses are every upstream that could have been named, and the two
-// rules that would pick one: declaration order, and the earliest release. A
-// cause is derivable only if some rule always agrees with the picture.
-func impliedSkipCauses(o Observation) (all, byOrder, bySettle string) {
-	answered := answeredFromDurable(o)
-	byID := map[string]int{}
-	for i, e := range o.Executions {
-		byID[e.ID] = i
-	}
-	var every, first, earliest []string
-	for _, e := range o.Executions {
-		if e.Started() || e.Queued() || e.SettledAt.IsZero() || len(e.DependsOn) == 0 {
-			continue
-		}
-		var unanswered []string
-		for _, up := range e.DependsOn {
-			if !answered[up] {
-				unanswered = append(unanswered, up)
-			}
-		}
-		if len(unanswered) == 0 {
-			continue
-		}
-		every = append(every, e.ID+"<-{"+strings.Join(unanswered, ",")+"}")
-		first = append(first, e.ID+"<-"+unanswered[0])
-		pick := unanswered[0]
-		for _, up := range unanswered[1:] {
-			if i, j := byID[up], byID[pick]; o.Executions[i].SettledAt.Before(o.Executions[j].SettledAt) {
-				pick = up
-			}
-		}
-		earliest = append(earliest, e.ID+"<-"+pick)
-	}
-	sort.Strings(every)
-	sort.Strings(first)
-	sort.Strings(earliest)
-	return join(every), join(first), join(earliest)
-}
-
-// actualSkipCause is the upstream the fan-out actually named, read off the node
-// the graph settled. It is the picture's own answer, not a derivation.
-func actualSkipCause(o Observation) string {
-	var out []string
-	for _, n := range o.Graph.Nodes {
-		if n.State == agentgraph.StateSkipped && n.Err != "" {
-			out = append(out, n.ID+": "+n.Err)
-		}
-	}
-	sort.Strings(out)
-	return orNone(join(out))
-}
-
-// impliedContextFromDurable predicts the delivery edges from durable facts
-// alone: a started item receives an answer from every upstream that answered.
-func impliedContextFromDurable(o Observation) string {
-	answered := answeredFromDurable(o)
-	var out []string
-	for _, e := range o.Executions {
-		if !e.Started() {
-			continue
-		}
-		for _, up := range e.DependsOn {
-			if answered[up] {
-				out = append(out, e.ID+"<-"+up)
-			}
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// deriveRows are the three questions, kept apart. A state that derives says
-// nothing about whether the cause does, and one arm's agreement says nothing
-// about a rule that has only ever seen one sample.
-func deriveRows(arm string, before, after Observation) []row {
-	rows := []row{
-		joinRow(before, after),
-		skipStateRow(before, after),
-		contextDeriveRow(before, after),
-	}
-	if arm != armDeriveAnswered {
-		rows = append(rows, skipCauseRow(before, after))
-	}
-	return rows
-}
-
-// joinRow is what every derivation below stands on: the child records name the
-// execution they ran for, so a restart can ask the store what each item ended
-// as. Without it none of the other rows could be computed at all.
-func joinRow(before, after Observation) row {
-	return valueRow("executions the store can be asked about", "child metadata's parent tool-call id",
-		"subagents/<ref>.meta.json", "joined to the journal by execution id",
-		joinedExecutions(before), joinedExecutions(after), true)
-}
-
-func joinedExecutions(o Observation) string {
-	var out []string
-	for _, f := range o.Children.Facts {
-		if f.ParentToolCallID != "" {
-			out = append(out, f.ParentToolCallID+"="+f.Status)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// skipStateRow compares the state a restart would derive against the one the
-// picture drew. Agreement means StateSkipped needs no record of its own.
-func skipStateRow(before, after Observation) row {
-	actual := join(nodesInState(graphOf(before), agentgraph.StateSkipped))
-	implied := impliedSkipped(after)
-	verdict := verdictHolds
-	if implied != actual {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "skipped state: derived against drawn", Authority: "the journal joined to the store",
-		Artifact:       "<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		Reconstruction: "settled, never started, ordered behind something that did not answer",
-		Before:         "drawn " + orNone(actual), After: "derived " + orNone(implied), Verdict: verdict,
-	}
-}
-
-// skipCauseRow keeps the harder question separate, and reads it three ways: the
-// upstream the picture named, the two rules this file works out on its own, and
-// what the production fold concludes. The independence matters — a probe that
-// only asked the implementation would agree with it by construction.
-func skipCauseRow(before, after Observation) row {
-	all, byOrder, bySettle := impliedSkipCauses(after)
-	return row{
-		Semantic:       "skipped cause: what the picture named, what the facts permit",
-		Authority:      "the fan-out's own choice against the durable upstream set",
-		Artifact:       "none: the reason text is not recorded",
-		Reconstruction: "declaration order, or earliest released, over the upstreams that did not answer",
-		Before:         actualSkipCause(before),
-		After: "candidates " + orNone(all) + " | by-order " + orNone(byOrder) +
-			" | by-settle " + orNone(bySettle) + " | rebuild " + orNone(rebuiltSkipCauses(after)),
-		Verdict: verdictStable,
-	}
-}
-
-// rebuiltSkipCauses is what the production fold names, asked the same question
-// through its own code path.
-func rebuiltSkipCauses(o Observation) string {
-	children := make([]execgraph.ChildOutcome, 0, len(o.Children.Facts))
-	for _, f := range o.Children.Facts {
-		children = append(children, execgraph.ChildOutcome{
-			Execution: f.ParentToolCallID, Ref: f.Ref, Status: f.Status,
-		})
-	}
-	rebuilt := execgraph.Rebuild(o.Executions, children, nil)
-	var out []string
-	for _, n := range rebuilt.Graph.Nodes {
-		if n.State != agentgraph.StateSkipped {
-			continue
-		}
-		if cause := execgraph.SkipCause(o.Executions, children, n.ID); cause != "" {
-			out = append(out, n.ID+"<-"+cause)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// contextDeriveRow asks the same question of the delivery edge, now that a
-// restart can tell which upstreams answered.
-func contextDeriveRow(before, after Observation) row {
-	actual, implied := contextEdges(before), impliedContextFromDurable(after)
-	verdict := verdictHolds
-	if implied != actual {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "context edges: derived against drawn", Authority: "the journal joined to the store",
-		Artifact:       "<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		Reconstruction: "a started item receives from every upstream that answered",
-		Before:         "drawn " + orNone(actual), After: "derived " + orNone(implied), Verdict: verdict,
-	}
-}
-
-// deriveArmInvalid refuses an arm that did not reach the shape it compares.
-func deriveArmInvalid(arm string, before Observation) string {
-	if arm == armDeriveAnswered {
-		if len(nodesInState(graphOf(before), agentgraph.StateAdopted)) == 0 {
-			return "no item was adopted, so the arm never had an answered upstream that did not run"
-		}
-		if contextEdges(before) == "" {
-			return "no answer was delivered, so there is no context edge to derive"
-		}
-		return ""
-	}
-	if len(nodesInState(graphOf(before), agentgraph.StateSkipped)) == 0 {
-		return "no dependent was skipped; the fan-out ended as " + orNone(terminalSummary(before))
-	}
-	if len(nodesInState(graphOf(before), agentgraph.StateFailed)) < 2 {
-		return "only one upstream ended without an answer, so the cause was never ambiguous"
-	}
-	return ""
-}
@@ -1,239 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"os"
-	"sort"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/session/control"
-)
-
-// The two ephemeral entry points, read against the execution read model's own
-// authority: the snapshot is what a reader starts from and what a gap restores,
-// and the stream is a projection of the same facts. A node only the stream has
-// is therefore not a lighter node but one the authority cannot account for, and
-// these arms ask whether either entry point produces one.
-const (
-	armEphemeralTask  = "ephemeral-task-authority"
-	armEphemeralSkill = "ephemeral-skill-authority"
-)
-
-func ephemeralArm(name string) bool {
-	return name == armEphemeralTask || name == armEphemeralSkill
-}
-
-const (
-	ephemeralSentinel = "PROBE-EPHEMERAL"
-	// ephemeralCallID is the call the model makes. Both entry points use it, so
-	// the two arms differ by which tool ran and by nothing else.
-	ephemeralCallID = "probe_ephemeral"
-)
-
-// ephemeralCall dispatches the entry point this arm measures. Its child hangs,
-// because every row here is about a run that is still happening: an ephemeral
-// execution that has ended owes no reader anything.
-func ephemeralCall(arm string) []provider.Chunk {
-	name, args := "read_only_task", map[string]any{
-		"prompt": childHang + " ephemeral research", "description": "still executing at death",
-	}
-	if arm == armEphemeralSkill {
-		name, args = "read_only_skill", map[string]any{
-			"name": probeSkillName, "arguments": childHang + " ephemeral research",
-		}
-	}
-	raw, _ := json.Marshal(args)
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: ephemeralCallID, Name: name, Arguments: string(raw)},
-	}}
-}
-
-// runEphemeralConstruct drives one ephemeral run and reads three surfaces at the
-// same instant, while its child is inside the provider: what the stream drew,
-// what the authority holds, and what a reader is left with once a gap makes it
-// start from the authority again.
-func runEphemeralConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	go func() { _ = ctrl.Run(ctx, fanOutTurn(turn+1, ephemeralSentinel)) }()
-	if err := waitForEphemeral(prov); err != nil {
-		return err
-	}
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	obs.Progress[probeChildEntered] = []string{fmt.Sprint(prov.fleets.held(childEntered(childHang)))}
-	// One instant, three readings. Taken here because two of them cannot be
-	// taken later at all: the authority is recomputed per call, and what a gap
-	// leaves is a claim about the moment the gap happens.
-	live, _ := sink.snapshot()
-	authority := ctrl.ExecutionGraph()
-	obs.Progress[probeEphemeralLive] = []string{ephemeralNodes(live)}
-	obs.Progress[probeEphemeralAuthority] = []string{ephemeralNodes(authority.Graph)}
-	obs.Progress[probeEphemeralAfterGap] = []string{ephemeralNodes(afterGap(authority.Graph))}
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// The three readings one instant produces. They ride the observation because
-// none of them survives the process, and the arm's whole question is what a
-// reader could see while the run was happening.
-const (
-	probeEphemeralLive      = "probe:ephemeral-live"
-	probeEphemeralAuthority = "probe:ephemeral-authority"
-	probeEphemeralAfterGap  = "probe:ephemeral-after-gap"
-)
-
-// afterGap is what a view holds once a dropped stream makes it start from the
-// authority again: the snapshot, and only the snapshot. A resync replaces the
-// folded state rather than merging into it — that is what makes the snapshot the
-// authority — so whatever the fold held and it does not has left the screen.
-func afterGap(authority agentgraph.Graph) agentgraph.Graph { return authority }
-
-func waitForEphemeral(prov *scripted) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		if prov.fleets.held(childEntered(childHang)) {
-			// Let the picture settle: a reading taken the instant the child
-			// arrives could miss a delta that was already on its way.
-			time.Sleep(time.Second)
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return errUnexpected("an ephemeral child inside the provider", "it never reached one")
-}
-
-// ephemeralNodes renders the nodes belonging to this arm's run, worker and group
-// alike: a group card drawn for work the authority cannot account for is the
-// same defect as a worker one.
-func ephemeralNodes(g agentgraph.Graph) string {
-	var out []string
-	for _, n := range g.Nodes {
-		if strings.HasPrefix(n.ID, ephemeralCallID) {
-			out = append(out, n.ID+":"+orDash(string(n.State)))
-		}
-	}
-	sort.Strings(out)
-	return orNone(join(out))
-}
-
-func ephemeralRows(arm string, before, after Observation) []row {
-	return []row{
-		ephemeralLiveRow(before),
-		ephemeralAuthorityRow(before),
-		ephemeralGapRow(before),
-		ephemeralRestartRow(before, after),
-		ephemeralDurableRow(before, after),
-		ephemeralContractRow(arm, before),
-	}
-}
-
-const authorityName = "Controller.ExecutionGraph (the snapshot a reader starts from)"
-
-func ephemeralLiveRow(before Observation) row {
-	drawn := firstProgress(before, probeEphemeralLive)
-	return row{
-		Semantic: "what the stream drew while it ran", Authority: liveAuthority,
-		Artifact: "none observed", Reconstruction: "in-process only",
-		Before: drawn, After: "—", Verdict: presenceVerdict(drawn != "none"),
-	}
-}
-
-func ephemeralAuthorityRow(before Observation) row {
-	held := firstProgress(before, probeEphemeralAuthority)
-	return row{
-		Semantic: "what the authority held at the same instant", Authority: authorityName,
-		Artifact:       "<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		Reconstruction: "execgraph.Rebuild over durable facts, recomputed on demand",
-		Before:         held, After: "—", Verdict: presenceVerdict(held != "none"),
-	}
-}
-
-// ephemeralGapRow is the one that decides. A reader that loses the stream starts
-// from the authority again, so whatever the fold held and the authority does not
-// leaves the screen — while the child it named is still holding a slot.
-func ephemeralGapRow(before Observation) row {
-	left := firstProgress(before, probeEphemeralAfterGap)
-	return row{
-		Semantic: "what a reader is left with after a gap", Authority: authorityName,
-		Artifact:       "none: the fold is replaced, never merged",
-		Reconstruction: "a resync starts from the authority",
-		Before:         left, After: "—", Verdict: presenceVerdict(left != "none"),
-	}
-}
-
-func ephemeralRestartRow(before, after Observation) row {
-	return row{
-		Semantic: "what a restart can say about it", Authority: rebuildAuthority,
-		Artifact:       "<stem>.execution.jsonl",
-		Reconstruction: "execgraph.Rebuild in the next process",
-		Before:         "executing inside the provider",
-		After:          ephemeralNodes(rebuiltGraph(after).Graph),
-		Verdict:        presenceVerdict(ephemeralNodes(rebuiltGraph(after).Graph) != "none"),
-	}
-}
-
-// ephemeralDurableRow is the contract both entry points already promise. It is
-// read rather than assumed, because a convergence that made one of them visible
-// by recording it would satisfy the rows above and break this one.
-func ephemeralDurableRow(before, after Observation) row {
-	return presenceRow("journal and store", crossAuthority,
-		"<stem>.execution.jsonl, subagents/<ref>.meta.json",
-		"what an ephemeral entry point promises to leave: nothing",
-		ephemeralDurable(before), ephemeralDurable(after))
-}
-
-func ephemeralDurable(o Observation) []string {
-	var out []string
-	for _, e := range o.Executions {
-		if strings.HasPrefix(e.ID, ephemeralCallID) {
-			out = append(out, "journal:"+e.ID)
-		}
-	}
-	for _, f := range o.Children.Facts {
-		if strings.HasPrefix(f.ParentToolCallID, ephemeralCallID) {
-			out = append(out, "store:"+f.ParentToolCallID+"="+f.Status)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// ephemeralContractRow states the authority contract as a conjunction. A node
-// the stream drew, an authority that cannot account for it, and a child still
-// executing: that is not a lighter kind of node, it is one that leaves the
-// screen at the next gap while the work it named goes on holding a slot.
-func ephemeralContractRow(arm string, before Observation) row {
-	live := firstProgress(before, probeEphemeralLive) != "none"
-	authority := firstProgress(before, probeEphemeralAuthority) != "none"
-	running := firstProgress(before, probeChildEntered) == "true"
-	answer, verdict := "no node was drawn: nothing claims what the authority cannot hold", verdictHolds
-	switch {
-	case !running:
-		answer, verdict = "the child never reached the provider", verdictNotMeasured
-	case live && !authority:
-		answer, verdict = "drawn by the stream, absent from the authority, and still executing", verdictViolated
-	case live && authority:
-		answer, verdict = "drawn, and the authority accounts for it", verdictHolds
-	}
-	return row{
-		Semantic: "a node the authority cannot account for", Authority: crossAuthority,
-		Artifact:       "none, which is the point",
-		Reconstruction: "the stream and the snapshot, read at one instant",
-		Before:         answer, After: "—", Verdict: verdict,
-	}
-}
-
-// ephemeralArmInvalid reports the premise this arm could not establish.
-func ephemeralArmInvalid(before Observation) string {
-	if firstProgress(before, probeChildEntered) != "true" {
-		return "the ephemeral child never reached the provider, so nothing was executing at the reading"
-	}
-	return ""
-}
@@ -1,641 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"os"
-	"regexp"
-	"slices"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/session/control"
-)
-
-// The three fan-out arms. Each dies at a different point in one fleet's life,
-// because "what survives" has no single answer: an item that finished, one
-// still executing, and one the graph knows about without ever running are
-// carried by different host state, and only separating them says which.
-const (
-	// armGraphCompleted lets the fleet finish and the turn close before the
-	// process exits. Nothing is interrupted, so a loss here is not about
-	// interruption at all — it says execution provenance itself is ephemeral.
-	armGraphCompleted = "graph-completed"
-	// armGraphRunning kills the process with a child mid-execution. It asks the
-	// one question a completed fleet cannot: what does the new process believe
-	// about work whose owner is gone?
-	armGraphRunning = "graph-running"
-	// armGraphMixed carries every state at once — completed, running, adopted,
-	// failed, skipped — plus a read grant, a write grant, and a dependency edge.
-	// It is the density arm: one death, every semantic the graph claims to hold.
-	armGraphMixed = "graph-mixed"
-)
-
-func graphArm(name string) bool {
-	return name == armGraphCompleted || name == armGraphRunning || name == armGraphMixed
-}
-
-// Child sentinels. A child's prompt decides what its run does, and the probe
-// reads the prompt off the request's last user message rather than scanning the
-// whole conversation: the parent's own transcript quotes every fleet argument,
-// so a scan would make the parent answer as its own children.
-const (
-	childDone = "PROBE-CHILD-DONE"
-	childHang = "PROBE-CHILD-HANG"
-	childFail = "PROBE-CHILD-FAIL"
-)
-
-// Parent sentinels, one per fleet the arm dispatches.
-const (
-	fleetWarmSentinel  = "PROBE-FLEET-WARM"
-	fleetMixedSentinel = "PROBE-FLEET-MIXED"
-	fleetPairSentinel  = "PROBE-FLEET-PAIR"
-	// The scheduler arms. Holder fills a ceiling and never ends; refused asks
-	// for admission the scheduler must deny. Two sentinels rather than one
-	// because the claim and transition arms need the holder to be already
-	// admitted before the second fleet is dispatched.
-	fleetHolderSentinel  = "PROBE-FLEET-HOLDER"
-	fleetRefusedSentinel = "PROBE-FLEET-REFUSED"
-	// The terminal arms. Their fan-outs run to completion, because a skip is
-	// published only in the closing delta; the turn is then held open so the
-	// process still dies mid-turn.
-	fleetTerminalSentinel = "PROBE-FLEET-TERMINAL"
-	// Named so it is not contained in PROBE-PARALLEL-IDENTITY. Sentinels are
-	// matched by substring, so one inside another dispatches a fan-out its arm
-	// never asked for — this pair silently ran cancelParallel in the identity
-	// arm, whose rows then read a scenario nobody wrote.
-	parallelSentinel      = "PROBE-PARALLEL-CANCEL"
-	fleetOutcomesSentinel = "PROBE-FLEET-OUTCOMES"
-	fleetDeriveSentinel   = "PROBE-FLEET-DERIVE"
-	fleetIdentitySentinel = "PROBE-FLEET-IDENTITY"
-	parallelIdentity      = "PROBE-PARALLEL-IDENTITY"
-	// The active-store arm. Its fan-out shares one ceiling with a lone
-	// delegation, so the four populations it needs are reached in one death.
-	fleetActiveSentinel = "PROBE-FLEET-ACTIVE"
-)
-
-// childHold is a child that reports holding its slot before it blocks, so a
-// later fleet can be dispatched knowing the ceiling is already occupied. Timing
-// it any other way races the scheduler.
-const (
-	childHold = "PROBE-CHILD-HOLD"
-	// childRelease finishes only when the arm frees it, which is how capacity
-	// is returned while a refusal is still standing.
-	childRelease = "PROBE-CHILD-RELEASE"
-	// childFailLate fails only once the arm frees it, so two failures in one
-	// fan-out have a decided order instead of a raced one. Its wording shares no
-	// prefix with childFail: the sentinels are matched by substring, so a name
-	// containing another's makes the shorter one win and the delay disappear.
-	childFailLate = "PROBE-CHILD-DELAYEDFAIL"
-)
-
-// probeClaimPath is the write path the claim arm overlaps. The refused writer
-// asks for a path inside it, which is a conflict no ceiling change releases.
-const (
-	probeClaimPath  = "probe-claim"
-	probeClaimInner = "probe-claim/inner"
-)
-
-// subagentRefPattern finds the references an earlier fleet's aggregate reported.
-// The mixed arm adopts one of them, which is the only way to reach StateAdopted
-// through the tool's own contract rather than by constructing the state.
-var subagentRefPattern = regexp.MustCompile(`sa_[A-Za-z0-9_-]+`)
-
-// childScript answers one delegated run. Hanging is expressed as a context read
-// so the child holds its scheduler slot until the process dies — a sleep long
-// enough to look the same would still be a race against the probe's deadline.
-func (s *scripted) childScript(ctx context.Context, sentinel string) []provider.Chunk {
-	// Reaching here is the child's own proof that it started, and returning is
-	// its proof that something let it go: an arm asking whether a cancellation
-	// reached a running child has no other way to know.
-	s.fleets.first(childEntered(sentinel))
-	defer s.fleets.first(childLeft(sentinel))
-	switch sentinel {
-	case childHold:
-		// Reaching a provider call means the slot is already held, which is the
-		// only moment a later fleet can be dispatched against a full ceiling
-		// without racing it.
-		s.holding.Do(func() { close(s.held) })
-		<-ctx.Done()
-		return nil
-	case childRelease:
-		<-s.release
-		return append(text("Child released."), done())
-	case childFailLate:
-		<-s.release
-		return []provider.Chunk{{Type: provider.ChunkError, Err: fmt.Errorf("probe child failed on release")}}
-	case childHang:
-		<-ctx.Done()
-		// The context closed, and this is the only place that can say so: a
-		// caller that stopped waiting leaves the same absence behind.
-		s.fleets.first(childCtxDone(sentinel))
-		return nil
-	case childFail:
-		return []provider.Chunk{{Type: provider.ChunkError, Err: fmt.Errorf("probe child failed on purpose")}}
-	default:
-		return append(text("Child answered."), done())
-	}
-}
-
-// askedInPrompt reports whether this request's own user turns carry a sentinel.
-// The user side only: a fleet's arguments live in an assistant message, so a
-// whole-conversation scan makes a parent answer as one of its own children.
-// Every user turn, not the last: the host appends its own notes after the
-// prompt, and the last user message in a request is frequently one of those.
-func askedInPrompt(req provider.Request, sentinel string) bool {
-	for _, m := range req.Messages {
-		if m.Role == provider.RoleUser && strings.Contains(m.Content, sentinel) {
-			return true
-		}
-	}
-	return false
-}
-
-// childSentinel reports which child script this request is, empty when the
-// request is a parent turn.
-func childSentinel(req provider.Request) string {
-	for _, sentinel := range []string{childDone, childHang, childFail, childHold, childRelease, childFailLate} {
-		if askedInPrompt(req, sentinel) {
-			return sentinel
-		}
-	}
-	return ""
-}
-
-// fanOutTurnRecorded reports whether the dispatching turn is in these messages.
-// The probe reads its own sentinel rather than counting rows: a turn appended
-// at some later point would still be there, and only the prompt says whether
-// this is the turn the fan-out was opened in.
-func fanOutTurnRecorded(msgs []provider.Message) bool {
-	for _, m := range msgs {
-		if strings.Contains(m.Content, fleetPairSentinel) || strings.Contains(m.Content, fleetMixedSentinel) {
-			return true
-		}
-	}
-	return false
-}
-
-// fleetCall renders one fleet dispatch. Nothing closes the round it opens: all
-// three arms die inside the turn, so the host's final-answer readiness check
-// never runs and the arms measure a process boundary rather than a task list.
-func fleetCall(id string, tasks []map[string]any) []provider.Chunk {
-	args, _ := json.Marshal(map[string]any{"tasks": tasks})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: id, Name: "fleet", Arguments: string(args)},
-	}}
-}
-
-// warmFleet finishes two children so a later fleet has a completed reference to
-// adopt. Both are read-only: nothing about the warm-up is under measurement.
-func warmFleet() []provider.Chunk {
-	return fleetCall("probe_fleet_warm", []map[string]any{
-		{"id": "w1", "prompt": childDone + " warm one", "description": "warm one", "read_only": true},
-		{"id": "w2", "prompt": childDone + " warm two", "description": "warm two", "read_only": true},
-	})
-}
-
-// pairFleet is the two-item shape both single-state arms use. The running arm
-// dies with the second item still executing; the completed arm lets both
-// finish, so the two differ by when the process exits and nothing else.
-func pairFleet(second string) []provider.Chunk {
-	return fleetCall("probe_fleet_pair", []map[string]any{
-		{"id": "p1", "prompt": childDone + " pair one", "description": "pair one", "read_only": true},
-		{"id": "p2", "prompt": second + " pair two", "description": "pair two", "read_only": true},
-	})
-}
-
-// mixedFleet is the density arm's dispatch: one item per terminal the graph can
-// record, plus one still running, and the two grants told apart. The write item
-// declares a path so preflight admits it beside the readers; it never writes,
-// because the grant is the envelope under measurement, not the effect.
-func mixedFleet(adoptRef string) []provider.Chunk {
-	return fleetCall("probe_fleet_mixed", []map[string]any{
-		// Model and effort are named so the identity row measures something: two
-		// nodes agreeing that a field is empty proves nothing about recovering it.
-		{"id": "m1", "prompt": childDone + " mixed one", "description": "completes", "read_only": true,
-			"model": probeModelRef, "effort": "high"},
-		{"id": "m2", "prompt": childFail + " mixed two", "description": "fails", "read_only": true},
-		{"id": "m3", "prompt": childDone + " mixed three", "description": "skipped by m2", "read_only": true, "depends_on": []string{"m2"}},
-		{"id": "m4", "adopt_ref": adoptRef, "description": "adopted"},
-		{"id": "m5", "prompt": childHang + " mixed five", "description": "running at death", "write_paths": []string{"probe-fanout-out"}},
-	})
-}
-
-// opens reports whether this round is the one that dispatches that fleet.
-func (s *scripted) opens(req provider.Request, sentinel string) bool {
-	return askedInPrompt(req, sentinel) && s.fleets.first(sentinel)
-}
-
-// opensRun is the same for the fan-out an arm is built around: the terminal and
-// derive arms hold the turn open once theirs has gone out.
-func (s *scripted) opensRun(req provider.Request, sentinel string) bool {
-	if !s.opens(req, sentinel) {
-		return false
-	}
-	s.opened.Store(true)
-	return true
-}
-
-// fanOut answers a parent turn that opens a fleet, at most once per sentinel.
-// The sentinel is read off the current user turn, and the latch is what stops
-// the round that receives the aggregate from dispatching the same fleet again.
-func (s *scripted) fanOut(req provider.Request) ([]provider.Chunk, bool) {
-	switch {
-	case s.opens(req, fleetWarmSentinel):
-		return append(warmFleet(), done()), true
-	case s.opens(req, fleetPairSentinel):
-		if s.arm == armWaitSlots {
-			return append(slotsFleet(), done()), true
-		}
-		second := childDone
-		if s.arm == armGraphRunning {
-			second = childHang
-		}
-		return append(pairFleet(second), done()), true
-	case s.opens(req, fleetSettledSentinel):
-		return append(settledFleet(), done()), true
-	case s.opens(req, fleetLiveSentinel):
-		return append(liveFleet(), done()), true
-	case s.opens(req, fleetMixedSentinel):
-		if s.arm == armTerminalAdopted {
-			return append(adoptOnlyFleet(adoptableRef(req)), done()), true
-		}
-		if s.arm == armUIGraphMixed {
-			return append(uiMixedFleet(adoptableRef(req)), done()), true
-		}
-		return append(mixedFleet(adoptableRef(req)), done()), true
-	case s.opensRun(req, fleetIdentitySentinel):
-		return append(identityFleet(adoptableRef(req)), done()), true
-	case s.opens(req, parallelIdentity):
-		return append(identityParallel(), done()), true
-	case s.opensRun(req, fleetDeriveSentinel):
-		return append(deriveFleet(s.arm, adoptableRef(req)), done()), true
-	case s.opensRun(req, fleetOutcomesSentinel):
-		return append(outcomesFleet(), done()), true
-	case s.opensRun(req, fleetTerminalSentinel):
-		return append(terminalFleet(s.arm), done()), true
-	case s.opensRun(req, parallelSentinel):
-		return append(cancelParallel(), done()), true
-	case s.fleets.held(taskSentinel) && s.opens(req, fleetActiveSentinel):
-		// The delegation goes first and its own child says when the slot is
-		// taken. Dispatching before that lets the fan-out take the whole
-		// ceiling, and the arm loses the entry point it exists to compare.
-		<-s.held
-		return append(activeStoreFleet(), done()), true
-	case s.opens(req, fleetHolderSentinel):
-		return append(holderFleet(s.arm), done()), true
-	case s.opens(req, fleetRefusedSentinel):
-		// The holder's own child says when its ceiling is occupied. Dispatching
-		// before that lets the refused item win the race and start.
-		<-s.held
-		return append(refusedFleet(s.arm), done()), true
-	}
-	return nil, false
-}
-
-// holderFleet occupies the ceiling this arm measures. The slots arm needs no
-// holder — its refused item is in the same fleet as the runs that fill the
-// ceiling — so only the arms that need one dispatch it in the background.
-func holderFleet(arm string) []provider.Chunk {
-	tasks := []map[string]any{
-		{"id": "h1", "prompt": childHold + " holder", "description": "holds the ceiling", "write_paths": []string{probeClaimPath}},
-		{"id": "h2", "prompt": childHang + " filler", "description": "filler", "read_only": true},
-	}
-	if arm == armWaitTransition {
-		// The second run is what the arm later releases: freeing total capacity
-		// while the writer ceiling stays full is how the blocker changes.
-		tasks[1] = map[string]any{"id": "h2", "prompt": childRelease + " releasable", "description": "released mid-wait", "read_only": true}
-	}
-	if bothHoldersReport(arm) {
-		tasks[1] = map[string]any{"id": "h2", "prompt": childHold + " second holder", "description": "holds the ceiling", "read_only": true}
-	}
-	return backgroundFleetCall("probe_fleet_holder", tasks)
-}
-
-// bothHoldersReport names the arms that leave one slot free. Either holder may
-// win it, so both report holding, and the refused delegation's own child is then
-// the only arrival that means anything — a filler that hung under the same
-// sentinel would be read as the subject having started.
-func bothHoldersReport(arm string) bool {
-	return queuedTaskArm(arm) || queuedSkillArm(arm) || queuedHostArm(arm)
-}
-
-// refusedFleet asks for admission the scheduler must deny, with every check
-// above this arm's cause already cleared.
-func refusedFleet(arm string) []provider.Chunk {
-	refused := map[string]any{"id": "r1", "prompt": childHang + " refused", "description": "refused admission"}
-	switch arm {
-	case armWaitWriters:
-		// Disjoint from the holder's path, so only the writer ceiling can
-		// refuse it.
-		refused["write_paths"] = []string{"probe-writers-own"}
-	case armWaitClaim:
-		refused["write_paths"] = []string{probeClaimInner}
-	default:
-		refused["write_paths"] = []string{"probe-transition-own"}
-	}
-	return fleetCall("probe_fleet_refused", []map[string]any{
-		refused,
-		{"id": "r2", "prompt": childHang + " sibling", "description": "sibling", "read_only": true},
-	})
-}
-
-// terminalFleet runs to completion holding the disposition its arm measures.
-// The skipped arm cuts a branch with a failure; the context arm delivers a real
-// answer across an ordering edge, which is the only way a context edge appears.
-func terminalFleet(arm string) []provider.Chunk {
-	upstream := map[string]any{"id": "up", "prompt": childDone + " upstream", "description": "upstream", "read_only": true}
-	if arm == armTerminalSkippedDep {
-		upstream["prompt"] = childFail + " upstream"
-	}
-	return fleetCall("probe_fleet_terminal", []map[string]any{
-		upstream,
-		{"id": "down", "prompt": childDone + " downstream", "description": "downstream", "read_only": true, "depends_on": []string{"up"}},
-	})
-}
-
-// cancelParallel opens more items than the session admits at once, so the arm
-// can interrupt the group while one has started and another has not. That split
-// is what parallel_tasks reports as cancelled against skipped.
-func cancelParallel() []provider.Chunk {
-	args, _ := json.Marshal(map[string]any{"tasks": []map[string]any{
-		{"prompt": childHang + " first", "description": "started"},
-		{"prompt": childHang + " second", "description": "never started"},
-		{"prompt": childHang + " third", "description": "never started"},
-	}})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: "probe_parallel", Name: "parallel_tasks", Arguments: string(args)},
-	}}
-}
-
-// adoptOnlyFleet runs to completion with one item adopted and one that ran.
-// Nothing hangs: the arm compares two settled items whose only difference is
-// that one executed, which a fan-out still in flight could not show.
-func adoptOnlyFleet(adoptRef string) []provider.Chunk {
-	return fleetCall("probe_fleet_terminal", []map[string]any{
-		{"id": "a1", "adopt_ref": adoptRef, "description": "adopted"},
-		{"id": "a2", "prompt": childDone + " ran", "description": "ran", "read_only": true},
-	})
-}
-
-// outcomesFleet reaches all three executed terminals in one group: one item
-// completes, one fails, and two are still running when the arm interrupts. The
-// store's record of each is then compared under identical conditions.
-func outcomesFleet() []provider.Chunk {
-	return fleetCall("probe_fleet_outcomes", []map[string]any{
-		{"id": "o1", "prompt": childDone + " completes", "description": "completes", "read_only": true},
-		{"id": "o2", "prompt": childFail + " fails", "description": "fails", "read_only": true},
-		{"id": "o3", "prompt": childHang + " cancelled", "description": "cancelled", "read_only": true},
-		{"id": "o4", "prompt": childHang + " cancelled too", "description": "cancelled too", "read_only": true},
-	})
-}
-
-// identityFleet names a worker identity four ways: not at all, by model, by
-// effort, and not at all because nothing runs. What the graph then shows for
-// each is what says which layer of the resolution it is reporting.
-func identityFleet(adoptRef string) []provider.Chunk {
-	return fleetCall("probe_fleet_identity", []map[string]any{
-		{"id": "i1", "prompt": childDone + " default", "description": "inherits", "read_only": true},
-		{"id": "i2", "prompt": childDone + " override", "description": "names a model", "read_only": true,
-			"model": probeAltModelRef, "effort": "high"},
-		{"id": "i3", "prompt": childDone + " effort only", "description": "names an effort", "read_only": true,
-			"effort": "low"},
-		{"id": "i4", "adopt_ref": adoptRef, "description": "runs nothing"},
-	})
-}
-
-// identityParallel asks the same of the other producer, which resolves through
-// a shorter chain: it has no profile layer to consult.
-func identityParallel() []provider.Chunk {
-	args, _ := json.Marshal(map[string]any{"tasks": []map[string]any{
-		{"prompt": childDone + " p default", "description": "inherits"},
-		{"prompt": childDone + " p override", "description": "names a model",
-			"model": probeAltModelRef, "effort": "high"},
-	}})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: "probe_parallel_identity", Name: "parallel_tasks", Arguments: string(args)},
-	}}
-}
-
-// deriveFleet orders one dependent behind two upstreams. The skip arms end both
-// upstreams without an answer, one immediately and one on release, so which of
-// them the fan-out names is decided by the order rather than by a race; the
-// answered arm gives the dependent one completed and one adopted upstream, so
-// it must run and both deliveries must be accounted for.
-func deriveFleet(arm, adoptRef string) []provider.Chunk {
-	first, second := childFail+" first", childFailLate+" second"
-	if arm == armDeriveSkipFlip {
-		first, second = childFailLate+" first", childFail+" second"
-	}
-	up := []map[string]any{
-		{"id": "a", "prompt": first, "description": "upstream a", "read_only": true},
-		{"id": "b", "prompt": second, "description": "upstream b", "read_only": true},
-	}
-	if arm == armDeriveAnswered {
-		up = []map[string]any{
-			{"id": "a", "prompt": childDone + " first", "description": "upstream a", "read_only": true},
-			{"id": "b", "adopt_ref": adoptRef, "description": "upstream b"},
-		}
-	}
-	return fleetCall("probe_fleet_derive", append(up, map[string]any{
-		"id": "c", "prompt": childDone + " dependent", "description": "dependent",
-		"read_only": true, "depends_on": []string{"a", "b"},
-	}))
-}
-
-// activeStoreFleet reaches three fan-out populations under one ceiling. The
-// first item settles and frees the slot the other two then contend for, so one
-// executes and one is refused — which of them is the scheduler's to decide, and
-// the arm reads the populations off the journal rather than naming them here.
-func activeStoreFleet() []provider.Chunk {
-	return fleetCall(activeFleetCallID, []map[string]any{
-		{"id": "a1", "prompt": childDone + " settles", "description": "settles", "read_only": true},
-		{"id": "a2", "prompt": childHang + " holds", "description": "contends for the freed slot",
-			"read_only": true, "depends_on": []string{"a1"}},
-		{"id": "a3", "prompt": childHang + " contends", "description": "contends for the freed slot",
-			"read_only": true, "depends_on": []string{"a1"}},
-	})
-}
-
-// slotsFleet fills the total ceiling and asks for one more, all read-only: a
-// reader clears every writer check by not being a writer.
-func slotsFleet() []provider.Chunk {
-	return fleetCall("probe_fleet_refused", []map[string]any{
-		{"id": "s1", "prompt": childHang + " one", "description": "fills a slot", "read_only": true},
-		{"id": "s2", "prompt": childHang + " two", "description": "fills a slot", "read_only": true},
-		{"id": "s3", "prompt": childHang + " three", "description": "refused admission", "read_only": true},
-	})
-}
-
-// backgroundFleetCall dispatches a fleet that returns a job id at once. The
-// holder has to keep holding while a later fleet is dispatched, which a
-// foreground call cannot do: it does not return until every item has settled.
-func backgroundFleetCall(id string, tasks []map[string]any) []provider.Chunk {
-	args, _ := json.Marshal(map[string]any{"tasks": tasks, "run_in_background": true})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: id, Name: "fleet", Arguments: string(args)},
-	}}
-}
-
-// adoptableRef picks a reference the warm-up fleet reported. It reads the
-// conversation the model reads: the aggregate names each child's reference, and
-// an arm that cannot find one has not established its premise.
-func adoptableRef(req provider.Request) string {
-	for _, m := range slices.Backward(req.Messages) {
-		if found := subagentRefPattern.FindAllString(m.Content, -1); len(found) > 0 {
-			return found[len(found)-1]
-		}
-	}
-	return ""
-}
-
-// fanOutTurn is the prompt that opens one fleet.
-func fanOutTurn(n int, sentinel string) string {
-	return fmt.Sprintf("%s %s: dispatch the fan-out.", marker(n), sentinel)
-}
-
-// fanOutSentinels is what this arm's one prompt carries. The mixed arm names
-// both of its fleets in the same turn: the warm-up has to finish before a
-// reference can be adopted, and a fleet's aggregate returns into the same turn
-// that dispatched it, so the second round dispatches the mixed fleet without
-// the turn ever closing.
-func fanOutSentinels(arm string) string {
-	if arm == armUIGraphMixed {
-		return fleetWarmSentinel + " " + fleetSettledSentinel + " " + fleetMixedSentinel
-	}
-	if arm == armGraphMixed {
-		return fleetWarmSentinel + " " + fleetMixedSentinel
-	}
-	return fleetPairSentinel
-}
-
-// runFanOutConstruct drives the arm's fan-out and exits while the turn is still
-// open. Every arm dies mid-turn, including the completed one: what separates
-// them is which states the graph holds at that moment, and a clean shutdown
-// would cancel a running child and settle its node — the one thing a process
-// that dies mid-fan-out does not do.
-func runFanOutConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, turn int) error {
-	go func() { _ = ctrl.Run(ctx, fanOutTurn(turn+1, fanOutSentinels(arm))) }()
-	if err := waitForFanOut(sink, arm); err != nil {
-		return err
-	}
-	if err := writeObservation(root, capture("construct", arm, bootSystem, ctrl, sink, root)); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// waitForFanOut blocks until the graph shows the shape this arm was built to
-// die with. Waiting on the shape rather than on a duration is what keeps the
-// arm honest: a timer that fired early would report a smaller fan-out as if
-// that were what the restart inherited.
-func waitForFanOut(sink *graphSink, arm string) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		graph, _ := sink.snapshot()
-		if fanOutReady(graph, arm) {
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected("a fan-out holding "+wantedText(arm), nodeStates(graph))
-}
-
-// fanOutReady is the arm's premise expressed as a graph shape. The completed
-// arm waits for every worker to settle, which is the opposite condition from
-// the other two rather than a weaker version of it.
-func fanOutReady(g agentgraph.Graph, arm string) bool {
-	if arm == armGraphCompleted {
-		return allWorkersSettled(g)
-	}
-	return holdsAll(g, wantedStates(arm))
-}
-
-// wantedStates are the states a mid-flight arm must hold at once. Skipped is
-// not among them and cannot be: a fan-out publishes a skip only in its closing
-// outcome delta, so nothing reads skipped until the group has ended and nothing
-// is running. An item whose dependency already failed sits at pending until
-// then, which is what the mixed arm dies holding.
-func wantedStates(arm string) []agentgraph.NodeState {
-	want := []agentgraph.NodeState{agentgraph.StateCompleted, agentgraph.StateRunning}
-	if arm == armGraphMixed || arm == armUIGraphMixed {
-		want = append(want, agentgraph.StateFailed, agentgraph.StateAdopted)
-	}
-	// The UI arm dies holding a settled fan-out beside a live one, so it also
-	// reaches the two states the mixed arm cannot: a skip, which is published
-	// only when a group closes, and an item a ceiling never admitted.
-	if arm == armUIGraphMixed {
-		want = append(want, agentgraph.StateSkipped, agentgraph.StatePending, agentgraph.StateQueued)
-	}
-	return want
-}
-
-func wantedText(arm string) string {
-	if arm == armGraphCompleted {
-		return "every worker settled"
-	}
-	return statesText(wantedStates(arm))
-}
-
-// allWorkersSettled reports whether every worker reached a terminal state. An
-// empty state is not terminal here even though NodeState.Terminal calls it one:
-// a node declared before it ran carries no state at all, and reading that as
-// settled would let the arm die before the fan-out started.
-func allWorkersSettled(g agentgraph.Graph) bool {
-	seen := false
-	for _, n := range g.Nodes {
-		if n.Kind != agentgraph.KindWorker {
-			continue
-		}
-		seen = true
-		if n.State == "" || !n.State.Terminal() {
-			return false
-		}
-	}
-	return seen
-}
-
-// holdsAll reports whether the graph shows at least one worker in each state.
-// Group nodes are excluded: a group is the call that owns the fan-out, and its
-// own running state would satisfy the running requirement without any child
-// being in flight.
-func holdsAll(g agentgraph.Graph, want []agentgraph.NodeState) bool {
-	for _, state := range want {
-		found := false
-		for _, n := range g.Nodes {
-			if n.Kind == agentgraph.KindWorker && n.State == state {
-				found = true
-				break
-			}
-		}
-		if !found {
-			return false
-		}
-	}
-	return true
-}
-
-func statesText(states []agentgraph.NodeState) string {
-	out := make([]string, 0, len(states))
-	for _, s := range states {
-		out = append(out, string(s))
-	}
-	return join(out)
-}
-
-func nodeStates(g agentgraph.Graph) string {
-	out := make([]string, 0, len(g.Nodes))
-	for _, n := range g.Nodes {
-		out = append(out, n.ID+":"+string(n.State))
-	}
-	return orNone(join(out))
-}
@@ -1,646 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"reasonix/internal/state/sessionstore"
-	"slices"
-	"sort"
-	"strings"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/runtime/delegation"
-	"reasonix/internal/state/execgraph"
-	"reasonix/internal/state/execjournal"
-)
-
-// How a run graph can stand after the process that built it exits. Three are
-// designs a host may hold; two are defects, and not the same defect. Losing the
-// graph costs a reader an explanation; a node still drawn as running with no
-// owner costs them the truth — reviving actionable state is more dangerous than
-// dropping it, and a running node is actionable state.
-const (
-	graphExact       = "reconstructed-exact"  // nodes, edges, grants, waits, states all back
-	graphInterrupted = "interrupted-explicit" // the fan-out is legible and the live child is marked cut
-	graphLossy       = "reconstructed-lossy"  // it is back, minus semantics nothing can re-derive
-	graphLostSilent  = "LOST-SILENT"          // gone, with no provenance that it ever existed
-	graphGhost       = "GHOST-RUNNING"        // still says running, with no live owner anywhere
-	graphNotMeasured = "not-measured"         // the construct phase never established the shape
-)
-
-// graphRows classify what a restart inherited from a fan-out. There are two
-// outcome rows rather than one because a fan-out that died mid-flight holds two
-// populations with different answers: the items that had already settled, whose
-// answers the store owns, and the one still executing, which nothing on disk
-// has ever been told about. One label over both reports the better half.
-func graphRows(before, after Observation) []row {
-	return []row{
-		ghostRow(before, after),
-		dispatchTurnRow(before, after),
-		settledOutcomeRow(before, after),
-		interruptedOutcomeRow(before, after),
-		graphStructureRow(before, after),
-		graphSemanticsRow(before, after),
-		childProvenanceRow(before, after),
-		executionJournalRow(before, after),
-		rebuildStateRow(before, after),
-		rebuildTopologyRow(before, after),
-		rebuildEnvelopeRow(before, after),
-		rebuildInterruptedRow(before, after),
-		rebuildIdentityRow(before, after),
-		executionTopologyRow(before, after),
-		executionKindRow(before, after),
-		executionContextRow(before, after),
-		graphObligationRow(before, after),
-	}
-}
-
-// runningWorkers are the nodes that were executing when the process died. They
-// are the arm's premise: an arm that establishes none has measured a fan-out
-// nobody interrupted, whatever else it recorded.
-func runningWorkers(o Observation) []string {
-	var out []string
-	for _, n := range o.Graph.Nodes {
-		if n.Kind == agentgraph.KindWorker && n.State == agentgraph.StateRunning {
-			out = append(out, n.ID)
-		}
-	}
-	return out
-}
-
-// executedWorkers are the items that ran to a terminal state. Adopted nodes are
-// excluded on purpose: nothing executed under them, so the store owes them no
-// record and counting them would read a correct store as incomplete.
-func executedWorkers(o Observation) []string {
-	var out []string
-	for _, n := range o.Graph.Nodes {
-		if n.Kind != agentgraph.KindWorker || n.State == "" || n.State == agentgraph.StateAdopted {
-			continue
-		}
-		if n.State.Terminal() {
-			out = append(out, n.ID)
-		}
-	}
-	return out
-}
-
-// ghostRow asks the one question with a single acceptable answer: does anything
-// the new process can read still present this work as running? A graph node and
-// a store record are both such a claim, and neither has an owner behind it —
-// the goroutine that would have settled them died with the first process.
-func ghostRow(before, after Observation) row {
-	ghosts := append(runningWorkers(after), childrenInState(after, string(sessionstore.SubagentRunning))...)
-	verdict := verdictHolds
-	if len(ghosts) > 0 {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "work still presented as running", Authority: "the graph and the sub-agent store",
-		Artifact:       "subagents/<ref>.meta.json",
-		Reconstruction: "nothing may read as running once its owner is gone",
-		Before:         orNone(join(runningWorkers(before))) + " running at death",
-		After:          orNone(join(ghosts)), Verdict: verdict,
-	}
-}
-
-// dispatchTurnRow is the row underneath every other loss here: whether the turn
-// that opened the fan-out reached the transcript at all. A turn is appended
-// when it ends, so a process that dies inside one takes the request with it —
-// and a graph reconstruction has nothing to attach to even in principle.
-func dispatchTurnRow(before, after Observation) row {
-	verdict := verdictPersisted
-	switch {
-	case !before.FanOutTurn:
-		verdict = verdictNotMeasured
-	case !after.FanOutTurn:
-		verdict = verdictLost
-	}
-	return row{
-		Semantic: "the turn that dispatched the fan-out", Authority: "session transcript + event log",
-		Artifact: "<stem>.jsonl, <stem>.events.jsonl", Reconstruction: "appended when the turn ends",
-		Before: recordedText(before), After: recordedText(after), Verdict: verdict,
-	}
-}
-
-func recordedText(o Observation) string {
-	if o.FanOutTurn {
-		return fmt.Sprintf("in the transcript (%d rows)", o.Transcript.Messages)
-	}
-	return fmt.Sprintf("absent (%d rows)", o.Transcript.Messages)
-}
-
-// settledOutcomeRow classifies the items that had already finished. Their
-// answers are what a later turn reads back, so this is the row that decides
-// whether a fan-out's history is durable at all.
-func settledOutcomeRow(before, after Observation) row {
-	outcome, verdict := settledOutcome(before, after)
-	return row{
-		Semantic: "items that had settled before the death", Authority: "the graph fold and the sub-agent store",
-		Artifact:       "subagents/<ref>.meta.json",
-		Reconstruction: "one durable child record per item that executed",
-		Before:         fmt.Sprintf("%d executed: %s", len(executedWorkers(before)), graphSummary(before)),
-		After:          outcome, Verdict: verdict,
-	}
-}
-
-func settledOutcome(before, after Observation) (string, string) {
-	executed := len(executedWorkers(before))
-	switch {
-	case len(before.Graph.Nodes) == 0:
-		return graphNotMeasured, verdictNotMeasured
-	case graphIdentical(before, after):
-		return graphExact, verdictExact
-	case executed == 0:
-		return graphNotMeasured, verdictNotMeasured
-	case len(after.Children.Facts) == 0:
-		return graphLostSilent, verdictViolated
-	case len(after.Children.Facts) < executed:
-		return fmt.Sprintf("%s (%d of %d children)", graphLossy, len(after.Children.Facts), executed), verdictLossy
-	default:
-		return graphLossy, verdictLossy
-	}
-}
-
-// interruptedOutcomeRow classifies the item that was still executing. This is
-// the row the arm exists for: the graph is gone either way, and what separates
-// a defensible design from a durable gap is whether anything the new process
-// can read still says that work was opened and cut.
-func interruptedOutcomeRow(before, after Observation) row {
-	outcome, verdict := interruptedOutcome(before, after)
-	return row{
-		Semantic: "the item still executing at the death", Authority: "the sub-agent store and the transcript",
-		Artifact:       "subagents/<ref>.meta.json, <stem>.jsonl",
-		Reconstruction: "an interrupted child record, or an unsettled call the transcript still owes",
-		Before:         orNone(join(runningWorkers(before))), After: outcome, Verdict: verdict,
-	}
-}
-
-func interruptedOutcome(before, after Observation) (string, string) {
-	switch {
-	case len(before.Graph.Nodes) == 0 || len(runningWorkers(before)) == 0:
-		return graphNotMeasured, verdictNotMeasured
-	case len(runningWorkers(after)) > 0 || len(childrenInState(after, string(sessionstore.SubagentRunning))) > 0:
-		return graphGhost, verdictViolated
-	case len(after.InterruptedExecutions) > 0:
-		return graphInterrupted, verdictLossy
-	case len(childrenInState(after, string(sessionstore.SubagentInterrupted))) > 0:
-		return graphInterrupted, verdictLossy
-	case len(after.Obligation.UnansweredCalls) > 0 || after.Obligation.InterruptionMarked > 0:
-		return graphInterrupted, verdictLossy
-	default:
-		return graphLostSilent, verdictViolated
-	}
-}
-
-// graphIdentical compares the whole graph, not its node count. Two graphs with
-// the same nodes and different grants are not the same graph, and a comparison
-// that only counted would have called that exact.
-func graphIdentical(before, after Observation) bool {
-	return graphNodes(before) == graphNodes(after) &&
-		graphEdges(before) == graphEdges(after) &&
-		graphMap(before.Graph.Grants) == graphMap(after.Graph.Grants) &&
-		graphMap(before.Graph.Waits) == graphMap(after.Graph.Waits)
-}
-
-// graphStructureRow is the shape a reader asks for: which nodes, and the typed
-// edges between them. Spawn, depends, context and adopt are separate facts, and
-// a restart that kept the nodes while losing the edges kept a list, not a graph.
-func graphStructureRow(before, after Observation) row {
-	return valueRow("run graph edges", "event sink (GraphDelta)",
-		"none observed", "fold GraphDelta events, if any survive",
-		graphEdges(before), graphEdges(after), false)
-}
-
-// graphSemanticsRow is the authority envelope and the reason a node waited.
-// They are called out apart from the structure because nothing else in the
-// system records them: a child's persisted metadata carries its model and its
-// profile, and says nothing about what it was allowed to touch.
-func graphSemanticsRow(before, after Observation) row {
-	return valueRow("run graph grants and waits", "event sink (GraphDelta)",
-		"none observed", "no durable source carries an authority envelope",
-		graphEnvelope(before), graphEnvelope(after), false)
-}
-
-func graphEnvelope(o Observation) string {
-	grants, waits := graphMap(o.Graph.Grants), graphMap(o.Graph.Waits)
-	if grants == "" && waits == "" {
-		return ""
-	}
-	return fmt.Sprintf("grants[%s] waits[%s]", orNone(grants), orNone(waits))
-}
-
-// childProvenanceRow is what the store still owns, which is the evidence behind
-// an interrupted classification. It is reported as its own row so a reader can
-// see the reconstruction the label stands on rather than trusting the label.
-func childProvenanceRow(before, after Observation) row {
-	return valueRow("delegated runs the store owns", "SubagentStore, swept at boot",
-		"subagents/<ref>.meta.json", "ListSubagentsByParent; CleanupStaleRunning marks stale running children interrupted",
-		childSummary(before), childSummary(after), true)
-}
-
-// graphObligationRow is the last fallback a reader has: a fleet call the
-// transcript records with no result is the host saying a fan-out was opened and
-// never settled, even when nothing else remembers it.
-func graphObligationRow(before, after Observation) row {
-	return row{
-		Semantic: "fan-out calls the transcript never settled", Authority: "the canonical transcript",
-		Artifact: "<stem>.jsonl", Reconstruction: "tool calls with no result",
-		Before: orNone(join(before.Obligation.UnansweredCalls)),
-		After:  orNone(join(after.Obligation.UnansweredCalls)), Verdict: verdictStable,
-	}
-}
-
-// rebuiltGraph is what the production fold makes of this session's durable
-// facts, join included: a record names its own execution, and an older one is
-// aliased to its parent call only where the journal opened one by that id.
-func rebuiltGraph(o Observation) execgraph.Result {
-	opened := make(map[string]bool, len(o.Executions))
-	for _, e := range o.Executions {
-		opened[e.ID] = true
-	}
-	children := make([]execgraph.ChildOutcome, 0, len(o.Children.Facts))
-	for _, f := range o.Children.Facts {
-		identity := delegation.ResolveExecutionIdentity(
-			sessionstore.SubagentMeta{ExecutionID: f.ExecutionID, ParentToolCallID: f.ParentToolCallID},
-			func(id string) bool { return opened[id] })
-		children = append(children, execgraph.ChildOutcome{
-			Execution: identity.Execution, Ref: f.Ref, Status: f.Status,
-		})
-	}
-	return execgraph.Rebuild(o.Executions, children, nil)
-}
-
-// rebuildStateRow compares the states the fold produces against the ones the
-// dying process drew. Nodes whose owner disappeared are excluded and reported
-// by their own row: putting those back as running is the one outcome this
-// whole line of work exists to prevent.
-func rebuildStateRow(before, after Observation) row {
-	rebuilt := rebuiltGraph(after)
-	cut := map[string]bool{}
-	for _, i := range rebuilt.Interrupted {
-		cut[i.Execution] = true
-	}
-	return valueRow("rebuilt node states", "execgraph.Rebuild over journal + store",
-		"<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		"recomputed from durable facts, never restored",
-		statesExcept(graphOf(before), cut), statesExcept(rebuilt.Graph, cut), false)
-}
-
-// rebuildTopologyRow compares every typed edge. Spawn and depends are stated by
-// the openings, adopt names its source, and context is derived — so a mismatch
-// here says which of those four is not yet recoverable.
-func rebuildTopologyRow(before, after Observation) row {
-	return valueRow("rebuilt topology", "execgraph.Rebuild over journal + store",
-		"<stem>.execution.jsonl", "spawn and depends stated, adopt named, context derived",
-		edgeSummary(graphOf(before)), edgeSummary(rebuiltGraph(after).Graph), false)
-}
-
-// rebuildEnvelopeRow is the authority and the wait provenance together: what
-// each item was allowed to touch, and why admission was first refused.
-func rebuildEnvelopeRow(before, after Observation) row {
-	return valueRow("rebuilt grants and wait causes", "execgraph.Rebuild over journal + store",
-		"<stem>.execution.jsonl", "recorded at the opening and at the first refusal",
-		envelopeSummary(graphOf(before)), envelopeSummary(rebuiltGraph(after).Graph), false)
-}
-
-// rebuildInterruptedRow accounts for the nodes the state row leaves out. They
-// are not dropped: the fold names them as interruptions, split by whether they
-// had reached a slot. A node the dying process showed live and the rebuild
-// mentions nowhere would be a loss the state row could not see.
-func rebuildInterruptedRow(before, after Observation) row {
-	rebuilt := rebuiltGraph(after)
-	var got []string
-	for _, i := range rebuilt.Interrupted {
-		kind := execjournal.InterruptedBeforeStart
-		if i.Started {
-			kind = execjournal.InterruptedDuringExecution
-		}
-		got = append(got, i.Execution+":"+kind)
-	}
-	sort.Strings(got)
-	return valueRow("nodes the rebuild will not call live", "execgraph.Rebuild over journal + store",
-		"<stem>.execution.jsonl", "an execution with no owner is named, never redrawn as running",
-		liveAtDeath(before), join(got), false)
-}
-
-// liveAtDeath is what the dying process still showed as someone's work.
-func liveAtDeath(o Observation) string {
-	var out []string
-	for _, n := range o.Graph.Nodes {
-		if n.Kind == agentgraph.KindGroup {
-			continue
-		}
-		switch n.State {
-		case agentgraph.StateRunning:
-			out = append(out, n.ID+":"+execjournal.InterruptedDuringExecution)
-		case agentgraph.StateQueued, agentgraph.StatePending:
-			out = append(out, n.ID+":"+execjournal.InterruptedBeforeStart)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// rebuildIdentityRow is what a reader sees on a node beyond its state: the
-// label it was dispatched under, and which worker ran it. Only one of those is
-// recorded, so this row is where a remaining gap shows up rather than hiding
-// inside an exact verdict about states.
-func rebuildIdentityRow(before, after Observation) row {
-	return valueRow("rebuilt node identity", "execgraph.Rebuild over journal + store",
-		"<stem>.execution.jsonl", "label from the opening; the worker override is recorded nowhere",
-		identitySummary(graphOf(before)), identitySummary(rebuiltGraph(after).Graph), false)
-}
-
-func identitySummary(g agentgraph.Graph) string {
-	var out []string
-	for _, n := range g.Nodes {
-		if n.Kind != agentgraph.KindWorker {
-			continue
-		}
-		out = append(out, fmt.Sprintf("%s[label=%s profile=%s model=%s effort=%s]",
-			n.ID, orNone(n.Label), orNone(n.Profile), orNone(n.Model), orNone(n.Effort)))
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// statesExcept renders worker states, skipping the executions whose owner is
-// gone. Those are not a state the rebuild may assert, and comparing them would
-// demand exactly the ghost this refuses to draw.
-func statesExcept(g agentgraph.Graph, skip map[string]bool) string {
-	var out []string
-	for _, n := range g.Nodes {
-		if n.Kind == agentgraph.KindGroup || skip[n.ID] {
-			continue
-		}
-		out = append(out, n.ID+":"+string(n.State))
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func edgeSummary(g agentgraph.Graph) string {
-	out := make([]string, 0, len(g.Edges))
-	for _, e := range g.Edges {
-		out = append(out, fmt.Sprintf("%s-%s->%s", e.From, e.Kind, e.To))
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func envelopeSummary(g agentgraph.Graph) string {
-	var out []string
-	for _, n := range g.Nodes {
-		if n.Grant != "" {
-			out = append(out, n.ID+"="+string(n.Grant))
-		}
-		if n.Wait != "" {
-			out = append(out, n.ID+"@"+string(n.Wait))
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// executionJournalRow is the record written before the work started, which is
-// the only artifact here that can outlive an unfinished turn. It is what a
-// restart reads to say a delegation existed at all.
-func executionJournalRow(before, after Observation) row {
-	return valueRow("delegations the journal recorded", "execution journal, written before dispatch",
-		"<stem>.execution.jsonl", "folded from openings and settlings",
-		executionSummary(before), executionSummary(after), true)
-}
-
-// executionTopologyRow is the ordering a restart inherits. An item that never
-// started is explained by what it was held behind, so the topology has to
-// survive the process that knew it — and survive unchanged, or the explanation
-// is a different graph's.
-func executionTopologyRow(before, after Observation) row {
-	return valueRow("declared ordering between items", "the fan-out plan, recorded at the opening",
-		"<stem>.execution.jsonl", "read back with the openings, never re-derived",
-		topologySummary(before), topologySummary(after), true)
-}
-
-func topologySummary(o Observation) string {
-	var out []string
-	for _, e := range o.Executions {
-		if len(e.DependsOn) == 0 {
-			continue
-		}
-		up := append([]string(nil), e.DependsOn...)
-		sort.Strings(up)
-		out = append(out, e.ID+"<-"+strings.Join(up, "+"))
-	}
-	if len(out) == 0 {
-		return ""
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// executionKindRow holds the two interruptions apart against what the dying
-// process actually showed. The item the graph had running must come back as one
-// that reached a slot; the one its dependency blocked must come back as one
-// that did not. A classification agreeing with neither is a label, not a fact.
-func executionKindRow(before, after Observation) row {
-	running := runningWorkers(before)
-	verdict := verdictHolds
-	if len(after.InterruptedExecutions) == 0 {
-		verdict = verdictNotMeasured
-	}
-	for _, e := range after.InterruptedExecutions {
-		want := execjournal.InterruptedBeforeStart
-		if slices.Contains(running, e.ID) {
-			want = execjournal.InterruptedDuringExecution
-		}
-		if e.Interruption() != want {
-			verdict = verdictViolated
-		}
-	}
-	return row{
-		Semantic: "how each interruption is classified", Authority: "the execution journal's start records",
-		Artifact: "<stem>.execution.jsonl", Reconstruction: "started at death -> during-execution, never started -> before-start",
-		Before: orNone(join(running)) + " running at death", After: interruptionKinds(after), Verdict: verdict,
-	}
-}
-
-func interruptionKinds(o Observation) string {
-	if len(o.InterruptedExecutions) == 0 {
-		return "none"
-	}
-	out := make([]string, 0, len(o.InterruptedExecutions))
-	for _, e := range o.InterruptedExecutions {
-		out = append(out, e.ID+":"+e.Interruption())
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// executionContextRow is the other side of that record: what the next request
-// is told. It must state provenance and never offer a continuation, so the
-// count is checked against the interruptions rather than against a fixed
-// number — a block with nothing behind it is a ghost in prose.
-func executionContextRow(before, after Observation) row {
-	verdict := verdictHolds
-	if blocksOwed(before) != before.ModelSeesInterruptedExecution ||
-		blocksOwed(after) != after.ModelSeesInterruptedExecution {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "what the next request says about them", Authority: "control, projected per request",
-		Artifact: "none: derived, never stored", Reconstruction: "one block while an interrupted execution stands, none otherwise",
-		Before: executionBlocks(before), After: executionBlocks(after), Verdict: verdict,
-	}
-}
-
-func blocksOwed(o Observation) int {
-	if len(o.InterruptedExecutions) > 0 {
-		return 1
-	}
-	return 0
-}
-
-func executionBlocks(o Observation) string {
-	return fmt.Sprintf("%d block(s) for %d interrupted", o.ModelSeesInterruptedExecution, len(o.InterruptedExecutions))
-}
-
-func executionSummary(o Observation) string {
-	if len(o.Executions) == 0 {
-		return ""
-	}
-	open, started := 0, 0
-	for _, e := range o.Executions {
-		if e.Open() {
-			open++
-		}
-		if e.Started() {
-			started++
-		}
-	}
-	return fmt.Sprintf("%d opened, %d started, %d still open, %d with no owner here",
-		len(o.Executions), started, open, len(o.InterruptedExecutions))
-}
-
-func childrenInState(o Observation, status string) []string {
-	var out []string
-	for _, f := range o.Children.Facts {
-		if f.Status == status {
-			out = append(out, f.Ref)
-		}
-	}
-	return out
-}
-
-func childSummary(o Observation) string {
-	if len(o.Children.Facts) == 0 {
-		if o.Children.Err != "" {
-			return "error: " + o.Children.Err
-		}
-		return ""
-	}
-	counts := map[string]int{}
-	for _, f := range o.Children.Facts {
-		counts[f.Status]++
-	}
-	keys := make([]string, 0, len(counts))
-	for k := range counts {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
-	parts := make([]string, 0, len(keys))
-	for _, k := range keys {
-		parts = append(parts, fmt.Sprintf("%s=%d", k, counts[k]))
-	}
-	return fmt.Sprintf("%d child(ren): %s", len(o.Children.Facts), strings.Join(parts, " "))
-}
-
-// graphSummary is the fan-out as the dying process last saw it, which is the
-// "before" every graph row is compared against.
-func graphSummary(o Observation) string {
-	if len(o.Graph.Nodes) == 0 {
-		return "no fan-out"
-	}
-	counts := map[string]int{}
-	for _, n := range o.Graph.Nodes {
-		if n.Kind == agentgraph.KindWorker {
-			counts[string(n.State)]++
-		}
-	}
-	keys := make([]string, 0, len(counts))
-	for k := range counts {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
-	parts := make([]string, 0, len(keys))
-	for _, k := range keys {
-		parts = append(parts, fmt.Sprintf("%s=%d", k, counts[k]))
-	}
-	return fmt.Sprintf("%d node(s), %d edge(s): %s",
-		len(o.Graph.Nodes), len(o.Graph.Edges), strings.Join(parts, " "))
-}
-
-// graphArmInvalid refuses an arm that never reached the state it names. The
-// running arms need a child in flight at death; every fan-out arm needs a
-// fan-out at all. Reporting a smaller shape as the thing that was interrupted
-// is how a probe answers a question nobody asked.
-func graphArmInvalid(arm string, before Observation) string {
-	if len(before.Graph.Nodes) == 0 {
-		return "no fan-out was published before the process exited, so nothing was measured"
-	}
-	if arm == armGraphCompleted {
-		if len(runningWorkers(before)) > 0 {
-			return "the fan-out was still running at exit, so this is not the completed arm"
-		}
-		return ""
-	}
-	if len(runningWorkers(before)) == 0 {
-		return "no child was executing when the process died, so nothing was interrupted"
-	}
-	if arm == armGraphMixed {
-		return mixedArmInvalid(before)
-	}
-	return ""
-}
-
-// mixedArmInvalid holds the density arm to its own premise. Its whole claim is
-// that one death carried every reachable state at once, and an arm that reached
-// three of them measures a different scenario than the one it reports.
-func mixedArmInvalid(before Observation) string {
-	var missing []string
-	for _, state := range wantedStates(armGraphMixed) {
-		if !holdsAll(agentgraph.Graph{Nodes: before.Graph.Nodes}, []agentgraph.NodeState{state}) {
-			missing = append(missing, string(state))
-		}
-	}
-	if len(missing) > 0 {
-		return "the fan-out never reached " + join(missing) + ", so the arm is not the mixed shape it reports"
-	}
-	if !carriesGrant(before, agentgraph.GrantRead) || !carriesGrant(before, agentgraph.GrantWrite) {
-		return "the fan-out carried only one kind of grant, so the envelope row measures one side of it"
-	}
-	return ""
-}
-
-func carriesGrant(o Observation, want agentgraph.Grant) bool {
-	for _, got := range o.Graph.Grants {
-		if got == string(want) {
-			return true
-		}
-	}
-	return false
-}
-
-// graphEdges renders the typed relations in a stable order. Sorting is not
-// cosmetic: the fold keeps first-seen order, which two processes reach by
-// different routes, and an unsorted comparison would report a reordering as a
-// changed graph.
-func graphEdges(o Observation) string {
-	if len(o.Graph.Edges) == 0 {
-		return ""
-	}
-	out := make([]string, 0, len(o.Graph.Edges))
-	for _, e := range o.Graph.Edges {
-		out = append(out, fmt.Sprintf("%s-%s->%s", e.From, e.Kind, e.To))
-	}
-	sort.Strings(out)
-	return join(out)
-}
@@ -1,602 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"sort"
-	"strings"
-	"sync/atomic"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/session/control"
-)
-
-// The interactive host-initiated arms, named for the one surface they measure: a
-// person typing /<skill> into a live session. The headless counterpart is not
-// measured here — it is caller-owned and its cancellation need not match, and
-// reading one for the other is how a real number gets the wrong subject.
-const (
-	armHostCompleted = "interactive-host-skill-completed"
-	armHostQueued    = "interactive-host-skill-queued-crash"
-	armHostRunning   = "interactive-host-skill-running-crash"
-	armHostCancel    = "interactive-host-skill-cancel"
-	// The headless counterpart: the controller's own synchronous entry point,
-	// whose caller owns both the context and the cancellation. It is measured
-	// apart from the four above and compared to them, never read for them.
-	armHeadlessCompleted     = "headless-host-skill-completed"
-	armHeadlessRunningCancel = "headless-host-skill-running-cancel"
-	armHeadlessQueuedCancel  = "headless-host-skill-queued-cancel"
-)
-
-func headlessHostArm(name string) bool {
-	switch name {
-	case armHeadlessCompleted, armHeadlessRunningCancel, armHeadlessQueuedCancel:
-		return true
-	}
-	return false
-}
-
-// callerStopArm names the arms a caller stops through its own context. The
-// controller never admitted these turns, so its own cancel holds nothing for
-// them — reading one surface with the other's stop is a mistake this probe has
-// already made once.
-func callerStopArm(name string) bool {
-	return name == armHeadlessRunningCancel || name == armHeadlessQueuedCancel
-}
-
-func hostExecArm(name string) bool {
-	switch name {
-	case armHostCompleted, armHostQueued, armHostRunning, armHostCancel:
-		return true
-	}
-	return headlessHostArm(name)
-}
-
-// queuedHostArm names the arms whose ceiling is filled before the invocation.
-func queuedHostArm(name string) bool {
-	return name == armHostQueued || name == armHeadlessQueuedCancel
-}
-
-// The four identities one host-started run can be known by. They are printed
-// apart because this is the first surface where they may not coincide: for a
-// model-delegated run the execution and the parent call are one string, and
-// that has only ever been true because a tool call produced every durable
-// execution there was.
-const (
-	probeHostDispatch  = "probe:host-dispatch-id"
-	probeHostChildRef  = "probe:host-child-ref"
-	probeHostParentID  = "probe:host-parent-lineage"
-	probeHostExecution = "probe:host-execution-id"
-	// probeHostStanding is what the call itself was doing at the reading:
-	// admitted or held back, its child in the provider or not.
-	probeHostStanding = "probe:host-standing"
-	// probeHostAuthority is the snapshot at the same instant: what a reader
-	// starting from the authority would hold, on either surface.
-	probeHostAuthority = "probe:host-authority"
-)
-
-// hostSlashInput is what a person types. The task carries the child sentinel so
-// the scripted provider knows how this arm's child behaves.
-func hostSlashInput(arm string) string {
-	return "/" + probeSkillName + " " + hostChildSentinel(arm) + " host-started work"
-}
-
-// runHostExecConstruct drives one interactive slash invocation and dies where
-// this arm is about. The queued arm fills the ceiling first with a background
-// fleet, which is dispatched through Run rather than Submit: Run is never
-// admitted through the turn gate, so the slash the arm measures can still be.
-func runHostExecConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	if queuedHostArm(arm) {
-		go func() { _ = ctrl.Run(ctx, fanOutTurn(turn+1, fleetHolderSentinel)) }()
-		if err := waitForCeiling(prov); err != nil {
-			return err
-		}
-	}
-	// Each surface is entered and stopped by the owner it has: the controller's
-	// gate for a turn it admitted, and the caller's own context for a
-	// synchronous call it never did.
-	callerCtx, stopCaller := context.WithCancel(ctx)
-	defer stopCaller()
-	var answered atomic.Bool
-	if headlessHostArm(arm) {
-		go func() {
-			_, _ = ctrl.RunSubagentProfile(callerCtx, probeSkillName, hostHeadlessTask(arm), false)
-			answered.Store(true)
-		}()
-	} else {
-		go ctrl.Submit(hostSlashInput(arm))
-	}
-	if err := waitForHostExec(ctrl, prov, arm, &answered); err != nil {
-		return err
-	}
-	stop := ""
-	switch {
-	case arm == armHostCancel:
-		stop = stopHostExec(ctrl, prov)
-	case callerStopArm(arm):
-		stop = stopHostCaller(ctrl, prov, arm, stopCaller)
-	}
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	recordHostIdentities(&obs, ctrl, sink, prov, arm)
-	if stop != "" {
-		obs.Progress[probeStopReached] = []string{stop}
-	}
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// recordHostIdentities writes down all four names this run could be known by,
-// at one instant. A single "identity: none" would hide which of them exists and
-// which does not, and that difference is the whole question.
-func recordHostIdentities(obs *Observation, ctrl *control.Controller, sink *graphSink, prov *scripted, arm string) {
-	obs.Progress[probeChildEntered] = []string{fmt.Sprint(prov.fleets.held(childEntered(hostChildSentinel(arm))))}
-	obs.Progress[probeHostDispatch] = []string{orNone(join(hostDispatchIDs(sink)))}
-	obs.Progress[probeHostChildRef] = []string{orNone(join(hostChildRefs(*obs)))}
-	obs.Progress[probeHostParentID] = []string{orNone(join(hostParentLineage(*obs)))}
-	obs.Progress[probeHostExecution] = []string{orNone(join(hostExecutions(*obs)))}
-	obs.Progress[probeHostStanding] = []string{hostStanding(ctrl, prov, arm)}
-	// The authority, read at the same instant. It is what a reader starts from
-	// on either surface, so it is where the two are compared — the delta stream
-	// exists only where a sink does, and a synchronous caller has none.
-	obs.Progress[probeHostAuthority] = []string{hostAuthorityNodes(ctrl)}
-}
-
-// hostHeadlessTask is what the controller's own entry point is asked to do.
-func hostHeadlessTask(arm string) string {
-	return hostChildSentinel(arm) + " host-started work"
-}
-
-func hostAuthorityNodes(ctrl *control.Controller) string {
-	var out []string
-	for _, n := range ctrl.ExecutionGraph().Graph.Nodes {
-		if ofHostExecution(n.ID) {
-			out = append(out, n.ID+":"+orDash(string(n.State)))
-		}
-	}
-	sort.Strings(out)
-	return orNone(join(out))
-}
-
-// stopHostCaller stops work through the only owner a synchronous call has: the
-// context its caller passed in. The controller's gate never admitted this turn,
-// so asking it to stop one would be reading a surface with another's stop.
-func stopHostCaller(ctrl *control.Controller, prov *scripted, arm string, stop func()) string {
-	stopped := time.Now()
-	stop()
-	child := hostChildSentinel(arm)
-	deadline := time.Now().Add(20 * time.Second)
-	for time.Now().Before(deadline) && !prov.fleets.held(childLeft(child)) {
-		time.Sleep(50 * time.Millisecond)
-	}
-	live := "gone"
-	if prov.fleets.held(childEntered(child)) && !prov.fleets.held(childLeft(child)) {
-		live = "still running"
-	}
-	return fmt.Sprintf("turn=%s work=%s ctx=%t after=%s", turnStanding(ctrl, arm), live,
-		prov.fleets.held(childCtxDone(child)), time.Since(stopped).Round(time.Second))
-}
-
-func hostChildSentinel(arm string) string {
-	if arm == armHostCompleted || arm == armHeadlessCompleted {
-		return childDone
-	}
-	return childHang
-}
-
-// hostDispatchIDs are the synthetic event ids the host minted so the child's
-// activity could nest under something on screen. They are UI identity, not a
-// call the model made, which is why no persisted lineage may carry one.
-func hostDispatchIDs(sink *graphSink) []string {
-	var out []string
-	for id := range sink.toolIDs() {
-		if strings.HasPrefix(id, "slash-skill-") {
-			out = append(out, id)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// hostChildRefs are the store's own artifact ids. They exist only once a child
-// transcript has been prepared, which is why they cannot serve as the identity
-// of work the scheduler is still holding back.
-func hostChildRefs(o Observation) []string {
-	var out []string
-	for _, f := range o.Children.Facts {
-		if f.Kind == "skill" {
-			out = append(out, f.Ref+"="+f.Status)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-func hostParentLineage(o Observation) []string {
-	var out []string
-	for _, f := range o.Children.Facts {
-		if f.Kind == "skill" {
-			out = append(out, f.Ref+"→"+orDash(f.ParentToolCallID))
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-func hostExecutions(o Observation) []string {
-	var out []string
-	for _, e := range o.Executions {
-		if ofHostExecution(e.ID) || strings.Contains(e.Name, probeSkillName) {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// ofHostExecution matches both names this work can go by: the execution the host
-// owns, and the synthetic event id it mints for the screen. Reading only the
-// second is how an arm reports absence when what changed is the identity.
-func ofHostExecution(id string) bool {
-	return strings.HasPrefix(id, "host-") || strings.HasPrefix(id, "slash-skill-")
-}
-
-func hostStanding(ctrl *control.Controller, prov *scripted, arm string) string {
-	return fmt.Sprintf("turn=%v ceiling=%t child=%t", ctrl.Running(),
-		prov.fleets.held(childEntered(childHold)),
-		prov.fleets.held(childEntered(hostChildSentinel(arm))))
-}
-
-func waitForCeiling(prov *scripted) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		if prov.fleets.held(childEntered(childHold)) {
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return errUnexpected("a fan-out holding the ceiling", "none reached the provider")
-}
-
-// waitForHostExec blocks until the invocation stands where this arm dies. The
-// queued arm asks a negative — the child never arrived while the ceiling was
-// full — so it settles and asks again rather than reading a race.
-func waitForHostExec(ctrl *control.Controller, prov *scripted, arm string, answered *atomic.Bool) error {
-	child := hostChildSentinel(arm)
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		switch {
-		case queuedHostArm(arm):
-			// In flight, read through whichever owner this surface has: a turn
-			// the controller admitted answers its gate, and a synchronous call
-			// it never admitted answers only by returning.
-			if hostInFlight(ctrl, arm, answered) && !prov.fleets.held(childEntered(child)) {
-				time.Sleep(2 * time.Second)
-				if prov.fleets.held(childEntered(child)) {
-					return errUnexpected("a host invocation the scheduler was still holding back", "its child ran")
-				}
-				return nil
-			}
-		case arm == armHostCompleted:
-			if prov.fleets.held(childLeft(child)) && !ctrl.Running() {
-				return nil
-			}
-		case arm == armHeadlessCompleted:
-			if prov.fleets.held(childLeft(child)) {
-				return nil
-			}
-		default:
-			if prov.fleets.held(childEntered(child)) {
-				return nil
-			}
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return errUnexpected("a host invocation "+hostWanted(arm), hostStanding(ctrl, prov, arm))
-}
-
-// hostInFlight reports that the invocation has not answered yet, asked of the
-// owner the surface actually has.
-func hostInFlight(ctrl *control.Controller, arm string, answered *atomic.Bool) bool {
-	if headlessHostArm(arm) {
-		return !answered.Load()
-	}
-	return ctrl.Running()
-}
-
-func hostWanted(arm string) string {
-	switch {
-	case queuedHostArm(arm):
-		return "held back by a full ceiling"
-	case arm == armHostCompleted || arm == armHeadlessCompleted:
-		return "that finished"
-	}
-	return "with its child inside the provider"
-}
-
-// stopHostExec stops the work the way a person does: the turn's own cancel. A
-// slash invocation is admitted through the same gate a typed message is, so
-// that gate is the owner of its stop.
-func stopHostExec(ctrl *control.Controller, prov *scripted) string {
-	stopped := time.Now()
-	ctrl.Cancel()
-	deadline := time.Now().Add(20 * time.Second)
-	for time.Now().Before(deadline) && ctrl.Running() {
-		time.Sleep(50 * time.Millisecond)
-	}
-	live := "gone"
-	if prov.fleets.held(childEntered(childHang)) && !prov.fleets.held(childLeft(childHang)) {
-		live = "still running"
-	}
-	return fmt.Sprintf("turn=%s work=%s ctx=%t after=%s", turnStanding(ctrl, armHostCancel), live,
-		prov.fleets.held(childCtxDone(childHang)), time.Since(stopped).Round(time.Second))
-}
-
-// The verdicts this arm needs. They are separate words because the failure
-// modes are separate facts, and one label over all of them would say less than
-// any of them does.
-const (
-	// verdictIdentityGap is work the host admitted that no identity outliving
-	// the process names. It is not about a lost record: none of the candidate
-	// identities can carry one.
-	verdictIdentityGap = "IDENTITY-GAP"
-	// verdictResultDurable is an invocation whose child record survives while
-	// the execution that produced it does not. A reader can see that a child of
-	// this name was cut; nothing says which invocation it belonged to.
-	verdictResultDurable = "record-only"
-)
-
-func hostExecRows(arm string, before, after Observation) []row {
-	return []row{
-		hostAdmissionRow(arm, before),
-		hostIdentityRow(before),
-		hostGraphRow(before),
-		hostJournalRow(before, after),
-		hostStoreRow(before, after),
-		hostLineageOracleRow(before, after),
-		hostPreStartIdentityRow(arm, before, after),
-		hostCancelRow(arm, before),
-		hostRestartRow(arm, before, after),
-	}
-}
-
-func hostAdmissionRow(arm string, before Observation) row {
-	entered := firstProgress(before, probeChildEntered) == "true"
-	answer, verdict := "admitted: the child reached the provider", verdictHolds
-	switch {
-	case queuedHostArm(arm) && !entered:
-		answer = "refused: the ceiling was full and the child never arrived"
-	case queuedHostArm(arm):
-		answer, verdict = "not held back: the child ran with the ceiling full", verdictViolated
-	case !entered:
-		answer, verdict = "the child never arrived", verdictNotMeasured
-	}
-	return row{
-		Semantic: "does it ask the session scheduler", Authority: skillAuthority,
-		Artifact: "none observed", Reconstruction: "in-process only",
-		Before: answer + " [" + firstProgress(before, probeHostStanding) + "]",
-		After:  "—", Verdict: verdict,
-	}
-}
-
-// hostIdentityRow prints all four names at once. This is the row the arm exists
-// for: for a model-delegated run the execution and the parent call are the same
-// string, and that has only ever held because a tool call produced every durable
-// execution there was.
-func hostIdentityRow(before Observation) row {
-	return row{
-		Semantic: "the four names this run could be known by", Authority: crossAuthority,
-		Artifact:       "the event stream, subagents/<ref>.meta.json, <stem>.execution.jsonl",
-		Reconstruction: "read apart, because they may not coincide here",
-		Before: fmt.Sprintf("dispatch=%s child=%s lineage=%s execution=%s",
-			firstProgress(before, probeHostDispatch), firstProgress(before, probeHostChildRef),
-			firstProgress(before, probeHostParentID), firstProgress(before, probeHostExecution)),
-		After: "—", Verdict: verdictHolds,
-	}
-}
-
-func hostGraphRow(before Observation) row {
-	var drawn []string
-	for _, n := range before.Graph.Nodes {
-		if ofHostExecution(n.ID) {
-			drawn = append(drawn, n.ID+":"+string(n.State))
-		}
-	}
-	sort.Strings(drawn)
-	// The stream and the authority are reported together because only one of
-	// them is surface-independent: a synchronous caller has no sink to draw on,
-	// and a reader on either surface starts from the snapshot regardless.
-	authority := firstProgress(before, probeHostAuthority)
-	return row{
-		Semantic:  "the graph, drawn and reconstructable",
-		Authority: liveAuthority + " and " + authorityName,
-		Artifact:  "<stem>.execution.jsonl", Reconstruction: "the stream where a sink exists; the authority always",
-		Before: fmt.Sprintf("stream=%s authority=%s", orNone(join(drawn)), authority),
-		After:  "—", Verdict: presenceVerdict(authority != "none"),
-	}
-}
-
-func hostJournalRow(before, after Observation) row {
-	return presenceRow("does the journal record it", journalAuthority, "<stem>.execution.jsonl",
-		"ExecutionHistory over the whole session", hostExecutions(before), hostExecutions(after))
-}
-
-func hostStoreRow(before, after Observation) row {
-	return presenceRow("what the store holds for it", storeAuthority, "subagents/<ref>.meta.json",
-		"ListSubagentsByParent over the transcript stem", hostChildRefs(before), hostChildRefs(after))
-}
-
-// hostLineageOracleRow is the contract Step 8 respected and this must not
-// weaken. A synthetic host event id is UI identity; persisting one as a parent
-// would say the model issued a call it never made.
-func hostLineageOracleRow(before, after Observation) row {
-	leaked := hostSyntheticLeaks(before)
-	leaked = append(leaked, hostSyntheticLeaks(after)...)
-	empty := hostLineageEmpty(before) && hostLineageEmpty(after)
-	answer, verdict := "no persisted lineage names a synthetic host id, and every parent is empty", verdictHolds
-	switch {
-	case len(leaked) > 0:
-		answer, verdict = "a synthetic host id was persisted as lineage: "+join(leaked), verdictViolated
-	case !empty:
-		answer, verdict = "a host-started run was filed under a parent call: "+
-			join(hostParentLineage(after)), verdictViolated
-	case len(hostChildRefs(before))+len(hostChildRefs(after)) == 0:
-		verdict = verdictNotMeasured
-		answer = "no child record exists to carry a lineage"
-	}
-	return row{
-		Semantic: "synthetic host id persisted as lineage", Authority: storeAuthority,
-		Artifact: "subagents/<ref>.meta.json", Reconstruction: "every persisted parent, read for host event ids",
-		Before: answer, After: "—", Verdict: verdict,
-	}
-}
-
-func hostSyntheticLeaks(o Observation) []string {
-	var out []string
-	for _, f := range o.Children.Facts {
-		if strings.HasPrefix(f.ParentToolCallID, "slash-skill-") {
-			out = append(out, f.Ref+"→"+f.ParentToolCallID)
-		}
-	}
-	return out
-}
-
-func hostLineageEmpty(o Observation) bool {
-	for _, f := range o.Children.Facts {
-		if f.Kind == "skill" && strings.TrimSpace(f.ParentToolCallID) != "" {
-			return false
-		}
-	}
-	return true
-}
-
-// hostPreStartIdentityRow is what decides where an identity would have to be
-// created. Work the scheduler is holding back has no child artifact yet, so a
-// store ref cannot name it; if nothing else survives the process either, no
-// candidate identity exists at the moment the work first became real.
-func hostPreStartIdentityRow(arm string, before, after Observation) row {
-	survives := len(hostExecutions(after)) > 0
-	answer := "nothing that outlives the process names this work"
-	verdict := verdictIdentityGap
-	if survives {
-		answer, verdict = "an execution the journal opened: "+join(hostExecutions(after)), verdictHolds
-	}
-	if !queuedHostArm(arm) && firstProgress(before, probeChildEntered) != "true" {
-		verdict = verdictNotMeasured
-	}
-	return row{
-		Semantic: "an identity that exists before the child does", Authority: crossAuthority,
-		Artifact:       "<stem>.execution.jsonl",
-		Reconstruction: "the candidates, read against what the work needs",
-		Before:         hostAtDeath(arm), After: answer, Verdict: verdict,
-	}
-}
-
-func hostAtDeath(arm string) string {
-	switch {
-	case queuedHostArm(arm):
-		return "admitted-domain work, held back, no child artifact yet"
-	case arm == armHostCompleted || arm == armHeadlessCompleted:
-		return "ran and returned"
-	}
-	return "executing inside the provider"
-}
-
-func hostCancelRow(arm string, before Observation) row {
-	if arm != armHostCancel && !callerStopArm(arm) {
-		return row{
-			Semantic: "what a stop reaches", Authority: "the turn's own cancel",
-			Artifact: "none observed", Reconstruction: "in-process only",
-			Before: "not stopped in this arm", After: "—", Verdict: verdictNotMeasured,
-		}
-	}
-	return row{
-		Semantic: "what a stop reaches", Authority: "the turn's own cancel",
-		Artifact:       "subagents/<ref>.meta.json",
-		Reconstruction: "the live stop and the record it left, kept apart",
-		Before:         firstProgress(before, probeStopReached) + " store=" + orNone(join(hostChildRefs(before))),
-		After:          "—", Verdict: verdictHolds,
-	}
-}
-
-// hostRestartRow classifies what the next process inherits. An answer that
-// survived while its execution did not is not a loss to the user, so it is not
-// called one; work that was executing and left nothing is.
-func hostRestartRow(arm string, before, after Observation) row {
-	exec, children := hostExecutions(after), hostChildRefs(after)
-	rebuilt := hostRebuilt(after)
-	answer := fmt.Sprintf("execution=%s store=%s rebuilt=%s",
-		orNone(join(exec)), orNone(join(children)), orNone(join(rebuilt)))
-	return row{
-		Semantic: "what a restart inherits", Authority: rebuildAuthority,
-		Artifact:       "<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		Reconstruction: "execgraph.Rebuild, then the store beside it",
-		Before:         hostAtDeath(arm), After: answer,
-		Verdict: hostRestartVerdict(arm, before, exec, children),
-	}
-}
-
-// hostRestartVerdict is driven by what a next process can actually hold, not by
-// what the arm expected. A child record with nothing to place it is not silence:
-// a reader can still see that a child of this name was cut. What it cannot do is
-// say which invocation that was, and the identity row answers that separately —
-// folding the two would report a surviving record as a total loss.
-func hostRestartVerdict(arm string, before Observation, exec, children []string) string {
-	switch {
-	case len(exec) > 0:
-		return verdictHolds
-	case len(children) > 0:
-		return verdictResultDurable
-	case firstProgress(before, probeChildEntered) == "true" || queuedHostArm(arm):
-		return graphLostSilent
-	default:
-		return verdictNotMeasured
-	}
-}
-
-// hostRebuilt is what a restart draws for this invocation. The whole rebuilt
-// graph is wider than the arm — the queued arm fills its ceiling with a fan-out,
-// and a fan-out records everything — so reading its nodes as the invocation's is
-// how an arm reports a gap as closed by evidence belonging to something else.
-func hostRebuilt(o Observation) []string {
-	var out []string
-	for _, n := range rebuiltGraph(o).Graph.Nodes {
-		if n.Kind == agentgraph.KindGroup {
-			continue
-		}
-		if ofHostExecution(n.ID) || strings.Contains(n.Label, probeSkillName) {
-			out = append(out, n.ID+":"+string(n.State))
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// hostExecArmInvalid reports the premise this arm could not establish.
-func hostExecArmInvalid(arm string, before Observation) string {
-	entered := firstProgress(before, probeChildEntered) == "true"
-	switch {
-	case queuedHostArm(arm) && entered:
-		return "the invocation ran while the ceiling was meant to be full"
-	case !queuedHostArm(arm) && !entered:
-		return "the invocation's child never reached the provider"
-	case headlessHostArm(arm) && len(hostDispatchIDs2(before)) > 0:
-		return "a synchronous call minted a dispatch id, so this was not the headless surface"
-	case !headlessHostArm(arm) && len(hostDispatchIDs2(before)) == 0:
-		return "the host minted no dispatch id, so this was not the slash surface"
-	}
-	return ""
-}
-
-// hostDispatchIDs2 reads the dispatch id back off the observation, where the
-// construct phase recorded what only a live sink could see.
-func hostDispatchIDs2(o Observation) []string {
-	if got := firstProgress(o, probeHostDispatch); got != "none" && got != "unrecorded" {
-		return strings.Split(got, ", ")
-	}
-	return nil
-}
@@ -1,163 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"strings"
-	"time"
-
-	"reasonix/internal/session/control"
-)
-
-// Two trees, read at once. A stop aimed at a turn must close the turn's own
-// tree and leave the job's alone; an explicit job stop does the reverse, and
-// reading only one of them is how a probe calls a violated handoff correct. The
-// manager's own root needs no separate question: a job still running with its
-// child's context open is the root answering directly.
-const (
-	// armLineageForeground is the positive control: it proves the reading can
-	// see an ordinary cancellation arrive.
-	armLineageForeground = "lineage-foreground-cancel"
-	// armLineageBackground is the one under suspicion.
-	armLineageBackground = "lineage-background-cancel"
-	// armLineageJobKill is the other positive control: without it, a reading
-	// that could never observe a job's own cancellation would make the arm
-	// above pass for the wrong reason.
-	armLineageJobKill = "lineage-job-kill"
-)
-
-func lineageArm(name string) bool {
-	switch name {
-	case armLineageForeground, armLineageBackground, armLineageJobKill:
-		return true
-	}
-	return false
-}
-
-// runLineageConstruct drives one delegation, stops it the way this arm's owner
-// stops it, and reads both trees afterwards. It records six facts, kept apart
-// on purpose: a context that closed and a call that returned are different
-// events, and only their combination says whether a cancellation propagated or
-// something else ended the work.
-func runLineageConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	ctrl.Send(fanOutTurn(turn+1, taskSentinel))
-	if err := waitFor(60*time.Second, func() bool { return prov.fleets.held(childEntered(childHang)) }); err != nil {
-		return errUnexpected("a delegated child inside the provider", "none arrived")
-	}
-	before := lineageReading(ctrl, prov, "before")
-
-	switch arm {
-	case armLineageJobKill:
-		jobs := ctrl.Jobs()
-		if len(jobs) == 0 {
-			return errUnexpected("a background job to stop", "none running")
-		}
-		for _, job := range jobs {
-			ctrl.CancelJob(job.ID)
-		}
-	default:
-		ctrl.Cancel()
-	}
-	// Long enough for a cancellation that was going to arrive to have arrived.
-	time.Sleep(3 * time.Second)
-	after := lineageReading(ctrl, prov, "after")
-
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	obs.Progress[probeLineage] = []string{before, after}
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-const probeLineage = "probe:lineage"
-
-// lineageReading is the state of both trees at one instant, in the terms the
-// rows judge: the turn's gate, the job's own liveness, the child's context, the
-// child's call, and the durable record the orchestration keeps.
-func lineageReading(ctrl *control.Controller, prov *scripted, at string) string {
-	return fmt.Sprintf("%s turn=%s job=%s childCtx=%s childCall=%s record=%s",
-		at, turnGateStanding(ctrl), jobStanding(ctrl),
-		yesNo(prov.fleets.held(childCtxDone(childHang)), "cancelled", "open"),
-		yesNo(prov.fleets.held(childLeft(childHang)), "returned", "inside"),
-		loneRecordStanding(ctrl))
-}
-
-func turnGateStanding(ctrl *control.Controller) string {
-	if ctrl.Running() {
-		return "active"
-	}
-	return "ended"
-}
-
-func jobStanding(ctrl *control.Controller) string {
-	if len(ctrl.Jobs()) > 0 {
-		return "running"
-	}
-	return "gone"
-}
-
-func loneRecordStanding(ctrl *control.Controller) string {
-	for _, e := range control.ExecutionHistory(ctrl.SessionPath()) {
-		if e.ID == parallelLoneNodeID {
-			return lifecycleOf(e)[1:]
-		}
-	}
-	return "absent"
-}
-
-func yesNo(ok bool, yes, no string) string {
-	if ok {
-		return yes
-	}
-	return no
-}
-
-func waitFor(within time.Duration, ready func() bool) error {
-	deadline := time.Now().Add(within)
-	for time.Now().Before(deadline) {
-		if ready() {
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return fmt.Errorf("timed out")
-}
-
-// lineageRows read the two trees against the ownership each arm is about. The
-// three sets of expectations are deliberately different: a fix that closes
-// everything on a stop satisfies the foreground arm and violates both others.
-func lineageRows(arm string, before Observation) []row {
-	at := reading(before.Progress[probeLineage], 1)
-	turnEnded := strings.Contains(at, "turn=ended")
-	jobLive := strings.Contains(at, "job=running")
-	childCancelled := strings.Contains(at, "childCtx=cancelled")
-	recordOpen := !strings.Contains(at, "settled") && !strings.Contains(at, "record=absent")
-
-	rows := []row{lineageRow("the turn the stop was aimed at", at, "ends", turnEnded)}
-	if arm == armLineageForeground {
-		return append(rows,
-			lineageRow("the child's own context", at, "closes with the turn that owned it", childCancelled),
-			lineageRow("the delegation record", at, "is let go of", !recordOpen))
-	}
-	if arm == armLineageJobKill {
-		return append(rows[:0],
-			lineageRow("the job the stop was aimed at", at, "ends", !jobLive),
-			lineageRow("the child's own context", at, "closes with the job that owned it", childCancelled),
-			lineageRow("the delegation record", at, "is let go of", !recordOpen))
-	}
-	return append(rows,
-		lineageRow("the job the turn no longer owns", at, "keeps running", jobLive),
-		lineageRow("the child's own context", at, "is not closed by another owner's stop", !childCancelled),
-		lineageRow("the delegation record", at, "stays open while its child runs", recordOpen))
-}
-
-func lineageRow(semantic, got, want string, ok bool) row {
-	return row{
-		Semantic: semantic, Authority: "both trees, read at the same instant",
-		Artifact: "none: in-process state", Reconstruction: want,
-		Before: want, After: orNone(got), Verdict: held(ok),
-	}
-}
@@ -1,356 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"os"
-	"slices"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/contract/event"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/session/control"
-)
-
-// The four single-delegation arms. A lone task is not another executor: it
-// reaches the same RunProfileSpec a fleet item does and asks the same scheduler
-// for a slot. What these ask is whether that shared chain is wired to the
-// durable authority the fan-out arms established — and where, exactly, it stops
-// being wired, which is a different question from "does it have a graph".
-const (
-	// armTaskCompleted lets the delegation finish and the turn close. Nothing is
-	// interrupted, so a loss here is not about interruption: it says the run's
-	// provenance was never durable at all.
-	armTaskCompleted = "task-completed"
-	// armTaskRunning kills the process with the child mid-execution, the same
-	// death the fan-out arms take.
-	armTaskRunning = "task-running"
-	// armTaskFgQueued is the live-truth arm: the ceiling is full when the
-	// delegation asks, so the scheduler is holding it back. What the parent and
-	// the graph say while that is true is the whole question — a picture that
-	// says running while nothing has started is wrong before any restart.
-	armTaskFgQueued = "task-foreground-queued"
-	// armTaskBgQueued is the one the foreground arms cannot stand in for: a
-	// background run hands its execution to a job and asks for a slot inside it,
-	// so the parent tool call has already returned when the refusal happens.
-	armTaskBgQueued = "task-background-queued"
-	// armTaskBgRunning is the same handoff with the slot granted, which is where
-	// ownership of the terminal moves from the caller to the job.
-	armTaskBgRunning = "task-background-running"
-	// The cancellation arms. They ask the one question the five before them
-	// cannot: who owns the ending when work is stopped, on each side of the
-	// boundary that decides whether it ever ran. A child the scheduler never
-	// admitted produced nothing, so the store owes it no terminal at all.
-	armTaskFgQueuedCancel  = "task-foreground-queued-cancel"
-	armTaskBgQueuedCancel  = "task-background-queued-cancel"
-	armTaskFgRunningCancel = "task-foreground-running-cancel"
-	armTaskBgRunningCancel = "task-background-running-cancel"
-)
-
-func cancelTaskArm(name string) bool {
-	switch name {
-	case armTaskFgQueuedCancel, armTaskBgQueuedCancel, armTaskFgRunningCancel, armTaskBgRunningCancel:
-		return true
-	}
-	return false
-}
-
-// beforeStartArm names the arms cancelled while the scheduler still held the
-// delegation, which is the half of the boundary the store must stay out of.
-func beforeStartArm(name string) bool {
-	return name == armTaskFgQueuedCancel || name == armTaskBgQueuedCancel
-}
-
-func loneTaskArm(name string) bool {
-	switch name {
-	case armTaskCompleted, armTaskRunning, armTaskFgQueued, armTaskBgQueued, armTaskBgRunning:
-		return true
-	}
-	return cancelTaskArm(name)
-}
-
-// queuedTaskArm names the arms that fill the ceiling before delegating, so the
-// scheduler has to refuse the delegation admission.
-func queuedTaskArm(name string) bool {
-	return name == armTaskFgQueued || name == armTaskBgQueued || beforeStartArm(name)
-}
-
-// childEntered names the record a child leaves when it reaches the provider.
-// The queued arms ask a negative — that one never did — which nothing else in
-// the probe can answer.
-func childEntered(sentinel string) string { return "child-entered:" + sentinel }
-
-// childLeft names the record a child leaves when it returns. Entered without
-// left is a child still inside the provider — the state a cancellation that
-// never arrived leaves behind.
-func childLeft(sentinel string) string { return "child-left:" + sentinel }
-
-// childCtxDone names the record a child leaves when its own context closes,
-// which is a different fact from having returned: a run whose caller stopped
-// waiting returns to nobody with its context still open.
-func childCtxDone(sentinel string) string { return "child-ctxdone:" + sentinel }
-
-// probeChildEntered is where that answer rides into the observation. It is not
-// a status the kernel emits: the key names the probe so no row reads it as one.
-const probeChildEntered = "probe:delegated-child-entered"
-
-func backgroundTaskArm(name string) bool {
-	switch name {
-	case armTaskBgQueued, armTaskBgRunning, armTaskBgQueuedCancel, armTaskBgRunningCancel,
-		armLineageBackground, armLineageJobKill, armCancelBackground, armFleetActiveStore:
-		return true
-	}
-	return false
-}
-
-// The sentinels one turn carries. The queued arm names the holder as well: its
-// ceiling has to be occupied before the delegation asks for a slot, or the
-// refusal it is built to record never happens.
-const (
-	taskSentinel   = "PROBE-TASK"
-	taskCallID     = "probe_task"
-	taskHoldsFirst = fleetHolderSentinel + " " + taskSentinel
-)
-
-// loneTaskCall is the delegation itself, dispatched through the tool a model
-// calls. Nothing here reaches past the tool's own arguments: the probe measures
-// the production path, so it must enter it the way a model does.
-func loneTaskCall(arm string) []provider.Chunk {
-	prompt, description := childDone+" lone task", "completes"
-	switch {
-	case arm == armFleetActiveStore:
-		// This delegation is the other occupant of the ceiling its arm shares
-		// with a fan-out, and its own child is the only thing that can say the
-		// slot is taken. Holding is how it hangs and says so at once.
-		prompt, description = childHold+" lone task", "holds a slot to the death"
-	case arm != armTaskCompleted:
-		prompt, description = childHang+" lone task", "still executing at death"
-	}
-	args, _ := json.Marshal(map[string]any{
-		"prompt": prompt, "description": description,
-		"run_in_background": backgroundTaskArm(arm),
-	})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: taskCallID, Name: "task", Arguments: string(args)},
-	}}
-}
-
-// probeStepIDs are the host's open steps, in the order they have to be signed.
-// The arm whose turn ends has to close them: a delivery is refused while the
-// list is open, and a todo_write that marks an item completed on its own is
-// refused too — the sign-off is the only door.
-func probeStepIDs() []string { return []string{"probe_step_01", "probe_step_02", "probe_step_03"} }
-
-// signStep signs off one step per round, which is what the host accepts: the
-// sign-off promotes the next item, so two in one round sign against a list that
-// no longer stands.
-func signStep(id string) []provider.Chunk {
-	args, _ := json.Marshal(map[string]any{
-		"step_id": id,
-		"result":  "the probe drove this step to its end",
-		"evidence": []map[string]any{
-			{"kind": "manual", "summary": "the probe established this state directly"},
-		},
-	})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: "probe_sign_" + id, Name: "complete_step", Arguments: string(args)},
-	}}
-}
-
-// closesItsTurn names the arms whose turn has to reach its end. A turn that
-// called a tool cannot deliver while the host's step list is open, so these
-// sign it off one step per round — the sign-off promotes the next item, and two
-// in one round would sign against a list that no longer stands.
-func closesItsTurn(arm string) bool {
-	return arm == armTaskCompleted || arm == armSkillCompleted || arm == armReadOnlySkillCompleted
-}
-
-// loneTaskSentinels is what this arm's one prompt carries.
-func loneTaskSentinels(arm string) string {
-	if queuedTaskArm(arm) {
-		return taskHoldsFirst
-	}
-	return taskSentinel
-}
-
-// runLoneTaskConstruct drives one delegation and dies at the moment this arm is
-// about. The completed arm is the exception that proves the rest: it lets the
-// turn close, so what it loses was never about dying inside one.
-func runLoneTaskConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	// One turn, whatever the arm needs in it: a session admits no second one. The
-	// queued arm's ceiling is filled by a background fleet named in the same
-	// prompt, whose own child says when the slot is taken.
-	prompt := fanOutTurn(turn+1, loneTaskSentinels(arm))
-	switch {
-	case arm == armTaskCompleted:
-		if err := ctrl.Run(ctx, prompt); err != nil {
-			return fmt.Errorf("delegating turn: %w", err)
-		}
-	case guardedArm(arm):
-		// A stop belongs to whoever admitted the turn, so an arm that is going
-		// to be stopped is admitted the way a person's input is.
-		ctrl.Send(prompt)
-		if err := waitForLoneTask(sink, prov, arm); err != nil {
-			return err
-		}
-	default:
-		go func() { _ = ctrl.Run(ctx, prompt) }()
-		if err := waitForLoneTask(sink, prov, arm); err != nil {
-			return err
-		}
-	}
-	stop := ""
-	if cancelTaskArm(arm) {
-		stop = stopDelegation(ctrl, prov, arm)
-	}
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	if stop != "" {
-		obs.Progress[probeStopReached] = []string{stop}
-	}
-	if queuedTaskArm(arm) {
-		// Whether the delegation's own child ever ran is the fact the live rows
-		// are read against, and only the scripted provider knows it.
-		obs.Progress[probeChildEntered] = []string{fmt.Sprint(prov.fleets.held(childEntered(childHang)))}
-	}
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// stopDelegation stops the work the way a person does — the turn's own cancel
-// for a foreground run, the job list and a kill for a backgrounded one — and
-// waits for the orchestration to let go of it. What each layer holds then is
-// this arm's whole question, so nothing is read until the record closes.
-func stopDelegation(ctrl *control.Controller, prov *scripted, arm string) string {
-	stopped := time.Now()
-	if backgroundTaskArm(arm) {
-		for _, job := range ctrl.Jobs() {
-			ctrl.CancelJob(job.ID)
-		}
-	} else {
-		ctrl.Cancel()
-	}
-	// Wait for the record to close, but do not demand it: whether a stop reaches
-	// the work it was aimed at is what this arm measures, and an arm that failed
-	// when it did not would report the finding as its own premise breaking.
-	deadline := time.Now().Add(20 * time.Second)
-	for time.Now().Before(deadline) && !settledLone(ctrl) {
-		time.Sleep(50 * time.Millisecond)
-	}
-	live := "gone"
-	if prov.fleets.held(childEntered(childHang)) && !prov.fleets.held(childLeft(childHang)) {
-		live = "still running"
-	}
-	return fmt.Sprintf("turn=%s work=%s after=%s", turnStanding(ctrl, arm), live,
-		time.Since(stopped).Round(time.Second))
-}
-
-func settledLone(ctrl *control.Controller) bool {
-	for _, e := range control.ExecutionHistory(ctrl.SessionPath()) {
-		if e.ID == parallelLoneNodeID && !e.SettledAt.IsZero() {
-			return true
-		}
-	}
-	return false
-}
-
-// waitForLoneTask blocks until the delegation stands where this arm dies. The
-// foreground arms read the live graph; the background ones cannot — a handed-off
-// run publishes no node — so they read the progress the parent was told.
-func waitForLoneTask(sink *graphSink, prov *scripted, arm string) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		if loneTaskReady(sink, prov, arm) {
-			// A refusal is a state nothing announces, so the arm holds still and
-			// asks again: a child that was merely slow to start would have
-			// arrived by now, and the death would be measuring a race.
-			if beforeStartArm(arm) || arm == armTaskFgQueued {
-				time.Sleep(2 * time.Second)
-				return refusalHolding(prov, arm)
-			}
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected("a lone delegation "+loneTaskWanted(arm),
-		nodeStates(graph)+" progress="+fmt.Sprint(sink.phaseSeries()))
-}
-
-// refusalHolding reports the arm's premise once more, after the settle: the
-// delegation's child must still not have run.
-func refusalHolding(prov *scripted, arm string) error {
-	if prov.fleets.held(childEntered(childHang)) {
-		return errUnexpected("a delegation the scheduler was still holding back", "its child ran")
-	}
-	return nil
-}
-
-func loneTaskReady(sink *graphSink, prov *scripted, arm string) bool {
-	graph, _ := sink.snapshot()
-	switch arm {
-	case armTaskFgQueued, armTaskFgQueuedCancel:
-		// The node is drawn — the delegation reached the graph — and its child
-		// has not started, which with the ceiling full is the refusal.
-		return loneNodeDrawn(graph) && !prov.fleets.held(childEntered(childHang))
-	case armTaskRunning, armTaskFgRunningCancel, armTaskBgRunningCancel:
-		return holdsAll(graph, []agentgraph.NodeState{agentgraph.StateRunning})
-	case armTaskBgQueued, armTaskBgQueuedCancel:
-		phases := sink.phaseSeries()[taskCallID]
-		return len(phases) > 0 && phases[len(phases)-1] == string(subagentQueued)
-	case armTaskBgRunning:
-		return slices.Contains(sink.phaseSeries()[taskCallID], subagentRunning)
-	}
-	return false
-}
-
-// loneNodeDrawn reports whether the delegation's own node is on the graph.
-func loneNodeDrawn(g agentgraph.Graph) bool {
-	for _, n := range g.Nodes {
-		if n.Kind == agentgraph.KindWorker && n.ID == parallelLoneNodeID {
-			return true
-		}
-	}
-	return false
-}
-
-// parallelLoneNodeID is what a lone delegation's node is called: the parent
-// call, and the first slot of the fan-out numbering it shares.
-const parallelLoneNodeID = taskCallID + "/sub-1"
-
-func loneTaskWanted(arm string) string {
-	switch arm {
-	case armTaskRunning, armTaskFgRunningCancel, armTaskBgRunningCancel:
-		return "running"
-	case armTaskFgQueued, armTaskFgQueuedCancel, armTaskBgQueuedCancel:
-		return "drawn while the scheduler holds it back"
-	case armTaskBgQueued:
-		return "queued for a slot and not started"
-	}
-	return "running in its job"
-}
-
-// The two phases the probe waits on. They are the parent-facing statuses a
-// background delegation emits, and for a handed-off run they are the only live
-// evidence there is.
-const (
-	subagentQueued  = "queued"
-	subagentRunning = "running"
-)
-
-// progressPhase reads one subagent status event. The tool id is the parent call
-// the delegation was made from, which is also what the graph node hangs under
-// and what the sub-agent store records as its parent — the join this arm exists
-// to check rather than assume.
-func progressPhase(e event.Event) (id, phase string, ok bool) {
-	if e.Kind != event.ToolProgress || e.Tool.Name != event.SubagentProgressStatusName {
-		return "", "", false
-	}
-	return e.Tool.ID, e.Tool.Output, e.Tool.ID != ""
-}
@@ -1,536 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"sort"
-	"strings"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/state/execgraph"
-	"reasonix/internal/state/execjournal"
-)
-
-// What a lone delegation leaves behind, layer by layer. The columns are the
-// ordinary ones: what the dying process held, and what the next one can still
-// say. Nothing here is a fix — the arm exists to find which transition stops
-// being recorded, because "single task has no graph" was the roadmap's guess
-// and the live chain has since been unified.
-func loneTaskRows(arm string, before, after Observation) []row {
-	return []row{
-		liveDelegationRow(before),
-		parentStatusRow(before),
-		durableOpeningRow(before, after),
-		durableQueueRow(before, after),
-		durableStartRow(before, after),
-		durableSettleRow(before, after),
-		delegationStoreRow(before, after),
-		rebuiltDelegationRow(before, after),
-		rebuiltDelegationIdentityRow(after),
-		delegationInterruptionRow(before, after),
-		identityJoinRow(before),
-		heldBackRow(arm, before),
-		stoppedOwnershipRow(arm, before),
-		startedTerminalRow(before, after),
-		fanOutControlRow(before),
-	}
-}
-
-// ofLoneTask reports whether an id belongs to the delegation under measurement.
-// The queued arm fills its ceiling with a fan-out, and a fan-out records
-// everything — so without this the rows would read that record as the lone
-// task's own and report a gap as closed.
-func ofLoneTask(id string) bool {
-	return id == taskCallID || strings.HasPrefix(id, taskCallID+"/")
-}
-
-const (
-	crossAuthority   = "the journal and the store, read against each other"
-	liveAuthority    = "event sink (GraphDelta), in-process"
-	journalAuthority = "execjournal, written by the orchestration"
-	storeAuthority   = "sub-agent store"
-	rebuildAuthority = "execgraph.Rebuild over journal + store"
-)
-
-// liveDelegationRow is what the dying process actually drew. A backgrounded run
-// draws nothing at all — openDelegationGraph returns early for it — so this row
-// separates "the graph was lost" from "there was never a node".
-func liveDelegationRow(before Observation) row {
-	var drawn []string
-	for _, n := range before.Graph.Nodes {
-		if n.Kind == agentgraph.KindWorker && ofLoneTask(n.ID) {
-			drawn = append(drawn, n.ID+":"+string(n.State))
-		}
-	}
-	sort.Strings(drawn)
-	return row{
-		Semantic: "the delegation the live graph drew", Authority: liveAuthority,
-		Artifact: "none observed", Reconstruction: "in-process only",
-		Before: orNone(join(drawn)), After: "—", Verdict: liveVerdict(len(drawn) > 0),
-	}
-}
-
-func liveVerdict(drawn bool) string {
-	if drawn {
-		return verdictStable
-	}
-	return verdictNotMeasured
-}
-
-// parentStatusRow is the other live surface, and the only one a handed-off run
-// has: the statuses the parent tool call was told.
-func parentStatusRow(before Observation) row {
-	var told []string
-	for id, phases := range before.Progress {
-		if ofLoneTask(id) {
-			told = append(told, id+":"+strings.Join(phases, "→"))
-		}
-	}
-	sort.Strings(told)
-	return row{
-		Semantic: "what the parent call was told", Authority: "subagent progress (ToolProgress)",
-		Artifact: "none observed", Reconstruction: "in-process only",
-		Before: orNone(join(told)), After: "—", Verdict: liveVerdict(len(told) > 0),
-	}
-}
-
-func durableOpeningRow(before, after Observation) row {
-	return valueRow("durable opening", journalAuthority, "<stem>.execution.jsonl",
-		"read back by ExecutionHistory", executionIDs(before), executionIDs(after), true)
-}
-
-func durableQueueRow(before, after Observation) row {
-	return valueRow("durable queue refusal", journalAuthority, "<stem>.execution.jsonl",
-		"the scheduler's first refusal, with its cause",
-		queuedCauses(before), queuedCauses(after), true)
-}
-
-func durableStartRow(before, after Observation) row {
-	return valueRow("durable slot grant", journalAuthority, "<stem>.execution.jsonl",
-		"written before the child can act", startedIDs(before), startedIDs(after), true)
-}
-
-func durableSettleRow(before, after Observation) row {
-	return valueRow("durable settlement", journalAuthority, "<stem>.execution.jsonl",
-		"the orchestration letting go of the execution",
-		settledIDs(before), settledIDs(after), true)
-}
-
-// delegationStoreRow is the layer that already survives for a fan-out item. It
-// is the reason this arm cannot be read off the graph alone: a store record with
-// no opening leaves the run's grant, worker and scheduling unaccounted for.
-func delegationStoreRow(before, after Observation) row {
-	return valueRow("child the store kept", storeAuthority, "subagents/<ref>.meta.json",
-		"ListSubagentsByParent over the transcript stem",
-		childFacts(before), childFacts(after), true)
-}
-
-func rebuiltDelegationRow(before, after Observation) row {
-	rebuilt := rebuiltGraph(after)
-	var nodes []string
-	for _, n := range rebuilt.Graph.Nodes {
-		if n.Kind != agentgraph.KindGroup && ofLoneTask(n.ID) {
-			nodes = append(nodes, n.ID+":"+string(n.State))
-		}
-	}
-	sort.Strings(nodes)
-	return valueRow("the node a restart rebuilds", rebuildAuthority, "<stem>.execution.jsonl",
-		"execgraph.Rebuild", loneDelegationAtDeath(before), join(nodes), false)
-}
-
-// rebuiltDelegationIdentityRow is what a reader sees on the rebuilt node beyond
-// its state. Grant and worker are opening facts: the store cannot answer them,
-// so an arm that only compared states would call a hollow node exact.
-func rebuiltDelegationIdentityRow(after Observation) row {
-	var got []string
-	for _, n := range rebuiltGraph(after).Graph.Nodes {
-		if n.Kind == agentgraph.KindGroup || !ofLoneTask(n.ID) {
-			continue
-		}
-		got = append(got, fmt.Sprintf("%s grant=%s model=%s effort=%s wait=%s",
-			n.ID, orDash(string(n.Grant)), orDash(n.Model), orDash(n.Effort), orDash(string(n.Wait))))
-	}
-	sort.Strings(got)
-	return row{
-		Semantic: "grant, worker and wait on the rebuilt node", Authority: rebuildAuthority,
-		Artifact: "<stem>.execution.jsonl", Reconstruction: "opening facts, which no store record carries",
-		Before: "declared at the opening", After: orNone(join(got)),
-		Verdict: rebuiltIdentityVerdict(got),
-	}
-}
-
-func rebuiltIdentityVerdict(got []string) string {
-	if len(got) == 0 {
-		return verdictLost
-	}
-	return verdictExact
-}
-
-func delegationInterruptionRow(before, after Observation) row {
-	var got []string
-	for _, i := range rebuiltGraph(after).Interrupted {
-		if ofLoneTask(i.Execution) {
-			got = append(got, i.Execution+":"+interruptionKind(i))
-		}
-	}
-	sort.Strings(got)
-	return valueRow("the interruption a restart names", rebuildAuthority, "<stem>.execution.jsonl",
-		"an execution with no owner is named, never redrawn as running",
-		loneDelegationAtDeath(before), join(got), false)
-}
-
-// loneDelegationAtDeath is the delegation as the dying process last knew it,
-// from whichever surface knew: a foreground run has a node, a handed-off one
-// has only the store's record, and an arm whose rows read the wrong surface
-// would report a loss the other one never had.
-func loneDelegationAtDeath(o Observation) string {
-	if live := liveWorkers(o); live != "" {
-		return live
-	}
-	return childFacts(o)
-}
-
-func interruptionKind(i execgraph.Interruption) string {
-	if i.Started {
-		return execjournal.InterruptedDuringExecution
-	}
-	return execjournal.InterruptedBeforeStart
-}
-
-// identityJoinRow is the question that decides whether Step 8 needs an identity
-// at all: the live node, the call it hung under, the parent the store recorded,
-// and the execution the journal opened. Four spellings, or one identity.
-func identityJoinRow(before Observation) row {
-	live, call, stored, opened := identitiesAtDeath(before)
-	verdict := verdictViolated
-	switch {
-	case stored != "" && (stored == opened || stored == call || strings.HasPrefix(live, stored)):
-		verdict = verdictHolds
-	case stored == "" && !startedAtDeath(before):
-		// Nothing executed, so the store owes no record and there is no join to
-		// make. A missing artifact here is the arm's own premise, not a defect.
-		verdict = verdictNotMeasured
-	}
-	return row{
-		Semantic: "one identity or four", Authority: "the graph, the call, the store and the journal",
-		Artifact:       "subagents/<ref>.meta.json, <stem>.execution.jsonl",
-		Reconstruction: "the store's parent call is what a rebuild joins on",
-		Before:         fmt.Sprintf("node=%s call=%s", orDash(live), orDash(call)),
-		After:          fmt.Sprintf("store parent=%s journal=%s", orDash(stored), orDash(opened)),
-		Verdict:        verdict,
-	}
-}
-
-// identitiesAtDeath reads the four spellings out of the dying process's own
-// observation, so a join that only holds after a restart cannot be claimed.
-func identitiesAtDeath(o Observation) (live, call, stored, opened string) {
-	for _, n := range o.Graph.Nodes {
-		if n.Kind == agentgraph.KindWorker && ofLoneTask(n.ID) {
-			live, call = n.ID, n.ParentID
-			break
-		}
-	}
-	for id := range o.Progress {
-		if call == "" && ofLoneTask(id) {
-			call = id
-		}
-	}
-	for _, f := range o.Children.Facts {
-		if ofLoneTask(f.ParentToolCallID) {
-			stored = f.ParentToolCallID
-			break
-		}
-	}
-	for _, e := range o.Executions {
-		if ofLoneTask(e.ID) {
-			opened = e.ID
-			break
-		}
-	}
-	return live, call, stored, opened
-}
-
-// startedAtDeath reports whether the delegation was ever granted a slot.
-func startedAtDeath(o Observation) bool {
-	for _, e := range o.Executions {
-		if ofLoneTask(e.ID) && e.Started() {
-			return true
-		}
-	}
-	return false
-}
-
-func executionIDs(o Observation) string {
-	var out []string
-	for _, e := range o.Executions {
-		if ofLoneTask(e.ID) {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func queuedCauses(o Observation) string {
-	var out []string
-	for _, e := range o.Executions {
-		if e.Queued() && ofLoneTask(e.ID) {
-			out = append(out, e.ID+":"+orDash(e.Cause))
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func startedIDs(o Observation) string {
-	var out []string
-	for _, e := range o.Executions {
-		if e.Started() && ofLoneTask(e.ID) {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func settledIDs(o Observation) string {
-	var out []string
-	for _, e := range o.Executions {
-		if !e.SettledAt.IsZero() && ofLoneTask(e.ID) {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func childFacts(o Observation) string {
-	var out []string
-	for _, f := range o.Children.Facts {
-		if ofLoneTask(f.ParentToolCallID) {
-			out = append(out, f.ParentToolCallID+":"+f.Status)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func liveWorkers(o Observation) string {
-	var out []string
-	for _, n := range o.Graph.Nodes {
-		if n.Kind == agentgraph.KindWorker && ofLoneTask(n.ID) {
-			out = append(out, n.ID+":"+string(n.State))
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func orDash(s string) string {
-	if strings.TrimSpace(s) == "" {
-		return "-"
-	}
-	return s
-}
-
-// heldBackRow is the live-truth row: with the ceiling full and the delegation's
-// own child never reaching the provider, the scheduler is holding it back — and
-// what the picture says while that is true is a claim about now, which no
-// restart can correct. Running is the answer that is wrong.
-func heldBackRow(arm string, before Observation) row {
-	if !queuedTaskArm(arm) {
-		return row{
-			Semantic: "what the picture says while the scheduler holds it back", Authority: liveAuthority,
-			Artifact: "none observed", Reconstruction: "this arm never fills the ceiling",
-			Before: "—", After: "—", Verdict: verdictNotMeasured,
-		}
-	}
-	drawn := orNone(liveWorkers(before))
-	told := orNone(loneStatus(before))
-	verdict := verdictHolds
-	if strings.Contains(drawn, ":running") || told == subagentRunning {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: "what the picture says while the scheduler holds it back", Authority: liveAuthority,
-		Artifact:       "none observed",
-		Reconstruction: "queued with its cause, or at least not started",
-		Before:         "held back: ceiling full, child never entered",
-		After:          "graph=" + drawn + " parent=" + told, Verdict: verdict,
-	}
-}
-
-func loneStatus(o Observation) string {
-	for id, phases := range o.Progress {
-		if ofLoneTask(id) && len(phases) > 0 {
-			return phases[len(phases)-1]
-		}
-	}
-	return ""
-}
-
-// stoppedOwnershipRow is the invariant a stop has to leave behind: once the
-// turn is over, nothing it owned may still be open with a live owner. A record
-// left open is what a restart reads as an interruption, so a turn that merely
-// stopped waiting looks, afterwards, exactly like a process that died.
-func stoppedOwnershipRow(arm string, before Observation) row {
-	if !cancelTaskArm(arm) {
-		return row{
-			Semantic: "work the stopped turn still owns", Authority: crossAuthority,
-			Artifact: "<stem>.execution.jsonl", Reconstruction: "this arm stops nothing",
-			Before: "—", After: "—", Verdict: verdictNotMeasured,
-		}
-	}
-	reading := ""
-	if seen := before.Progress[probeStopReached]; len(seen) > 0 {
-		reading = seen[0]
-	}
-	entry, _ := loneEntry(before)
-	open := entry.SettledAt.IsZero()
-	turnOver := strings.Contains(reading, "turn=ended")
-	// An open record implies a live owner: nothing else closes one. The child's
-	// liveness rides in the reading as evidence, not as the test — a delegation
-	// still waiting for a slot has no child and is just as much still owned.
-	return row{
-		Semantic: "work the stopped turn still owns", Authority: crossAuthority,
-		Artifact:       "<stem>.execution.jsonl",
-		Reconstruction: "a turn that has ended owns nothing still open",
-		Before:         "the turn owned it when the stop was issued",
-		After:          orNone(reading) + " record=" + entryStanding(entry), Verdict: held(!(turnOver && open)),
-	}
-}
-
-func entryStanding(e execjournal.Entry) string {
-	if e.SettledAt.IsZero() {
-		return "still open"
-	}
-	return "settled"
-}
-
-// startedTerminalRow refuses the combination neither layer may produce: a child
-// record for work the journal proves was never admitted. A state comparison
-// shows that only as "cancelled came back failed"; this names the cause, which
-// is a store asked to speak for an execution that never happened.
-func startedTerminalRow(before, after Observation) row {
-	if startedAtDeath(before) {
-		return row{
-			Semantic: "a child record for work that never started", Authority: crossAuthority,
-			Artifact:       "subagents/<ref>.meta.json, <stem>.execution.jsonl",
-			Reconstruction: "this delegation was granted a slot, so the store owes a record",
-			Before:         "—", After: "—", Verdict: verdictNotMeasured,
-		}
-	}
-	kept := orNone(join(compact(childFacts(before), childFacts(after))))
-	return row{
-		Semantic: "a child record for work that never started", Authority: crossAuthority,
-		Artifact:       "subagents/<ref>.meta.json, <stem>.execution.jsonl",
-		Reconstruction: "no slot was ever granted, so nothing ran and nothing owes a terminal",
-		Before:         "none", After: kept, Verdict: held(kept == "none"),
-	}
-}
-
-func compact(parts ...string) []string {
-	var out []string
-	for _, p := range parts {
-		if strings.TrimSpace(p) != "" {
-			out = append(out, p)
-		}
-	}
-	return out
-}
-
-// fanOutControlRow is the positive control the queued arm carries with it: the
-// fleet that fills its ceiling runs through the same scheduler in the same
-// session, and records everything. What this arm reports missing is missing at
-// the entry point, not in the machinery.
-func fanOutControlRow(before Observation) row {
-	var recorded []string
-	for _, e := range before.Executions {
-		if !ofLoneTask(e.ID) {
-			recorded = append(recorded, e.ID+lifecycleOf(e))
-		}
-	}
-	sort.Strings(recorded)
-	verdict := verdictNotMeasured
-	if len(recorded) > 0 {
-		verdict = verdictPersisted
-	}
-	return row{
-		Semantic: "a fan-out in the same session, same scheduler", Authority: journalAuthority,
-		Artifact: "<stem>.execution.jsonl", Reconstruction: "control: the machinery this arm reports missing",
-		Before: orNone(join(recorded)), After: "—", Verdict: verdict,
-	}
-}
-
-func lifecycleOf(e execjournal.Entry) string {
-	marks := ""
-	if e.Queued() {
-		marks += " queued=" + orDash(e.Cause)
-	}
-	if e.Started() {
-		marks += " started"
-	}
-	if !e.SettledAt.IsZero() {
-		marks += " settled"
-	}
-	return marks
-}
-
-// loneTaskArmInvalid guards each arm's premise: the state it is built to die in
-// has to have been reached, or its rows describe a different death.
-func loneTaskArmInvalid(arm string, before Observation) string {
-	phases := strings.Join(before.Progress[taskCallID], "→")
-	switch arm {
-	case armTaskCompleted:
-		if !strings.Contains(phases, "completed") && len(before.Children.Facts) == 0 {
-			return "the delegation never completed, so this arm measures a different death"
-		}
-	case armTaskRunning:
-		if !holdsState(before, string(agentgraph.StateRunning)) {
-			return "nothing was executing when the process died"
-		}
-	case armTaskFgQueued:
-		if entered := before.Progress[probeChildEntered]; len(entered) == 0 || entered[0] != "false" {
-			return "the delegation's child ran, so the scheduler was not holding it back"
-		}
-		if !strings.Contains(fanOutControlRow(before).Before, "queued=slots") {
-			return "no fan-out filled the ceiling, so nothing refused the delegation"
-		}
-	case armTaskBgQueued:
-		if !strings.HasSuffix(phases, subagentQueued) {
-			return "the delegation was not left queued for a slot: " + orNone(phases)
-		}
-	case armTaskBgRunning:
-		if !strings.Contains(phases, subagentRunning) {
-			return "the delegation never entered its job: " + orNone(phases)
-		}
-	}
-	if cancelTaskArm(arm) {
-		return stoppedArmInvalid(arm, before)
-	}
-	return ""
-}
-
-// stoppedArmInvalid holds a cancellation arm to the side of the boundary it was
-// built for. An arm that stopped the work somewhere else measures a different
-// ownership question and must say so rather than reporting the answer.
-func stoppedArmInvalid(arm string, before Observation) string {
-	entry, ok := loneEntry(before)
-	switch {
-	case !ok:
-		return "the delegation was never recorded, so there is no ownership to read"
-	case beforeStartArm(arm) && entry.Started():
-		return "the delegation was granted a slot before the stop, so this is not a pre-start cancellation"
-	case !beforeStartArm(arm) && !entry.Started():
-		return "the delegation never held a slot, so this is not a cancellation of work that ran"
-	}
-	return ""
-}
-
-func loneEntry(o Observation) (execjournal.Entry, bool) {
-	for _, e := range o.Executions {
-		if ofLoneTask(e.ID) {
-			return e, true
-		}
-	}
-	return execjournal.Entry{}, false
-}
@@ -1,48 +0,0 @@
-// Process-boundary resurrection probe. It answers one question and modifies
-// nothing: after the process that built a session exits, what can a new process
-// still prove about that session's runtime semantics without asking a model?
-package main
-
-import (
-	"flag"
-	"fmt"
-	"os"
-)
-
-func main() {
-	phase := flag.String("phase", "", "internal: construct|successor|resume, spawned by the orchestrator")
-	root := flag.String("root", "", "internal: this arm's root directory")
-	armName := flag.String("arm", "", "internal: this arm's name")
-	work := flag.String("work", "", "directory to hold arm roots; default is a temp dir")
-	only := flag.String("only", "", "run a single arm by name")
-	jsonOut := flag.String("json", "", "also write the machine-readable matrix here")
-	keep := flag.Bool("keep", false, "keep arm roots on disk after the run")
-	sabotage := flag.String("sabotage", "", "run under a deliberate defect the arm must fail on (ui-graph-mixed: publish|trajectory)")
-	flag.Parse()
-	if *sabotage != "" {
-		// Carried to the children through the environment they already inherit,
-		// so no phase has to learn about a flag it does not read.
-		if err := os.Setenv(envProbeSabotage, *sabotage); err != nil {
-			fmt.Fprintln(os.Stderr, "runtime-resume:", err)
-			os.Exit(1)
-		}
-	}
-
-	var err error
-	switch *phase {
-	case "construct":
-		err = runConstruct(*root, *armName)
-	case "successor":
-		err = runSuccessor(*root, *armName)
-	case "resume":
-		err = runResume(*root, *armName)
-	case "":
-		err = orchestrate(*work, *only, *jsonOut, *keep)
-	default:
-		err = fmt.Errorf("unknown phase %q", *phase)
-	}
-	if err != nil {
-		fmt.Fprintln(os.Stderr, "runtime-resume:", err)
-		os.Exit(1)
-	}
-}
@@ -1,584 +0,0 @@
-package main
-
-import (
-	"crypto/sha256"
-	"encoding/hex"
-	"encoding/json"
-	"os"
-	"path/filepath"
-	"reasonix/internal/state/sessionstore"
-	"regexp"
-	"sort"
-	"strings"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/runtime/delegation"
-	"reasonix/internal/safety/evidence"
-	"reasonix/internal/session/control"
-	"reasonix/internal/state/execjournal"
-	"reasonix/internal/state/store"
-)
-
-// Observation is everything one phase can read about a session without asking a
-// model. Both phases capture the same shape so the matrix is a comparison and
-// not two differently-shaped reports.
-type Observation struct {
-	Phase       string `json:"phase"`
-	Arm         string `json:"arm"`
-	SessionPath string `json:"session_path"`
-	SystemHash  string `json:"system_hash"`
-	SystemBytes int    `json:"system_bytes"`
-	// BootSystemHash is what this process's assembly composed, read before any
-	// resume; SystemHash is the session's own leading row. Different questions.
-	BootSystemHash  string              `json:"boot_system_hash"`
-	BootSystemBytes int                 `json:"boot_system_bytes"`
-	Transcript      TranscriptObs       `json:"transcript"`
-	Goal            GoalObs             `json:"goal"`
-	Todos           []evidence.TodoItem `json:"todos"`
-	Decisions       []control.Decision  `json:"decisions"`
-	Context         ContextObs          `json:"context"`
-	Sidecar         SidecarObs          `json:"sidecar"`
-	View            ViewObs             `json:"view"`
-	TodoNotes       TodoNoteObs         `json:"todo_notes"`
-	Deferred        DeferredObs         `json:"deferred"`
-	Obligation      ObligationObs       `json:"obligation"`
-	// Interrupted is what the host derives from a barrier it found open and did
-	// not open itself: evidence the wait happened, not a question to answer.
-	Interrupted []control.InterruptedAdjudication `json:"interrupted"`
-	// ModelSeesInterruption counts the interruption blocks the request carries,
-	// which is a different surface from the derived fact behind them.
-	ModelSeesInterruption int `json:"model_sees_interruption"`
-	// Journal is how each barrier ended. A superseded edge is the durable
-	// evidence that a turn received the interruption: nothing else writes one.
-	Journal []control.AdjudicationEntry `json:"journal,omitempty"`
-	Graph   GraphObs                    `json:"graph"`
-	// Children is the durable side of the same fan-out the graph draws. The
-	// graph has no read surface after a restart, so without this a lost graph
-	// and a lost delegation would be one indistinguishable row.
-	Children ChildrenObs `json:"children"`
-	// Listing is the same children through the read list_subagents performs.
-	// It filters by workspace and walks lineage, so it can answer differently
-	// from the one above — and it is the answer a model asking would be shown.
-	Listing ListingObs `json:"listing"`
-	// FanOutTurn is whether the dispatching turn is in this phase's transcript
-	// at all. A turn is appended when it ends, so a process that dies inside one
-	// leaves no record of the request, and a lost graph is the smaller half.
-	FanOutTurn bool `json:"fan_out_turn"`
-	// Executions is the durable record of delegations this turn opened, and
-	// InterruptedExecutions the host's judgement over it: open, with no owner
-	// here. Written before the work starts, so it outlives an unfinished turn.
-	Executions            []execjournal.Entry `json:"executions,omitempty"`
-	InterruptedExecutions []execjournal.Entry `json:"interrupted_executions,omitempty"`
-	// ModelSeesInterruptedExecution counts the blocks the next request carries
-	// about them, which is a different surface from the fact behind them.
-	ModelSeesInterruptedExecution int `json:"model_sees_interrupted_execution"`
-	// WaitCauses counts the refusals the graph carries, and WaitSeries keeps
-	// each node's causes in publication order. The fold keeps only the latest,
-	// which cannot say whether a cause was reported once or replaced.
-	WaitCauses map[string]int      `json:"wait_causes,omitempty"`
-	WaitSeries map[string][]string `json:"wait_series,omitempty"`
-	// Progress is every parent-facing subagent status, per call. A background
-	// delegation draws no node, so this is where its live evidence is.
-	Progress map[string][]string `json:"progress,omitempty"`
-	// SettledWorkers and QueuedWorkers are the transition arm's evidence that
-	// capacity was actually freed while a refusal still stood.
-	SettledWorkers int           `json:"settled_workers"`
-	QueuedWorkers  []string      `json:"queued_workers,omitempty"`
-	Artifacts      []ArtifactObs `json:"artifacts"`
-	// UI is what a frontend attached to this process saw, in order. Only the one
-	// arm that runs a client fills it; a value comparison cannot answer what it
-	// asks, which is through which door each fact arrived.
-	UI *UIObs `json:"ui,omitempty"`
-}
-
-type TranscriptObs struct {
-	Messages int    `json:"messages"`
-	Digest   string `json:"digest"`
-}
-
-type GoalObs struct {
-	Goal      string `json:"goal"`
-	Status    string `json:"status"`
-	TurnsUsed int    `json:"turns_used"`
-}
-
-// ContextObs is the live judgement: CheckpointState is "none" whenever the
-// projection did not survive validation, which is what the compaction arms are
-// actually asking about.
-type ContextObs struct {
-	ProjectionVersion uint64 `json:"projection_version"`
-	CheckpointState   string `json:"checkpoint_state"`
-	CanonicalTokens   int    `json:"canonical_tokens"`
-	ProjectedTokens   int    `json:"projected_tokens"`
-	Blocked           bool   `json:"blocked"`
-}
-
-// SidecarObs is the stored claim, read off disk rather than through the agent,
-// so a projection the host refuses to use is still visible as present.
-type SidecarObs struct {
-	Present           bool   `json:"present"`
-	Messages          int    `json:"messages"`
-	CoveredCount      int    `json:"covered_count"`
-	CoveredPrefixHash string `json:"covered_prefix_hash"`
-	ProjectionVersion uint64 `json:"projection_version"`
-	TranscriptVersion uint64 `json:"transcript_version"`
-	Generation        uint64 `json:"generation"`
-	Err               string `json:"err,omitempty"`
-}
-
-// ViewObs is the model-visible context, spliced the way ContextProjection
-// declares it: the stored body followed by canonical[CoveredCount:]. Markers
-// are what a fold can be held to — the scripted digest carries none, so a
-// marker that is gone was dropped, not summarised.
-type ViewObs struct {
-	Messages int `json:"messages"`
-	// Markers are the user turns still shown, Echoes the assistant replies. A
-	// fold takes the reply first, so the two answer different questions.
-	Markers         []string `json:"markers"`
-	Echoes          []string `json:"echoes"`
-	BodyMarkers     []string `json:"body_markers"`
-	BodyEchoes      []string `json:"body_echoes"`
-	SplicedFromTail int      `json:"spliced_from_tail"`
-}
-
-// TodoNoteObs locates the host's step-identity note inside the model-visible
-// view. BodyLen is where the frozen body ends, so an index below it says the
-// note was frozen ahead of the live tail rather than riding it.
-type TodoNoteObs struct {
-	Count   int      `json:"count"`
-	Indexes []int    `json:"indexes"`
-	IDs     []string `json:"ids"`
-	BodyLen int      `json:"body_len"`
-	ViewLen int      `json:"view_len"`
-	HostIDs []string `json:"host_ids"`
-	// ReadableInHistory says the conversation still shows every current id, so
-	// no note is owed. Without it a count of zero cannot be told apart from a
-	// lost identity, and the gate would demand a note nobody needs.
-	ReadableInHistory bool `json:"readable_in_history"`
-}
-
-// todoNoteMarker is the opening of todoIdentityNote. The probe reads the view
-// the way a model would, so it matches the text rather than a type.
-const todoNoteMarker = "Host task state."
-
-var stepIDPattern = regexp.MustCompile(`probe_step_\d+`)
-
-func todoNoteObs(view, history []provider.Message, todos []evidence.TodoItem) TodoNoteObs {
-	out := TodoNoteObs{BodyLen: len(history), ViewLen: len(view)}
-	for _, t := range todos {
-		if t.StepID != "" {
-			out.HostIDs = append(out.HostIDs, t.StepID)
-		}
-	}
-	out.ReadableInHistory = len(out.HostIDs) > 0
-	for _, id := range out.HostIDs {
-		if !mentionsID(history, id) {
-			out.ReadableInHistory = false
-		}
-	}
-	for i, m := range view {
-		if !strings.Contains(m.Content, todoNoteMarker) {
-			continue
-		}
-		out.Count++
-		out.Indexes = append(out.Indexes, i)
-		out.IDs = append(out.IDs, stepIDPattern.FindAllString(m.Content, -1)...)
-	}
-	return out
-}
-
-// DeferredObs is the effect a pending decision is holding back. Whether the
-// decision itself survives is one question; whether what it blocked stayed
-// blocked is the one that cannot be answered by a design choice.
-type DeferredObs struct {
-	MarkerPath string `json:"marker_path"`
-	Executed   bool   `json:"executed"`
-}
-
-func deferredObs(workspace string) DeferredObs {
-	path := filepath.Join(workspace, deferredEffect)
-	_, err := os.Stat(path)
-	return DeferredObs{MarkerPath: deferredEffect, Executed: err == nil}
-}
-
-// ObligationObs is what the transcript itself still says about a turn that
-// never finished: a tool call with no result is a claim the host made and did
-// not settle, and an interruption record is the host saying so outright. Either
-// tells a lost decision apart from a silently dropped one.
-type ObligationObs struct {
-	UnansweredCalls    []string `json:"unanswered_calls"`
-	InterruptionMarked int      `json:"interruption_marked"`
-}
-
-func obligationObs(msgs []provider.Message) ObligationObs {
-	answered := map[string]bool{}
-	var out ObligationObs
-	for _, m := range msgs {
-		if m.ToolCallID != "" {
-			answered[m.ToolCallID] = true
-		}
-		if m.InterruptedTurn != nil {
-			out.InterruptionMarked++
-		}
-	}
-	for _, m := range msgs {
-		for _, call := range m.ToolCalls {
-			if call.ID != "" && call.ID != provider.LocalOnlyToolID && !answered[call.ID] {
-				out.UnansweredCalls = append(out.UnansweredCalls, call.Name+":"+call.ID)
-			}
-		}
-	}
-	return out
-}
-
-type GraphObs struct {
-	Deltas int               `json:"deltas"`
-	Nodes  []agentgraph.Node `json:"nodes,omitempty"`
-	Edges  []agentgraph.Edge `json:"edges,omitempty"`
-	Grants map[string]string `json:"grants,omitempty"`
-	Waits  map[string]string `json:"waits,omitempty"`
-}
-
-// ChildFact is one delegated run as the durable store records it, which is a
-// different set of facts from the node the graph drew for it. Whatever the
-// graph carried and this does not is a semantic no reconstruction can recover.
-type ChildFact struct {
-	Ref    string `json:"ref"`
-	Status string `json:"status"`
-	Kind   string `json:"kind,omitempty"`
-	Name   string `json:"name,omitempty"`
-	// ExecutionID is which execution the record belongs to; ParentToolCallID is
-	// the provider-visible call it descends from. They are separate answers, and
-	// a host-started run has only the first.
-	ExecutionID      string `json:"executionId,omitempty"`
-	ParentToolCallID string `json:"parentToolCallId,omitempty"`
-	Model            string `json:"model,omitempty"`
-	Effort           string `json:"effort,omitempty"`
-}
-
-// ChildrenObs is every child the store still owns for this parent. It is the
-// only durable execution evidence a restart has been shown to read, so a graph
-// classification that ignores it would report a loss the host can in fact
-// still speak to.
-type ChildrenObs struct {
-	Facts []ChildFact `json:"facts,omitempty"`
-	Err   string      `json:"err,omitempty"`
-}
-
-// childrenObs reads the store the way a restart does: by parent session id,
-// which is the transcript's stem. A session with no path owns nothing.
-func childrenObs(root armRoot, sessionPath string) ChildrenObs {
-	if strings.TrimSpace(sessionPath) == "" {
-		return ChildrenObs{}
-	}
-	parent := strings.TrimSuffix(filepath.Base(sessionPath), ".jsonl")
-	artifacts, err := sessionstore.ListSubagentsByParent(root.Sessions, parent)
-	if err != nil {
-		return ChildrenObs{Err: err.Error()}
-	}
-	out := ChildrenObs{}
-	for _, a := range artifacts {
-		out.Facts = append(out.Facts, ChildFact{
-			Ref: a.Ref, Status: string(a.Meta.Status), Kind: a.Meta.Kind, Name: a.Meta.Name,
-			ExecutionID:      a.Meta.ExecutionID,
-			ParentToolCallID: a.Meta.ParentToolCallID, Model: a.Meta.Model, Effort: a.Meta.Effort,
-		})
-	}
-	sort.Slice(out.Facts, func(i, j int) bool { return out.Facts[i].Ref < out.Facts[j].Ref })
-	return out
-}
-
-// ListingRow is one row list_subagents would print. The tool call itself needs
-// a registry this probe has no way to build, so the read behind it is what is
-// exercised; the formatting adds no fact the row does not already carry.
-type ListingRow struct {
-	Ref    string `json:"ref"`
-	Status string `json:"status"`
-	From   string `json:"from,omitempty"`
-}
-
-// ListingObs is what the recovery surface would show for this conversation.
-type ListingObs struct {
-	Rows []ListingRow `json:"rows,omitempty"`
-	Err  string       `json:"err,omitempty"`
-}
-
-// listingObs reads the store the way the tool does: by the parent session the
-// child recorded, filtered to this workspace. The workspace filter is why this
-// cannot be folded into childrenObs — a root that does not match drops every
-// row, and a reader must be able to see that happen rather than read it as an
-// empty store.
-func listingObs(root armRoot, sessionPath string) ListingObs {
-	if strings.TrimSpace(sessionPath) == "" {
-		return ListingObs{}
-	}
-	store := delegation.NewSubagentStore(filepath.Join(root.Sessions, "subagents"))
-	parent := strings.TrimSuffix(filepath.Base(sessionPath), ".jsonl")
-	artifacts, err := store.ListForParent(parent, root.Workspace)
-	if err != nil {
-		return ListingObs{Err: err.Error()}
-	}
-	out := ListingObs{}
-	for _, a := range artifacts {
-		out.Rows = append(out.Rows, ListingRow{
-			Ref: a.Ref, Status: string(a.Meta.Status), From: a.Meta.ParentToolCallID,
-		})
-	}
-	return out
-}
-
-type ArtifactObs struct {
-	Name  string `json:"name"`
-	Bytes int64  `json:"bytes"`
-}
-
-func capture(phase, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, root armRoot) Observation {
-	path := ctrl.SessionPath()
-	history := ctrl.History()
-	system := systemText(history)
-	snap := ctrl.ContextMaintenanceSnapshot()
-	st, loaded, loadErr := sessionstore.LoadCompactionState(path)
-	view, viewMsgs := viewObs(st, loaded, history)
-	graph, deltas := sink.snapshot()
-	return Observation{
-		Phase:           phase,
-		Arm:             arm,
-		SessionPath:     path,
-		SystemHash:      shortSum(system),
-		SystemBytes:     len(system),
-		BootSystemHash:  shortSum(bootSystem),
-		BootSystemBytes: len(bootSystem),
-		Transcript:      TranscriptObs{Messages: len(history), Digest: transcriptDigest(history)},
-		Goal:            GoalObs{Goal: ctrl.Goal(), Status: ctrl.GoalStatus(), TurnsUsed: ctrl.GoalRuntime().TurnsUsed},
-		Todos:           ctrl.Todos(),
-		Decisions:       ctrl.Decisions(),
-		Context: ContextObs{
-			ProjectionVersion: snap.ProjectionVersion,
-			CheckpointState:   snap.CheckpointState,
-			CanonicalTokens:   snap.CanonicalTokens,
-			ProjectedTokens:   snap.ProjectedTokens,
-			Blocked:           snap.Blocked,
-		},
-		Sidecar: sidecarObs(st, loaded, loadErr),
-		View:    view,
-		// The note is derived for a request and stored nowhere, so the contracts
-		// about it are read off the request rather than off the stored view.
-		TodoNotes:             todoNoteObs(ctrl.ModelVisibleMessages(), viewMsgs, ctrl.Todos()),
-		Graph:                 graphObs(graph, deltas),
-		WaitCauses:            waitCauseCounts(graph),
-		WaitSeries:            sink.waitSeries(),
-		Progress:              sink.phaseSeries(),
-		SettledWorkers:        settledWorkers(graph),
-		QueuedWorkers:         queuedWorkers(graph),
-		Children:              childrenObs(root, path),
-		Listing:               listingObs(root, path),
-		FanOutTurn:            fanOutTurnRecorded(history),
-		Executions:            control.ExecutionHistory(path),
-		InterruptedExecutions: ctrl.InterruptedExecutions(),
-		ModelSeesInterruptedExecution: countBlocks(
-			ctrl.ModelVisibleMessages(), "<interrupted-execution>"),
-		Deferred:              deferredObs(root.Workspace),
-		Obligation:            obligationObs(history),
-		Interrupted:           ctrl.InterruptedAdjudications(),
-		Journal:               control.AdjudicationHistory(path),
-		ModelSeesInterruption: countBlocks(ctrl.ModelVisibleMessages(), "<interrupted-adjudication>"),
-		Artifacts:             readArtifacts(path),
-	}
-}
-
-// bootSystemText is what the freshly built assembly composed, read before a
-// resume replaces the session under it.
-func bootSystemText(ctrl *control.Controller) string { return systemText(ctrl.History()) }
-
-func systemText(msgs []provider.Message) string {
-	var b strings.Builder
-	for _, m := range msgs {
-		if m.Role == provider.RoleSystem {
-			b.WriteString(m.Content)
-			b.WriteByte('\n')
-		}
-	}
-	return b.String()
-}
-
-// transcriptDigest covers only what reaches a provider, so a local-only marker
-// appended by either process cannot read as a changed conversation.
-func transcriptDigest(msgs []provider.Message) string {
-	h := sha256.New()
-	for _, m := range provider.ModelMessages(msgs) {
-		b, _ := json.Marshal(struct {
-			R, C, N string
-		}{string(m.Role), m.Content, m.Name})
-		h.Write(b)
-		h.Write([]byte{'\n'})
-	}
-	return hex.EncodeToString(h.Sum(nil))[:16]
-}
-
-func shortSum(s string) string {
-	sum := sha256.Sum256([]byte(s))
-	return hex.EncodeToString(sum[:8])
-}
-
-func sidecarObs(st sessionstore.CompactionState, ok bool, err error) SidecarObs {
-	out := SidecarObs{Present: ok}
-	if err != nil {
-		out.Err = err.Error()
-		return out
-	}
-	if !ok {
-		return out
-	}
-	out.Messages = len(st.Projection.Messages)
-	out.CoveredCount = st.Projection.CoveredCount
-	out.CoveredPrefixHash = st.Projection.CoveredPrefixHash
-	out.ProjectionVersion = st.Projection.ProjectionVersion
-	out.TranscriptVersion = st.TranscriptVersion
-	out.Generation = st.Generation
-	return out
-}
-
-// viewObs splices the model-visible context. With no usable projection the
-// view is the canonical transcript, which is what the host would send.
-func viewObs(st sessionstore.CompactionState, ok bool, canonical []provider.Message) (ViewObs, []provider.Message) {
-	var out ViewObs
-	if !ok || len(st.Projection.Messages) == 0 {
-		out.Messages = len(canonical)
-		out.Markers = markersIn(contents(canonical), markerPattern)
-		out.Echoes = markersIn(contents(canonical), echoPattern)
-		return out, canonical
-	}
-	body := st.Projection.Messages
-	out.BodyMarkers = markersIn(contents(body), markerPattern)
-	out.BodyEchoes = markersIn(contents(body), echoPattern)
-	view := append([]provider.Message(nil), body...)
-	if n := st.Projection.CoveredCount; n >= 0 && n < len(canonical) {
-		view = append(view, canonical[n:]...)
-		out.SplicedFromTail = len(canonical) - n
-	}
-	out.Messages = len(view)
-	out.Markers = markersIn(contents(view), markerPattern)
-	out.Echoes = markersIn(contents(view), echoPattern)
-	return out, view
-}
-
-// mentionsID reads an id the way the host's own visibility check does: text or
-// tool arguments, anywhere in the messages the request carries.
-func mentionsID(msgs []provider.Message, id string) bool {
-	for _, m := range msgs {
-		if strings.Contains(m.Content, id) {
-			return true
-		}
-		for _, call := range m.ToolCalls {
-			if strings.Contains(call.Arguments, id) {
-				return true
-			}
-		}
-	}
-	return false
-}
-
-func countBlocks(msgs []provider.Message, marker string) int {
-	n := 0
-	for _, m := range msgs {
-		if strings.Contains(m.Content, marker) {
-			n++
-		}
-	}
-	return n
-}
-
-func contents(msgs []provider.Message) []string {
-	out := make([]string, 0, len(msgs))
-	for _, m := range msgs {
-		out = append(out, m.Content)
-	}
-	return out
-}
-
-// waitCauseCounts is how many nodes the graph shows under each refusal. It
-// counts nodes rather than reports: a node keeps its cause after it starts, so
-// this answers what was refused, not what is refused now.
-func waitCauseCounts(g agentgraph.Graph) map[string]int {
-	out := map[string]int{}
-	for _, n := range g.Nodes {
-		if n.Wait != "" {
-			out[string(n.Wait)]++
-		}
-	}
-	if len(out) == 0 {
-		return nil
-	}
-	return out
-}
-
-// queuedWorkers are the items still refused admission, in a stable order.
-func queuedWorkers(g agentgraph.Graph) []string {
-	var out []string
-	for _, n := range g.Nodes {
-		if n.Kind != agentgraph.KindGroup && n.State == agentgraph.StateQueued {
-			out = append(out, n.ID)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-func graphObs(g agentgraph.Graph, deltas int) GraphObs {
-	out := GraphObs{Deltas: deltas}
-	if len(g.Nodes) == 0 {
-		return out
-	}
-	out.Nodes = g.Nodes
-	out.Edges = g.Edges
-	out.Grants = map[string]string{}
-	out.Waits = map[string]string{}
-	for _, n := range g.Nodes {
-		if n.Grant != "" {
-			out.Grants[n.ID] = string(n.Grant)
-		}
-		if n.Wait != "" {
-			out.Waits[n.ID] = string(n.Wait)
-		}
-	}
-	return out
-}
-
-// readArtifacts lists the session sidecars that exist, which is the evidence
-// behind every "persisted" claim in the matrix.
-func readArtifacts(sessionPath string) []ArtifactObs {
-	if sessionPath == "" {
-		return nil
-	}
-	paths := append([]string{sessionPath}, store.SessionSidecarFiles(sessionPath)...)
-	var out []ArtifactObs
-	for _, p := range paths {
-		if p == "" {
-			continue
-		}
-		info, err := os.Stat(p)
-		if err != nil {
-			continue
-		}
-		out = append(out, ArtifactObs{Name: filepath.Base(p), Bytes: info.Size()})
-	}
-	return out
-}
-
-func writeObservation(root armRoot, obs Observation) error {
-	b, err := json.MarshalIndent(obs, "", "  ")
-	if err != nil {
-		return err
-	}
-	return os.WriteFile(filepath.Join(root.Dir, "obs-"+obs.Phase+".json"), b, 0o644)
-}
-
-func readObservation(root armRoot, phase string) (Observation, error) {
-	var obs Observation
-	b, err := os.ReadFile(filepath.Join(root.Dir, "obs-"+phase+".json"))
-	if err != nil {
-		return obs, err
-	}
-	return obs, json.Unmarshal(b, &obs)
-}
@@ -1,338 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"reasonix/internal/state/sessionstore"
-	"time"
-
-	"reasonix/internal/runtime/agent"
-	"reasonix/internal/session/control"
-	"reasonix/internal/state/checkpoint"
-)
-
-// probeTurns is enough conversation for a fold to have a region to work on:
-// the system row and the first user turn are pinned, and the recent tail is
-// kept verbatim, so a short session has nothing foldable between them.
-const probeTurns = 8
-
-// refoldTurns grows the transcript past the first fold by less than the fold
-// keeps verbatim, so the second fold has to reach into the stored body.
-const refoldTurns = 3
-
-// runTurns drives count turns, each carrying its own marker, and returns the
-// next turn number.
-func runTurns(ctx context.Context, ctrl *control.Controller, from, count int) (int, error) {
-	for i := range count {
-		n := from + i + 1
-		prompt := fmt.Sprintf("%s Probe turn %d: record the task list and keep working.", marker(n), n)
-		if err := ctrl.Run(ctx, prompt); err != nil {
-			return n, fmt.Errorf("turn %d: %w", n, err)
-		}
-	}
-	return from + count, nil
-}
-
-// openDecisionAndDie leaves a question open and kills the process while the
-// turn is still parked on it. It exits rather than returning: a clean shutdown
-// would resolve or cancel the question, which is the one thing a process that
-// dies mid-decision does not do.
-func openDecisionAndDie(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink) error {
-	// The gate a person answers through, wired the way chat and desktop wire
-	// it. Without it the ask tool decides for itself and nothing ever blocks.
-	ctrl.EnableInteractiveApproval()
-	go func() { _ = ctrl.Run(ctx, marker(99)+" "+askSentinel+": ask before writing.") }()
-	if err := waitForAsk(ctrl); err != nil {
-		return err
-	}
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	if obs.Deferred.Executed {
-		return errUnexpected("a write still held back by the open question", obs.Deferred.MarkerPath)
-	}
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// waitForAsk blocks until the host reports a question waiting on a person.
-func waitForAsk(ctrl *control.Controller) error {
-	deadline := time.Now().Add(60 * time.Second)
-	for time.Now().Before(deadline) {
-		for _, d := range ctrl.Decisions() {
-			if d.Kind == control.DecisionAsk {
-				return nil
-			}
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return errUnexpected("a question waiting on a person", ctrl.Decisions())
-}
-
-// rewindTurn drives the rewind a person drives, recording the projection
-// before it: the tail arm takes back the turn appended after the fold, the
-// covered arm lands below the fold boundary and must lose the projection
-// whatever coverage says.
-func rewindTurn(root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink) error {
-	points := ctrl.Checkpoints()
-	if len(points) == 0 {
-		return errUnexpected("a checkpoint to rewind to", 0)
-	}
-	target := points[len(points)-1].Turn
-	if arm == armCoveredRewind {
-		var err error
-		if target, err = turnBelowFold(ctrl, points); err != nil {
-			return err
-		}
-	}
-	if err := writeObservation(root, capture(extraPhase(arm), arm, bootSystem, ctrl, sink, root)); err != nil {
-		return err
-	}
-	if err := ctrl.Rewind(target, control.RewindConversation); err != nil {
-		return fmt.Errorf("rewind: %w", err)
-	}
-	return nil
-}
-
-// turnBelowFold picks the checkpoint whose boundary sits just below the fold,
-// read from the sidecar rather than assumed: a control aiming at a fixed turn
-// stops removing folded history the moment coverage moves, and says nothing.
-// A boundary at the first message is skipped — a session left holding only a
-// system row is never persisted, so the restart would read the arm's work back.
-func turnBelowFold(ctrl *control.Controller, points []checkpoint.Meta) (int, error) {
-	covered := 0
-	if st, ok, _ := sessionstore.LoadCompactionState(ctrl.SessionPath()); ok {
-		covered = st.Projection.CoveredCount
-	}
-	target := -1
-	for _, p := range points {
-		plan, err := ctrl.PrepareRewind(p.Turn, control.RewindConversation)
-		if err != nil || !plan.HasBoundary {
-			continue
-		}
-		if plan.BoundaryIndex > 1 && plan.BoundaryIndex < covered {
-			target = p.Turn
-		}
-	}
-	if target < 0 {
-		return 0, errUnexpected("a checkpoint below the fold boundary", covered)
-	}
-	return target, nil
-}
-
-// extraPhase names the in-process observation an arm records before its
-// process exits, empty when it records none.
-func extraPhase(arm string) string {
-	switch arm {
-	case armRefoldIntoBody:
-		return "prefold"
-	case armTailRewind, armCoveredRewind:
-		return "prerewind"
-	}
-	return ""
-}
-
-const probeGoal = "Prove which runtime semantics survive an OS process boundary."
-
-// armConstruct hands the arm to the construct path it needs, reporting whether
-// one took over. Each of them records its own observation and exits inside the
-// turn, so a caller that sees taken has nothing left to do but report the error.
-func armConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) (bool, error) {
-	switch {
-	case arm == armOpenDecision || successorArm(arm):
-		return true, openDecisionAndDie(ctx, root, arm, bootSystem, ctrl, sink)
-	case deriveArm(arm):
-		return true, runDeriveConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case terminalArm(arm):
-		return true, runTerminalConstruct(ctx, root, arm, bootSystem, ctrl, sink, turn)
-	case schedulerWaitArm(arm):
-		return true, runSchedulerWaitConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case lineageArm(arm):
-		return true, runLineageConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case cancelRoutingArm(arm):
-		return true, runCancelRoutingConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case loneTaskArm(arm):
-		return true, runLoneTaskConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case activeStoreArm(arm):
-		return true, runActiveStoreConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case skillArm(arm):
-		return true, runSkillClassConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case ephemeralArm(arm):
-		return true, runEphemeralConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case hostExecArm(arm):
-		return true, runHostExecConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn)
-	case graphArm(arm) || uiArm(arm):
-		return true, runFanOutConstruct(ctx, root, arm, bootSystem, ctrl, sink, turn)
-	}
-	return false, nil
-}
-
-// runConstruct establishes host state through the same calls a frontend makes,
-// records what this process can see, and returns so the process can exit. It
-// never asserts: an arm that failed to establish something is reported as not
-// measured, which is the one reading a missing "before" supports.
-func runConstruct(dir, arm string) error {
-	root := rootFor(dir)
-	if err := root.create(); err != nil {
-		return err
-	}
-	if skillArm(arm) || arm == armEphemeralSkill || hostExecArm(arm) {
-		if err := writeProbeSkill(root.Workspace); err != nil {
-			return fmt.Errorf("probe skill: %w", err)
-		}
-	}
-	if total, writers := schedulerLimits(arm); total > 0 {
-		if err := root.writeProjectConfig(total, writers); err != nil {
-			return fmt.Errorf("project config: %w", err)
-		}
-	}
-	ctx := context.Background()
-	sink := &graphSink{}
-	bus, into := uiBus(arm, sink)
-	ctrl, prov, err := buildRuntime(ctx, root, arm, into)
-	if err != nil {
-		return fmt.Errorf("build: %w", err)
-	}
-	defer ctrl.Close()
-	// The process that dies is a window in production, and a window records what
-	// it served. Without that the trajectory this session leaves behind holds
-	// none of its own frames.
-	recordLikeAWindow(ctrl, bus)
-
-	bootSystem := bootSystemText(ctrl)
-	ctrl.EnsureSessionPath()
-	if ctrl.SessionPath() == "" {
-		return errUnexpected("a session path", "")
-	}
-	turn, err := runTurns(ctx, ctrl, 0, probeTurns)
-	if err != nil {
-		return err
-	}
-	if err := ctrl.SetGoalDurable(probeGoal); err != nil {
-		return fmt.Errorf("goal: %w", err)
-	}
-	if _, err := ctrl.Compact(ctx, agent.CompactRequest{Instructions: "Fold the probe turn."}); err != nil {
-		return fmt.Errorf("compact: %w", err)
-	}
-	if taken, err := armConstruct(ctx, root, arm, bootSystem, ctrl, sink, prov, turn); taken {
-		return err
-	}
-	if arm == armTodoIdentity {
-		// Identity A is already written and folded away. Grow, replace the list
-		// with identity B, then fold again: the second fold is what decides
-		// whether the model is left holding one current identity or two.
-		if turn, err = runTurns(ctx, ctrl, turn, refoldTurns); err != nil {
-			return err
-		}
-		if err := ctrl.Run(ctx, fmt.Sprintf("%s %s: replace the task list.", marker(turn+1), retodoSentinel)); err != nil {
-			return fmt.Errorf("retodo turn: %w", err)
-		}
-		// Grow well past the rewrite so the fold takes the new ids out of view.
-		// A scenario that leaves them readable asks nothing of the note.
-		if _, err = runTurns(ctx, ctrl, turn+1, probeTurns); err != nil {
-			return err
-		}
-		if _, err := ctrl.Compact(ctx, agent.CompactRequest{Instructions: "Fold again, after the identity moved."}); err != nil {
-			return fmt.Errorf("refold: %w", err)
-		}
-	} else if arm == armRefoldIntoBody {
-		// Grow past the fold, record the view a second fold will operate on,
-		// then fold again: its region reaches back into the stored body, which
-		// has no canonical counterpart to map a new boundary onto.
-		if _, err = runTurns(ctx, ctrl, turn, refoldTurns); err != nil {
-			return err
-		}
-		if err := writeObservation(root, capture(extraPhase(arm), arm, bootSystem, ctrl, sink, root)); err != nil {
-			return err
-		}
-		if _, err := ctrl.Compact(ctx, agent.CompactRequest{Instructions: "Fold again, into the stored body."}); err != nil {
-			return fmt.Errorf("refold: %w", err)
-		}
-	} else if appendsAfterFold(arm) {
-		if _, err = runTurns(ctx, ctrl, turn, 1); err != nil {
-			return err
-		}
-	}
-	if arm == armTailRewind || arm == armCoveredRewind {
-		if err := rewindTurn(root, arm, bootSystem, ctrl, sink); err != nil {
-			return err
-		}
-	}
-	return writeObservation(root, capture("construct", arm, bootSystem, ctrl, sink, root))
-}
-
-// successorArm reports whether this arm starts from an interrupted barrier.
-func successorArm(arm string) bool {
-	_, ok := successorTurn(arm)
-	return ok
-}
-
-// runSuccessor is the middle process: it inherits the interruption and does
-// what a person would do next — nothing, unrelated work, or something that
-// reads like an answer to the dead question. It runs a model on purpose; the
-// phase that judges what the host knows still does not.
-func runSuccessor(dir, arm string) error {
-	root := rootFor(dir)
-	before, err := readObservation(root, "construct")
-	if err != nil {
-		return err
-	}
-	ctx := context.Background()
-	sink := &graphSink{}
-	ctrl, _, err := buildRuntime(ctx, root, arm, sink)
-	if err != nil {
-		return fmt.Errorf("build: %w", err)
-	}
-	defer ctrl.Close()
-	bootSystem := bootSystemText(ctrl)
-	session, err := sessionstore.LoadSession(before.SessionPath)
-	if err != nil {
-		return fmt.Errorf("load session: %w", err)
-	}
-	if err := ctrl.Resume(session, before.SessionPath); err != nil {
-		return fmt.Errorf("resume: %w", err)
-	}
-	if turn, _ := successorTurn(arm); turn != "" {
-		if err := ctrl.Run(ctx, turn); err != nil {
-			return fmt.Errorf("successor turn: %w", err)
-		}
-	}
-	return writeObservation(root, capture("successor", arm, bootSystem, ctrl, sink, root))
-}
-
-// runResume boots a second time against the same roots and reads host state.
-// It deliberately never calls Run: the question is what the host can prove
-// before a provider call, and a model that reconstructs a lost identity from
-// the transcript would turn a real loss into a passing row.
-func runResume(dir, arm string) error {
-	root := rootFor(dir)
-	before, err := readObservation(root, "construct")
-	if err != nil {
-		return fmt.Errorf("read construct observation: %w", err)
-	}
-	ctx := context.Background()
-	sink := &graphSink{}
-	bus, into := uiBus(arm, sink)
-	ctrl, _, err := buildRuntime(ctx, root, arm, into)
-	if err != nil {
-		return fmt.Errorf("build: %w", err)
-	}
-	defer ctrl.Close()
-
-	bootSystem := bootSystemText(ctrl)
-	session, err := sessionstore.LoadSession(before.SessionPath)
-	if err != nil {
-		return fmt.Errorf("load session: %w", err)
-	}
-	if err := ctrl.Resume(session, before.SessionPath); err != nil {
-		return fmt.Errorf("resume: %w", err)
-	}
-	obs := capture("resume", arm, bootSystem, ctrl, sink, root)
-	if uiArm(arm) {
-		ui := observeUI(root, ctrl, bus, obs)
-		obs.UI = &ui
-	}
-	return writeObservation(root, obs)
-}
@@ -1,133 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"sort"
-	"strconv"
-	"strings"
-)
-
-// A fold may only remove a run of markers from the middle: the pinned head
-// stays, the retained tail stays, and everything between them becomes the
-// digest. Any other shape means the view lost something no fold folded.
-const (
-	verdictContiguous = "contiguous"
-	verdictHole       = "HOLE"
-	markPrefix        = "PROBE-MARK-"
-	echoPrefix        = "PROBE-ECHO-"
-)
-
-// refoldRows report what the second fold did. They are their own set because
-// the comparison is inside one process — the view before the refold against
-// the view after it — not across the process boundary.
-func refoldRows(prefold, postfold, resumed Observation, total int) []row {
-	return []row{
-		coverageRow("fold 1: stored body and coverage", prefold),
-		coverageRow("fold 2: stored body and coverage", postfold),
-		continuityRow("assistant work visible after fold 1", prefold, total),
-		continuityRow("assistant work visible after fold 2", postfold, total),
-		continuityRow("assistant work visible after restart", resumed, total),
-		turnsRow("user turns visible after fold 2", postfold),
-	}
-}
-
-func coverageRow(semantic string, o Observation) row {
-	return row{
-		Semantic: semantic, Authority: "compaction sidecar", Artifact: "<stem>.context.json",
-		Reconstruction: "projection.Messages + canonical[CoveredCount:]",
-		Before:         fmt.Sprintf("canonical=%d", o.Transcript.Messages),
-		After: fmt.Sprintf("body=%d covered=%d spliced=+%d view=%d",
-			o.Sidecar.Messages, o.Sidecar.CoveredCount, o.View.SplicedFromTail, o.View.Messages),
-		Verdict: verdictStable,
-	}
-}
-
-// continuityRow judges the assistant side: a fold replaces one run of replies
-// with a digest, so anything but a single surviving run is a loss no fold made.
-func continuityRow(semantic string, o Observation, total int) row {
-	shape := markerShape(o.View.Echoes, echoPrefix)
-	verdict := verdictContiguous
-	if reason := markerHole(o.View.Echoes, echoPrefix, total); reason != "" {
-		verdict, shape = verdictHole, shape+" — "+reason
-	}
-	return row{
-		Semantic: semantic, Authority: "the spliced model-visible view",
-		Artifact: "none: derived", Reconstruction: "assistant replies surviving in the view",
-		Before: fmt.Sprintf("emitted 1-%d", total), After: shape, Verdict: verdict,
-	}
-}
-
-// turnsRow reports the user side beside it: the retention budget keeps those
-// verbatim, which is why a one-sided tag could not see a dropped reply.
-func turnsRow(semantic string, o Observation) row {
-	return row{
-		Semantic: semantic, Authority: "the spliced model-visible view",
-		Artifact: "none: derived", Reconstruction: "user turns surviving in the view",
-		Before: "kept by the retention budget",
-		After:  markerShape(o.View.Markers, markPrefix), Verdict: verdictStable,
-	}
-}
-
-// pad renders a tag number the way the scripted turns emit it.
-func pad(n int) string { return fmt.Sprintf("%03d", n) }
-
-func markerNumbers(markers []string, prefix string) []int {
-	seen := map[int]bool{}
-	var out []int
-	for _, m := range markers {
-		n, err := strconv.Atoi(strings.TrimPrefix(m, prefix))
-		if err != nil || seen[n] {
-			continue
-		}
-		seen[n] = true
-		out = append(out, n)
-	}
-	sort.Ints(out)
-	return out
-}
-
-// markerShape renders the surviving markers as runs, which is the form the
-// judgement is about: "1, 9-12" says a fold took 2-8.
-func markerShape(markers []string, prefix string) string {
-	nums := markerNumbers(markers, prefix)
-	if len(nums) == 0 {
-		return "none"
-	}
-	var runs []string
-	start := nums[0]
-	for i := 1; i <= len(nums); i++ {
-		if i < len(nums) && nums[i] == nums[i-1]+1 {
-			continue
-		}
-		if end := nums[i-1]; end == start {
-			runs = append(runs, strconv.Itoa(start))
-		} else {
-			runs = append(runs, fmt.Sprintf("%d-%d", start, end))
-		}
-		if i < len(nums) {
-			start = nums[i]
-		}
-	}
-	return strings.Join(runs, ", ")
-}
-
-// markerHole names the violation, or returns "" when the shape is one a fold
-// can produce: an optional pinned head starting at 1, then one run reaching the
-// last turn. Two gaps, or a tail that stops short of it, are losses.
-func markerHole(markers []string, prefix string, total int) string {
-	nums := markerNumbers(markers, prefix)
-	if len(nums) == 0 {
-		return "the view shows no turn at all"
-	}
-	runs := strings.Split(markerShape(markers, prefix), ", ")
-	if last := nums[len(nums)-1]; last != total {
-		return fmt.Sprintf("the newest turn %d is missing", total)
-	}
-	if len(runs) > 2 {
-		return fmt.Sprintf("%d separate runs; a fold removes one", len(runs))
-	}
-	if len(runs) == 2 && nums[0] != 1 {
-		return fmt.Sprintf("the kept head starts at %d, not the pinned first turn", nums[0])
-	}
-	return ""
-}
@@ -1,45 +0,0 @@
-package main
-
-import "testing"
-
-// The oracle has to be able to fail, or "no hole" says nothing. These are the
-// shapes a fold can produce and the shapes only a loss can.
-func TestMarkerHoleDiscriminates(t *testing.T) {
-	echo := func(ns ...int) []string {
-		out := make([]string, 0, len(ns))
-		for _, n := range ns {
-			out = append(out, echoPrefix+pad(n))
-		}
-		return out
-	}
-	cases := []struct {
-		name  string
-		in    []string
-		total int
-		holed bool
-	}{
-		{"nothing folded", echo(1, 2, 3, 4, 5), 5, false},
-		{"middle folded away", echo(4, 5), 5, false},
-		{"pinned head kept beside the tail", echo(1, 4, 5), 5, false},
-		{"a reply dropped inside the kept tail", echo(3, 5), 5, true},
-		{"two gaps", echo(1, 3, 5), 5, true},
-		{"newest turn missing", echo(3, 4), 5, true},
-		{"head kept but not the first turn", echo(2, 4, 5), 5, true},
-		{"empty view", nil, 5, true},
-	}
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			reason := markerHole(tc.in, echoPrefix, tc.total)
-			if holed := reason != ""; holed != tc.holed {
-				t.Fatalf("markerHole(%v) = %q, want holed=%v", tc.in, reason, tc.holed)
-			}
-		})
-	}
-}
-
-func TestMarkerShapeRendersRuns(t *testing.T) {
-	in := []string{echoPrefix + pad(1), echoPrefix + pad(9), echoPrefix + pad(10), echoPrefix + pad(11)}
-	if got := markerShape(in, echoPrefix); got != "1, 9-11" {
-		t.Fatalf("markerShape = %q, want %q", got, "1, 9-11")
-	}
-}
@@ -1,60 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"strings"
-)
-
-// renderMatrix prints one table per arm. It reports what each row's value was
-// on both sides of the process boundary, because a verdict without the two
-// readings behind it is an assertion.
-func renderMatrix(results []armResult, work string) string {
-	var b strings.Builder
-	b.WriteString("# Runtime resurrection matrix\n\n")
-	b.WriteString("Measurement boundary: OS process. Phase B loads and inspects; it never calls a model.\n\n")
-	fmt.Fprintf(&b, "Arm roots: %s\n\n", work)
-	for _, res := range results {
-		fmt.Fprintf(&b, "## arm: %s\n\n%s\n\n", res.Arm, res.Asks)
-		if res.Invalid != "" {
-			fmt.Fprintf(&b, "**ARM INVALID — %s.** Its rows below establish nothing.\n\n", res.Invalid)
-		}
-		b.WriteString("| Semantic | Before death | After new process | Verdict | Authority | Artifact |\n")
-		b.WriteString("| --- | --- | --- | --- | --- | --- |\n")
-		for _, r := range res.Rows {
-			fmt.Fprintf(&b, "| %s | %s | %s | %s | %s | %s |\n",
-				r.Semantic, cell(r.Before), cell(r.After), r.Verdict, r.Authority, r.Artifact)
-		}
-		b.WriteString("\n")
-		b.WriteString(summarize(res))
-		b.WriteString("\n")
-	}
-	return b.String()
-}
-
-func summarize(res armResult) string {
-	counts := map[string]int{}
-	for _, r := range res.Rows {
-		counts[r.Verdict]++
-	}
-	var parts []string
-	// Every verdict the matrix can produce. A summary that silently omits one
-	// is how a VIOLATED row reads as a clean arm.
-	for _, v := range []string{verdictPersisted, verdictExact, verdictLossy, verdictLost, verdictNotMeasured,
-		verdictChanged, verdictStable, verdictContiguous, verdictHole, verdictHolds, verdictViolated} {
-		if counts[v] > 0 {
-			parts = append(parts, fmt.Sprintf("%s=%d", v, counts[v]))
-		}
-	}
-	return strings.Join(parts, "  ") + "\n"
-}
-
-func cell(s string) string {
-	if strings.TrimSpace(s) == "" {
-		return "—"
-	}
-	s = strings.ReplaceAll(s, "|", "\\|")
-	if len(s) > 64 {
-		return s[:61] + "..."
-	}
-	return s
-}
@@ -1,323 +0,0 @@
-package main
-
-import (
-	"encoding/json"
-	"fmt"
-	"sort"
-	"strings"
-)
-
-// Verdicts. They are deliberately not booleans: "the host holds it verbatim"
-// and "a fold over canonical artifacts happens to agree" are different claims
-// about durability, and a row that was never established supports neither.
-const (
-	verdictPersisted   = "persisted-direct"
-	verdictExact       = "reconstructed-exact"
-	verdictLossy       = "reconstructed-lossy"
-	verdictLost        = "lost"
-	verdictNotMeasured = "not-measured"
-	verdictChanged     = "changed"
-	verdictStable      = "stable"
-)
-
-type row struct {
-	Semantic       string `json:"semantic"`
-	Authority      string `json:"authority"`
-	Artifact       string `json:"artifact"`
-	Reconstruction string `json:"reconstruction"`
-	Before         string `json:"before"`
-	After          string `json:"after"`
-	Verdict        string `json:"verdict"`
-}
-
-type armResult struct {
-	Arm     string       `json:"arm"`
-	Asks    string       `json:"asks"`
-	Invalid string       `json:"invalid,omitempty"`
-	Rows    []row        `json:"rows"`
-	Extra   *Observation `json:"extra,omitempty"`
-	Before  Observation  `json:"before"`
-	After   Observation  `json:"after"`
-}
-
-func classify(a arm, extra *Observation, before, after Observation) armResult {
-	res := armResult{Arm: a.name, Asks: a.asks, Extra: extra, Before: before, After: after}
-	prefix := prefixRow(before, after)
-	res.Rows = append(res.Rows, prefix)
-	if a.name == "system-swap" && prefix.Verdict != verdictChanged {
-		res.Invalid = "the lever did not move the stable prefix, so this arm measures nothing"
-	}
-	if a.name == "exact" && prefix.Verdict != verdictStable {
-		res.Invalid = "the stable prefix moved without a lever, so this arm is not the control it claims"
-	}
-	res.Rows = append(res.Rows,
-		valueRow("system row in the session", "the transcript's own leading system message",
-			"<stem>.jsonl", "loaded with the transcript",
-			before.SystemHash, after.SystemHash, true),
-		valueRow("canonical transcript", "session transcript + event log",
-			"<stem>.jsonl, <stem>.events.jsonl", "LoadSession replays the append log",
-			before.Transcript.Digest, after.Transcript.Digest, true),
-		valueRow("active goal identity", "goalMachine (control)",
-			"<stem>.goal-state.json", "restoreFromState on Resume",
-			before.Goal.Goal, after.Goal.Goal, true),
-		valueRow("goal status", "goalMachine (control)",
-			"<stem>.goal-state.json", "restoreFromState on Resume",
-			before.Goal.Status, after.Goal.Status, true),
-		valueRow("todo host identity", "agent todoState (host-owned, never in the prompt)",
-			"none verbatim; goal-state.json snapshots a terminal goal's list",
-			"rebuilt from the transcript's todo_write / complete_step receipts",
-			todoIDs(before), todoIDs(after), false),
-		valueRow("open decisions", "approvalManager (control, in-memory)",
-			"none", "none: derived from live pending approvals/asks",
-			decisionIDs(before), decisionIDs(after), false),
-		valueRow("compaction projection (stored)", "compaction sidecar",
-			"<stem>.context.json", "LoadProjectionSidecar on bind",
-			sidecarClaim(before), sidecarClaim(after), true),
-		useRow("compaction projection (in use)", before, after),
-		valueRow("run graph nodes", "event sink (GraphDelta)",
-			"none observed", "fold GraphDelta events, if any survive",
-			graphNodes(before), graphNodes(after), false),
-		valueRow("run graph grants", "event sink (GraphDelta)",
-			"none observed", "fold GraphDelta events, if any survive",
-			graphMap(before.Graph.Grants), graphMap(after.Graph.Grants), false),
-		valueRow("run graph wait causes", "event sink (GraphDelta)",
-			"none observed", "fold GraphDelta events, if any survive",
-			graphMap(before.Graph.Waits), graphMap(after.Graph.Waits), false),
-	)
-	switch {
-	case successorArm(a.name) && extra != nil:
-		res.Rows = append(res.Rows, successorRows(a.name, *extra, after)...)
-		// The premise is a barrier open at death. The construct process is its
-		// own waiter, so it reports no interruption of its own.
-		if len(before.Decisions) == 0 {
-			res.Invalid = "no decision was open when the first process died"
-		}
-	case a.name == armOpenDecision:
-		res.Rows = append(res.Rows, decisionRows(before, after)...)
-		if len(before.Decisions) == 0 {
-			res.Invalid = "no decision was open when the process died, so nothing was interrupted"
-		}
-	case deriveArm(a.name):
-		res.Rows = append(res.Rows, deriveRows(a.name, before, after)...)
-		res.Invalid = deriveArmInvalid(a.name, before)
-	case terminalArm(a.name):
-		res.Rows = append(res.Rows, terminalRows(a.name, before, after)...)
-		res.Invalid = terminalArmInvalid(a.name, before)
-	case schedulerWaitArm(a.name):
-		res.Rows = append(res.Rows, schedulerWaitRows(a.name, before, after)...)
-		res.Invalid = schedulerArmInvalid(a.name, before)
-	case graphArm(a.name):
-		res.Rows = append(res.Rows, graphRows(before, after)...)
-		res.Invalid = graphArmInvalid(a.name, before)
-	case uiArm(a.name):
-		res.Rows = append(res.Rows, uiRows(before, after)...)
-		res.Invalid = uiArmInvalid(before, after)
-	case lineageArm(a.name):
-		res.Rows = append(res.Rows, lineageRows(a.name, before)...)
-	case cancelRoutingArm(a.name):
-		res.Rows = append(res.Rows, cancelRoutingRows(a.name, before)...)
-	case loneTaskArm(a.name):
-		res.Rows = append(res.Rows, loneTaskRows(a.name, before, after)...)
-		res.Invalid = loneTaskArmInvalid(a.name, before)
-	case activeStoreArm(a.name):
-		res.Rows = append(res.Rows, activeStoreRows(before, after)...)
-		res.Invalid = activeStoreInvalid(before)
-	case skillArm(a.name):
-		res.Rows = append(res.Rows, skillClassRows(a.name, before, after)...)
-		res.Invalid = skillArmInvalid(a.name, before)
-	case ephemeralArm(a.name):
-		res.Rows = append(res.Rows, ephemeralRows(a.name, before, after)...)
-		res.Invalid = ephemeralArmInvalid(before)
-	case hostExecArm(a.name):
-		res.Rows = append(res.Rows, hostExecRows(a.name, before, after)...)
-		res.Invalid = hostExecArmInvalid(a.name, before)
-	case a.name == armTodoIdentity:
-		res.Rows = append(res.Rows, todoRows(before, after)...)
-	case a.name == armRefoldIntoBody && extra != nil:
-		res.Rows = append(res.Rows, refoldRows(*extra, before, after, probeTurns+refoldTurns)...)
-		if extra.Sidecar.Messages == 0 {
-			res.Invalid = "the first fold stored no body, so the second had none to reach into"
-		}
-	case (a.name == armTailRewind || a.name == armCoveredRewind) && extra != nil:
-		res.Rows = append(res.Rows, useRow("projection across the rewind, same process", *extra, before))
-		res.Rows = append(res.Rows, rewindLandingRow(a.name, *extra, before))
-		res.Invalid = rewindLandedWrong(a.name, *extra, before)
-	}
-	return res
-}
-
-// prefixRow is the arm's own control rather than a durability claim: it reports
-// whether the stable prefix moved, which is the thing each arm sets up.
-func prefixRow(before, after Observation) row {
-	verdict := verdictStable
-	if before.BootSystemHash != after.BootSystemHash {
-		verdict = verdictChanged
-	}
-	return row{
-		Semantic: "stable prefix identity (composed)", Authority: "boot assembly (recomposed per process)",
-		Artifact: "none (composed from config, memory and instructions)", Reconstruction: "recomposed at boot",
-		Before: before.BootSystemHash, After: after.BootSystemHash, Verdict: verdict,
-	}
-}
-
-func valueRow(semantic, authority, artifact, reconstruction, before, after string, persisted bool) row {
-	r := row{Semantic: semantic, Authority: authority, Artifact: artifact,
-		Reconstruction: reconstruction, Before: before, After: after}
-	switch {
-	case before == "" || before == "-":
-		r.Verdict = verdictNotMeasured
-	case after == "" || after == "-":
-		r.Verdict = verdictLost
-	case before != after:
-		r.Verdict = verdictLossy
-	case persisted:
-		r.Verdict = verdictPersisted
-	default:
-		r.Verdict = verdictExact
-	}
-	return r
-}
-
-func todoIDs(o Observation) string {
-	var ids []string
-	for _, t := range o.Todos {
-		id := t.StepID
-		if id == "" {
-			id = "(no step_id)"
-		}
-		ids = append(ids, id+":"+t.Status)
-	}
-	return join(ids)
-}
-
-func decisionIDs(o Observation) string {
-	var ids []string
-	for _, d := range o.Decisions {
-		ids = append(ids, string(d.Kind)+":"+d.ID)
-	}
-	return join(ids)
-}
-
-func sidecarClaim(o Observation) string {
-	if !o.Sidecar.Present {
-		return ""
-	}
-	return fmt.Sprintf("v%d msgs=%d covered=%d hash=%s",
-		o.Sidecar.ProjectionVersion, o.Sidecar.Messages, o.Sidecar.CoveredCount, short(o.Sidecar.CoveredPrefixHash))
-}
-
-// rewindLandingRow says where the rewind actually landed relative to the fold.
-// An arm that names a side and does not reach it is not measuring that side.
-func rewindLandingRow(arm string, prerewind, after Observation) row {
-	side := "above the fold boundary"
-	if after.Transcript.Messages < prerewind.Sidecar.CoveredCount {
-		side = "below the fold boundary"
-	}
-	return row{
-		Semantic: "where the rewind landed", Authority: "checkpoint boundary against the sidecar",
-		Artifact: "none: derived", Reconstruction: "canonical length after the rewind against CoveredCount before it",
-		Before: fmt.Sprintf("covered=%d", prerewind.Sidecar.CoveredCount),
-		After:  fmt.Sprintf("canonical=%d, %s", after.Transcript.Messages, side), Verdict: verdictStable,
-	}
-}
-
-// rewindLandedWrong invalidates an arm that aimed at one side of the fold and
-// reached the other. covered-rewind stopped removing folded history the moment
-// coverage became a real boundary, and nothing in the matrix said so.
-func rewindLandedWrong(arm string, prerewind, after Observation) string {
-	below := after.Transcript.Messages < prerewind.Sidecar.CoveredCount
-	switch {
-	case arm == armCoveredRewind && !below:
-		return "the rewind landed above the fold boundary, so it removed no folded history"
-	case arm == armTailRewind && below:
-		return "the rewind landed below the fold boundary, so it is not a tail-only truncation"
-	}
-	return ""
-}
-
-// useRow turns on whether the next request would ride the projection, not on
-// the token counts beside it: a tail edit legitimately moves those, and letting
-// them decide the verdict reported a surviving projection as a loss.
-func useRow(semantic string, before, after Observation) row {
-	r := valueRow(semantic, "projectionValid over the covered prefix",
-		"none: a judgement, not a stored value",
-		"revalidated against the canonical prefix on every read",
-		projectionState(before), projectionState(after), false)
-	r.Before, r.After = projectionUse(before), projectionUse(after)
-	return r
-}
-
-// projectionState is the compared value: in use, not in use, or never measured.
-func projectionState(o Observation) string {
-	switch o.Context.CheckpointState {
-	case "":
-		return ""
-	case "none":
-		return "NOT-in-use"
-	}
-	return "in-use"
-}
-
-// projectionUse reads the host's own judgement rather than the stored claim.
-// CheckpointState is "none" whenever validation failed, which is the difference
-// between a sidecar that loaded and one the next request will actually ride.
-// The label itself is not compared — "applied" and "restored" are the same
-// answer reached from different sides — so the row turns on use and size.
-func projectionUse(o Observation) string {
-	if o.Context.CheckpointState == "" {
-		return ""
-	}
-	use := "in-use"
-	if o.Context.CheckpointState == "none" {
-		use = "NOT-in-use"
-	}
-	return fmt.Sprintf("%s projected=%d canonical=%d", use,
-		o.Context.ProjectedTokens, o.Context.CanonicalTokens)
-}
-
-// graphNodes renders the fold's nodes in a stable order. The fold keeps
-// first-seen order, which two processes reach by different routes, so an
-// unsorted rendering would report a reordering as a changed graph.
-func graphNodes(o Observation) string {
-	if len(o.Graph.Nodes) == 0 {
-		return ""
-	}
-	ids := make([]string, 0, len(o.Graph.Nodes))
-	for _, n := range o.Graph.Nodes {
-		ids = append(ids, n.ID+":"+string(n.State))
-	}
-	sort.Strings(ids)
-	return join(ids)
-}
-
-// graphMap renders one per-node attribute set. Map iteration is unordered, so
-// without the sort two identical envelopes would compare as different.
-func graphMap(m map[string]string) string {
-	if len(m) == 0 {
-		return ""
-	}
-	out := make([]string, 0, len(m))
-	for k, v := range m {
-		out = append(out, k+"="+v)
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func join(in []string) string {
-	if len(in) == 0 {
-		return ""
-	}
-	return strings.Join(in, ", ")
-}
-
-func short(s string) string {
-	if len(s) <= 12 {
-		return s
-	}
-	return s[:12]
-}
-
-func marshalResults(results []armResult) ([]byte, error) {
-	return json.MarshalIndent(results, "", "  ")
-}
@@ -1,265 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"io"
-	"os"
-	"path/filepath"
-	"strings"
-	"sync"
-	"time"
-
-	"reasonix/internal/assembly/boot"
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/contract/event"
-	"reasonix/internal/session/control"
-)
-
-// armRoot names the four directories one arm owns. Home and Workspace are
-// separate so a lever can change what the prefix folds without touching the
-// transcript, and Sessions is explicit so the probe never writes to the
-// developer's own session directory.
-type armRoot struct{ Home, Workspace, Sessions, Dir string }
-
-func rootFor(dir string) armRoot {
-	return armRoot{
-		Dir:       dir,
-		Home:      filepath.Join(dir, "home"),
-		Workspace: filepath.Join(dir, "workspace"),
-		Sessions:  filepath.Join(dir, "sessions"),
-	}
-}
-
-func (r armRoot) create() error {
-	for _, d := range []string{r.Home, r.Workspace, r.Sessions} {
-		if err := os.MkdirAll(d, 0o755); err != nil {
-			return err
-		}
-	}
-	return nil
-}
-
-// writeProjectConfig puts a project config beside the workspace, which is where
-// a person configures these ceilings. The scheduler arms need them small enough
-// to reach a specific refusal, and a poked field would prove nothing about the
-// judgement production makes.
-func (r armRoot) writeProjectConfig(total, writers int) error {
-	body := fmt.Sprintf("[agent]\nmax_subagent_concurrency = %d\nmax_parallel_writers = %d\n", total, writers)
-	return os.WriteFile(filepath.Join(r.Workspace, "reasonix.toml"), []byte(body), 0o644)
-}
-
-// requireFresh refuses a root a previous run left behind. A reused root is how
-// a lever's effect ends up already present in the construct phase, which reads
-// as "the lever changed nothing" — a false negative that looks like a result.
-func (r armRoot) requireFresh() error {
-	entries, err := os.ReadDir(r.Dir)
-	if err != nil {
-		if os.IsNotExist(err) {
-			return nil
-		}
-		return err
-	}
-	if len(entries) > 0 {
-		return fmt.Errorf("arm root %s is not empty; each run needs a clean root", r.Dir)
-	}
-	return nil
-}
-
-// graphSink records what the run graph published while this process was alive.
-// It is the only "before" the graph row has: nothing else in a session's
-// artifacts is known to carry it, and a row with no before is not measured.
-type graphSink struct {
-	mu     sync.Mutex
-	graph  agentgraph.Graph
-	deltas int
-	// waits is every wait cause published per node, in order. The fold keeps
-	// only the latest, which cannot say whether a cause was reported once or
-	// replaced — and that is exactly what the transition arm asks.
-	waits map[string][]string
-	// phases is every subagent status the parent was told, per call. A run
-	// handed to a job publishes no graph node at all, so for those this is the
-	// only live evidence that anything happened.
-	phases map[string][]string
-	// tools is when each call was dispatched and when it answered. A stop that
-	// reached a running tool is visible as the second arriving early.
-	tools map[string][2]time.Time
-}
-
-func (s *graphSink) Emit(e event.Event) {
-	if e.Kind == event.ToolDispatch || e.Kind == event.ToolResult {
-		s.noteTool(e)
-		return
-	}
-	if id, phase, ok := progressPhase(e); ok {
-		s.mu.Lock()
-		if s.phases == nil {
-			s.phases = map[string][]string{}
-		}
-		s.phases[id] = append(s.phases[id], phase)
-		s.mu.Unlock()
-		return
-	}
-	if e.Kind != event.GraphDelta || e.Graph == nil {
-		return
-	}
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	s.deltas++
-	for _, n := range e.Graph.Nodes {
-		if n.Wait == "" {
-			continue
-		}
-		if s.waits == nil {
-			s.waits = map[string][]string{}
-		}
-		s.waits[n.ID] = append(s.waits[n.ID], string(n.Wait))
-	}
-	s.graph.Apply(*e.Graph)
-}
-
-func (s *graphSink) snapshot() (agentgraph.Graph, int) {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	return s.graph, s.deltas
-}
-
-// noteTool records the two edges of a tool call this probe may be asked about.
-func (s *graphSink) noteTool(e event.Event) {
-	if e.Tool.ID == "" {
-		return
-	}
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	if s.tools == nil {
-		s.tools = map[string][2]time.Time{}
-	}
-	at := s.tools[e.Tool.ID]
-	if e.Kind == event.ToolDispatch {
-		at[0] = time.Now()
-	} else {
-		at[1] = time.Now()
-	}
-	s.tools[e.Tool.ID] = at
-}
-
-func (s *graphSink) toolDispatched(id string) (time.Time, bool) {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	at, ok := s.tools[id]
-	return at[0], ok && !at[0].IsZero()
-}
-
-func (s *graphSink) toolResult(id string) (time.Time, bool) {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	at, ok := s.tools[id]
-	return at[1], ok && !at[1].IsZero()
-}
-
-// phaseSeries is every status published for one call, in publication order.
-func (s *graphSink) phaseSeries() map[string][]string {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	out := map[string][]string{}
-	for id, seq := range s.phases {
-		out[id] = append([]string(nil), seq...)
-	}
-	return out
-}
-
-// waitSeries is every cause published for one node, in publication order.
-func (s *graphSink) waitSeries() map[string][]string {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	out := map[string][]string{}
-	for id, seq := range s.waits {
-		out[id] = append([]string(nil), seq...)
-	}
-	return out
-}
-
-// buildRuntime assembles a real runtime bound to this arm's roots. Everything a
-// frontend would supply is supplied; only the provider is scripted, so the
-// probe never depends on a network or a key.
-func buildRuntime(ctx context.Context, root armRoot, arm string, sink event.Sink) (*control.Controller, *scripted, error) {
-	prov := newScripted(arm)
-	ctrl, err := boot.Build(ctx, boot.Options{
-		Version:              "runtime-resume-probe",
-		Model:                probeModelRef,
-		Sink:                 sink,
-		WorkspaceRoot:        root.Workspace,
-		Home:                 root.Home,
-		SessionDir:           root.Sessions,
-		ProviderResolver:     resolverFor(prov),
-		Stderr:               io.Discard,
-		HeadlessApprovalMode: approvalMode(arm),
-	})
-	if err != nil {
-		return nil, nil, err
-	}
-	ctrl.ApplyHeadlessApprovalMode(approvalMode(arm))
-	return ctrl, prov, nil
-}
-
-// approvalMode is the gate this arm needs. Everything else runs denied, the
-// strictest a headless frontend can be. A fleet is not read-only, so under a
-// denied gate the dispatch is refused before any child starts and the arm
-// measures a permission decision instead of a process boundary.
-func approvalMode(arm string) string {
-	if graphArm(arm) || uiArm(arm) || loneTaskArm(arm) || cancelRoutingArm(arm) || lineageArm(arm) ||
-		schedulerWaitArm(arm) || terminalArm(arm) || deriveArm(arm) || activeStoreArm(arm) ||
-		skillArm(arm) || ephemeralArm(arm) || hostExecArm(arm) {
-		return control.ToolApprovalAuto
-	}
-	return "deny"
-}
-
-// childEnv binds the child process to this arm's home through the environment
-// as well as through Options.Home: a Controller's later re-reads and the shared
-// history index still read the process, not the assembly.
-func childEnv(root armRoot) []string {
-	env := append(os.Environ(),
-		"REASONIX_HOME="+root.Home,
-		"REASONIX_STATE_HOME="+root.Home,
-	)
-	// A child runs in its own workspace, so the one arm that has to build the
-	// frontend is told where the checkout is rather than searching for it.
-	if repo := repoRoot(); repo != "" {
-		env = append(env, envProbeRepo+"="+repo)
-	}
-	return env
-}
-
-// repoRoot walks up from the orchestrator's working directory for the module
-// this probe lives in. Empty when it is not run from a checkout, which the arm
-// reports as not measured rather than as a failure.
-func repoRoot() string {
-	dir, err := os.Getwd()
-	if err != nil {
-		return ""
-	}
-	for {
-		if b, err := os.ReadFile(filepath.Join(dir, "go.mod")); err == nil &&
-			strings.HasPrefix(string(b), "module reasonix") {
-			return dir
-		}
-		parent := filepath.Dir(dir)
-		if parent == dir {
-			return ""
-		}
-		dir = parent
-	}
-}
-
-// toolIDs is every call this sink saw dispatched, which is where a host-minted
-// event id first appears: it is not in the graph and not in the transcript.
-func (s *graphSink) toolIDs() map[string]bool {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	out := make(map[string]bool, len(s.tools))
-	for id := range s.tools {
-		out[id] = true
-	}
-	return out
-}
@@ -1,376 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"sort"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/session/control"
-)
-
-// The scheduler-wait arms. Each dies with one item ready and refused admission,
-// and they are separate because the scheduler decides in a fixed order —
-// slots, then writers, then claim. An arm that let an earlier cause stand would
-// report it as the later one, so each has to clear every check above the one it
-// is measuring.
-const (
-	// armWaitSlots fills the session's total ceiling with runs that never end.
-	// The refused item is read-only: it clears every writer check by not being
-	// a writer at all.
-	armWaitSlots = "wait-slots"
-	// armWaitWriters leaves total capacity free and fills the writer ceiling.
-	// Its two writers declare disjoint paths, so a claim conflict cannot be
-	// what refuses the second one.
-	armWaitWriters = "wait-writers"
-	// armWaitClaim leaves both ceilings free and overlaps two write paths. A
-	// single fan-out cannot reach this: preflight refuses concurrent writers
-	// whose claims overlap, so the holder is dispatched as a background fleet
-	// and the refused writer comes from a later one.
-	armWaitClaim = "wait-claim"
-	// armWaitTransition asks what a wait cause means once the thing it named
-	// stops being what holds the item back. The scheduler reports the cause
-	// once, when the request queues; the blocker can change afterwards.
-	armWaitTransition = "wait-transition"
-)
-
-func schedulerWaitArm(name string) bool {
-	switch name {
-	case armWaitSlots, armWaitWriters, armWaitClaim, armWaitTransition:
-		return true
-	}
-	return false
-}
-
-// schedulerLimits are the ceilings an arm needs to reach its refusal. They are
-// written to the project config, so the run makes the same judgement a person
-// configuring the session would get.
-func schedulerLimits(arm string) (total, writers int) {
-	switch arm {
-	case armTerminalCancelled:
-		// One slot, so the group has an item running and items that never got
-		// one — the split a cancellation is reported against.
-		return 1, 1
-	case armLineageBackground, armLineageJobKill:
-		// Room for a second run: the alive check asks whether the manager can
-		// still start anything, and a ceiling of one would answer for the
-		// hanging child instead.
-		return 4, 4
-	case armTaskFgQueued, armTaskBgQueued, armTaskFgQueuedCancel, armTaskBgQueuedCancel:
-		// One slot, held by the fleet dispatched ahead of the delegation, so the
-		// refusal the arm records is the scheduler's and not a race.
-		return 1, 1
-	case armUIGraphMixed:
-		// One slot, so the item after the one that hangs is refused: the death
-		// then holds a node that was queued and never admitted.
-		return 1, 1
-	case armHostQueued, armHeadlessQueuedCancel:
-		// One slot, held by the fan-out dispatched ahead of the invocation.
-		return 1, 1
-	case armSkillQueued, armReadOnlySkillQueued:
-		// One slot, held by the fleet dispatched ahead of the skill, so the
-		// refusal the arm reads is the scheduler's and not a race.
-		return 1, 1
-	case armFleetActiveStore:
-		// Two slots: the delegation holds one to the death, and the fan-out's
-		// first item frees the other for exactly one of the two behind it.
-		return 2, 2
-	case armWaitSlots:
-		return 2, 1
-	case armWaitWriters:
-		return 6, 1
-	case armWaitClaim:
-		return 6, 3
-	case armWaitTransition:
-		return 3, 1
-	}
-	return 0, 0
-}
-
-// wantedCause is the refusal this arm exists to produce. The transition arm
-// starts at slots and asks what happens to that answer afterwards.
-func wantedCause(arm string) agentgraph.WaitCause {
-	switch arm {
-	case armWaitWriters:
-		return agentgraph.WaitWriters
-	case armWaitClaim:
-		return agentgraph.WaitClaim
-	}
-	return agentgraph.WaitSlots
-}
-
-// waitingWorkers are the nodes the scheduler refused and has not admitted: the
-// arm's premise. State is checked as well as the cause, because a node that
-// went on to run still carries the cause it waited under.
-func waitingWorkers(g agentgraph.Graph, cause agentgraph.WaitCause) []string {
-	var out []string
-	for _, n := range g.Nodes {
-		if n.Kind == agentgraph.KindGroup || n.State != agentgraph.StateQueued {
-			continue
-		}
-		if n.Wait == cause {
-			out = append(out, n.ID)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// waitForRefusal blocks until one item is queued under this arm's cause.
-func waitForRefusal(sink *graphSink, arm string) error {
-	cause := wantedCause(arm)
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		graph, _ := sink.snapshot()
-		if len(waitingWorkers(graph, cause)) > 0 {
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected("an item queued under "+string(cause), waitStates(graph))
-}
-
-func waitStates(g agentgraph.Graph) string {
-	out := make([]string, 0, len(g.Nodes))
-	for _, n := range g.Nodes {
-		out = append(out, fmt.Sprintf("%s:%s/%s", n.ID, n.State, orNone(string(n.Wait))))
-	}
-	sort.Strings(out)
-	return orNone(join(out))
-}
-
-// schedTurn is the prompt that opens this arm's fan-out. The slots arm needs
-// only one fleet; the others name both, and the latch order dispatches the
-// holder first while the refused fleet waits on the holder's own signal.
-func schedTurn(n int, arm string) string {
-	if arm == armWaitSlots {
-		return fmt.Sprintf("%s %s: dispatch the fan-out.", marker(n), fleetPairSentinel)
-	}
-	return fmt.Sprintf("%s %s %s: dispatch the fan-out.", marker(n), fleetHolderSentinel, fleetRefusedSentinel)
-}
-
-// runSchedulerWaitConstruct drives the arm to its refusal and exits while the
-// item is still queued. The transition arm goes further: it frees the capacity
-// the refusal named and asks what the reported cause does about it.
-func runSchedulerWaitConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	go func() { _ = ctrl.Run(ctx, schedTurn(turn+1, arm)) }()
-	if err := waitForRefusal(sink, arm); err != nil {
-		return err
-	}
-	if arm == armWaitTransition {
-		if err := releaseAndSettle(sink, prov); err != nil {
-			return err
-		}
-	}
-	if err := writeObservation(root, capture("construct", arm, bootSystem, ctrl, sink, root)); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// releaseAndSettle frees the run that filled total capacity, then waits for the
-// graph to show it gone. What the refused item's cause does after that is the
-// arm's question: the ceiling it named has room, and the writer ceiling that
-// now holds it back was never reported.
-func releaseAndSettle(sink *graphSink, prov *scripted) error {
-	prov.releaseOnce()
-	deadline := time.Now().Add(60 * time.Second)
-	for time.Now().Before(deadline) {
-		graph, _ := sink.snapshot()
-		if settledWorkers(graph) > 0 && len(waitingWorkers(graph, agentgraph.WaitSlots)) > 0 {
-			// Give any second report a chance to arrive before reading.
-			time.Sleep(500 * time.Millisecond)
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected("capacity freed while a refusal still stands", waitStates(graph))
-}
-
-func settledWorkers(g agentgraph.Graph) int {
-	n := 0
-	for _, node := range g.Nodes {
-		if node.Kind != agentgraph.KindGroup && node.State != "" && node.State.Terminal() {
-			n++
-		}
-	}
-	return n
-}
-
-// schedulerWaitRows report what a restart inherits about an item the scheduler
-// refused. The classification is already right — STARTED settled that — so the
-// question here is narrower: can the host still say why it never started?
-func schedulerWaitRows(arm string, before, after Observation) []row {
-	rows := []row{refusalRow(arm, before, after)}
-	if arm == armWaitTransition {
-		rows = append(rows, capacityFreedRow(before))
-	}
-	return append(rows,
-		waitSeriesRow(arm, before),
-		executionKindRow(before, after),
-		waitProvenanceRow(before, after),
-		schedulerGateRow(before, after),
-	)
-}
-
-// capacityFreedRow is the transition arm's evidence. It freed the ceiling the
-// refusal named and the item stayed queued, held by a different one — so what
-// the reported cause does next is a question about the contract, not a race.
-func capacityFreedRow(before Observation) row {
-	return row{
-		Semantic: "the ceiling the refusal named, once freed", Authority: "the arm, through a child it releases",
-		Artifact: "none", Reconstruction: "an item settles while the refused one stays queued",
-		Before: fmt.Sprintf("%d settled, still queued: %s", before.SettledWorkers, orNone(join(before.QueuedWorkers))),
-		After:  "not applicable: the scheduler is gone", Verdict: verdictStable,
-	}
-}
-
-// refusalRow is the arm's own control: it reached the refusal it names, and no
-// earlier one. Reaching a different cause means an upper check intercepted it,
-// and the arm measured someone else's question.
-func refusalRow(arm string, before, after Observation) row {
-	want := string(wantedCause(arm))
-	verdict := verdictHolds
-	if before.WaitCauses[want] == 0 {
-		verdict = verdictNotMeasured
-	}
-	return row{
-		Semantic: "the refusal this arm reached", Authority: "SubagentScheduler.canStartLocked",
-		Artifact: "none: reported once through OnQueued", Reconstruction: "slots, then writers, then claim",
-		Before: causeSummary(before), After: causeSummary(after), Verdict: verdict,
-	}
-}
-
-// waitSeriesRow is what the transition arm exists for: how many times a cause
-// was published for the item that never started. Once means the answer is the
-// cause that queued it, not the one that held it last.
-func waitSeriesRow(arm string, before Observation) row {
-	return row{
-		Semantic: "how often a wait cause was reported", Authority: "OnQueued, which fires once per request",
-		Artifact: "none", Reconstruction: causeContract(before),
-		Before: seriesSummary(before), After: "not applicable: the scheduler is gone",
-		Verdict: verdictStable,
-	}
-}
-
-// causeContract states what the observed reporting says the cause means. It is
-// read off the series rather than assumed: one report can only be the cause
-// that queued the request, and more than one would track the current blocker.
-func causeContract(o Observation) string {
-	for _, seq := range o.WaitSeries {
-		if len(seq) > 1 {
-			return "reported again as the blocker changes: the cause is the current one"
-		}
-	}
-	return "reported once: the cause is the one that queued the request, not the one holding it now"
-}
-
-// waitProvenanceRow is what the batch was built to expose and then close. The
-// refusal exists only at the moment the scheduler makes it, so it is recorded
-// there; this compares what the graph showed against what a later process can
-// still read back.
-func waitProvenanceRow(before, after Observation) row {
-	return valueRow("why the item never started", "the execution journal's queue records",
-		"<stem>.execution.jsonl", "the cause the scheduler gave when it first refused admission",
-		causeSummary(before), journalCauses(after), true)
-}
-
-// journalCauses is the refusal a restart reads back, counted the way the graph
-// counts it so the two sides of the row are the same measurement.
-func journalCauses(o Observation) string {
-	counts := map[string]int{}
-	for _, e := range o.Executions {
-		if e.Queued() {
-			counts[e.Cause]++
-		}
-	}
-	if len(counts) == 0 {
-		return ""
-	}
-	keys := make([]string, 0, len(counts))
-	for k := range counts {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
-	parts := make([]string, 0, len(keys))
-	for _, k := range keys {
-		parts = append(parts, fmt.Sprintf("%s=%d", k, counts[k]))
-	}
-	return strings.Join(parts, " ")
-}
-
-// schedulerGateRow is the other half of what a queue record proves. An item
-// only reaches the scheduler once its dependencies are answered, so a queued
-// entry is durable evidence it crossed that gate — the one thing a dependency
-// edge alone could never say.
-func schedulerGateRow(before, after Observation) row {
-	return valueRow("items proven to have reached the scheduler", "the execution journal's queue records",
-		"<stem>.execution.jsonl", "a queue record implies the dependency gate was crossed",
-		queuedIDs(before), queuedIDs(after), true)
-}
-
-func queuedIDs(o Observation) string {
-	var out []string
-	for _, e := range o.Executions {
-		if e.Queued() {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func causeSummary(o Observation) string {
-	if len(o.WaitCauses) == 0 {
-		return ""
-	}
-	keys := make([]string, 0, len(o.WaitCauses))
-	for k := range o.WaitCauses {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
-	parts := make([]string, 0, len(keys))
-	for _, k := range keys {
-		parts = append(parts, fmt.Sprintf("%s=%d", k, o.WaitCauses[k]))
-	}
-	return strings.Join(parts, " ")
-}
-
-func seriesSummary(o Observation) string {
-	if len(o.WaitSeries) == 0 {
-		return "none"
-	}
-	ids := make([]string, 0, len(o.WaitSeries))
-	for id := range o.WaitSeries {
-		ids = append(ids, id)
-	}
-	sort.Strings(ids)
-	out := make([]string, 0, len(ids))
-	for _, id := range ids {
-		out = append(out, id+":["+strings.Join(o.WaitSeries[id], "->")+"]")
-	}
-	return join(out)
-}
-
-// schedulerArmInvalid refuses an arm that did not reach its own refusal. Each
-// arm exists to isolate one cause, and one that reached a different cause was
-// intercepted by a check above it — reporting that as this arm's answer is how
-// a probe measures the wrong ceiling and calls it evidence.
-func schedulerArmInvalid(arm string, before Observation) string {
-	want := string(wantedCause(arm))
-	if before.WaitCauses[want] == 0 {
-		return "no item was refused under " + want + "; the arm reached " + orNone(causeSummary(before))
-	}
-	for cause := range before.WaitCauses {
-		if cause != want {
-			return "an item was also refused under " + cause + ", so this arm does not isolate " + want
-		}
-	}
-	return ""
-}
@@ -1,327 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"regexp"
-	"slices"
-	"strconv"
-	"strings"
-	"sync"
-	"sync/atomic"
-
-	"reasonix/internal/contract/provider"
-)
-
-// newScripted builds the provider with its synchronisation channels open. The
-// scheduler arms coordinate through them; every other arm never reads them.
-func newScripted(arm string) *scripted {
-	return &scripted{arm: arm, held: make(chan struct{}), release: make(chan struct{})}
-}
-
-// probeModelRef is both the catalog ref and the model name boot is given, so
-// syntheticEntryFromResolver matches it without touching any config file.
-const probeModelRef = "probe/scripted"
-
-// probeAltModelRef is a second identity the same scripted provider answers to.
-// Without one, every layer of the model resolution collapses onto the same
-// string and a probe cannot tell which layer produced it.
-const probeAltModelRef = "probe/alt"
-
-// scripted answers completions from a fixed script so the construct phase can
-// establish host state without a network call. A request carrying no tools is a
-// host-internal completion (the compaction summary); everything else is a turn.
-type scripted struct {
-	// arm is which run this provider is scripting. Only the fan-out arms read
-	// it: the pair fleet is the same dispatch either way, and the arm decides
-	// whether its second item finishes or is still executing at death.
-	arm       string
-	turnCalls atomic.Int64
-	// retodo latches the one reply that replaces the task list. The sentinel
-	// stays in history after the call, so without it every later round would
-	// rewrite the list again.
-	retodo atomic.Bool
-	// asked latches the one reply that opens a question and, in the same round,
-	// tries to write. Calling ask ends the round, so the write is the effect a
-	// decision is holding back.
-	asked atomic.Bool
-	// Which fleets have been opened. A fleet's aggregate returns into the same
-	// round that dispatched it, so without a latch the reply that follows reads
-	// its own sentinel off the still-current user turn and dispatches again.
-	fleets dispatched
-	// opened says the arm's own fan-out has gone out, whichever sentinel opened
-	// it: the terminal and derive arms then hold the turn rather than replying,
-	// so the process still dies inside one.
-	opened atomic.Bool
-	// The scheduler arms dispatch two fleets and must not race them. held is
-	// closed by the holder's own child, which only runs once its slot is
-	// granted; release lets the arm free capacity while a refusal stands.
-	holding  sync.Once
-	released sync.Once
-	held     chan struct{}
-	release  chan struct{}
-}
-
-// dispatched is the set of fleets already opened, keyed by the sentinel that
-// opened each. One named state rather than a latch per fleet: they share a
-// lifetime and a meaning, and every new fan-out was adding another flag.
-type dispatched struct {
-	mu   sync.Mutex
-	seen map[string]bool
-}
-
-// held reports whether something already recorded this name, without claiming
-// it. The lone-delegation arms read it to ask a negative: that a child the
-// scheduler should be holding back never reached the provider at all.
-func (d *dispatched) held(name string) bool {
-	d.mu.Lock()
-	defer d.mu.Unlock()
-	return d.seen[name]
-}
-
-// first reports whether this is the sentinel's first dispatch, and records it.
-func (d *dispatched) first(sentinel string) bool {
-	d.mu.Lock()
-	defer d.mu.Unlock()
-	if d.seen[sentinel] {
-		return false
-	}
-	if d.seen == nil {
-		d.seen = map[string]bool{}
-	}
-	d.seen[sentinel] = true
-	return true
-}
-
-// releaseOnce frees the child waiting on release. Idempotent: an arm may reach
-// the moment to release through more than one path.
-func (s *scripted) releaseOnce() { s.released.Do(func() { close(s.release) }) }
-
-func (s *scripted) Name() string { return "probe" }
-
-func (s *scripted) Stream(ctx context.Context, req provider.Request) (<-chan provider.Chunk, error) {
-	ch := make(chan provider.Chunk, 8)
-	go func() {
-		defer close(ch)
-		for _, c := range s.script(ctx, req) {
-			select {
-			case ch <- c:
-			case <-ctx.Done():
-				return
-			}
-		}
-	}()
-	return ch, nil
-}
-
-func (s *scripted) script(ctx context.Context, req provider.Request) []provider.Chunk {
-	if len(req.Tools) == 0 {
-		return append(text(summaryBody), done())
-	}
-	// A delegated run answers first: the parent's transcript quotes every fleet
-	// argument, so any later branch that scanned the conversation would make
-	// the parent answer as one of its own children.
-	if sentinel := childSentinel(req); sentinel != "" {
-		return s.childScript(ctx, sentinel)
-	}
-	if chunks, ok := s.fanOut(req); ok {
-		return chunks
-	}
-	if s.opens(req, sleepSentinel) {
-		return append(sleepCall(), done())
-	}
-	if s.opensRun(req, taskSentinel) {
-		if queuedTaskArm(s.arm) {
-			// The holder's own child says when the ceiling is occupied.
-			// Dispatching before that lets the delegation win the race and start.
-			<-s.held
-		}
-		return append(loneTaskCall(s.arm), done())
-	}
-	if s.opensRun(req, ephemeralSentinel) {
-		return append(ephemeralCall(s.arm), done())
-	}
-	if s.opensRun(req, skillSentinel) {
-		if queuedSkillArm(s.arm) {
-			// The holder's own child says when the ceiling is occupied.
-			// Dispatching before that lets the skill win the race and start.
-			<-s.held
-		}
-		return append(skillCall(s.arm), done())
-	}
-	// The one arm whose turn has to close signs the host's steps off first, one
-	// per round, because a turn that called a tool cannot deliver while the list
-	// is open. The latch is read last: a round that could not answer keeps it.
-	if closesItsTurn(s.arm) && s.opened.Load() && hasTool(req.Tools, "complete_step") {
-		for _, id := range probeStepIDs() {
-			if s.fleets.first("sign-" + id) {
-				return append(signStep(id), done())
-			}
-		}
-	}
-	// A terminal arm needs its fan-out to finish and its turn to stay open: a
-	// closed turn would settle the marker, and the process would no longer be
-	// dying inside one.
-	if (terminalArm(s.arm) || deriveArm(s.arm)) && s.opened.Load() {
-		<-ctx.Done()
-		return nil
-	}
-	if s.turnCalls.Add(1) == 1 && hasTool(req.Tools, "todo_write") {
-		return append(todoCall(firstTodos()), done())
-	}
-	if hasTool(req.Tools, "todo_write") && requestMentions(req, retodoSentinel) && !s.retodo.Swap(true) {
-		return append(todoCall(secondTodos()), done())
-	}
-	if requestMentions(req, askSentinel) && !s.asked.Swap(true) {
-		return append(askAndWrite(), done())
-	}
-	return append(text(echoMarker(req)+turnBody), done())
-}
-
-// A turn is tagged twice: the user side by markerPattern, the assistant reply
-// by echoPattern. A fold folds assistant work first, so a one-sided tag would
-// let the surviving user turn hide the reply that was dropped beside it.
-var (
-	markerPattern = regexp.MustCompile(`PROBE-MARK-\d+`)
-	echoPattern   = regexp.MustCompile(`PROBE-ECHO-\d+`)
-)
-
-func marker(n int) string { return fmt.Sprintf("PROBE-MARK-%03d", n) }
-
-// markersIn returns the tags a message set still shows, in order.
-func markersIn(texts []string, pat *regexp.Regexp) []string {
-	var out []string
-	for _, t := range texts {
-		out = append(out, pat.FindAllString(t, -1)...)
-	}
-	return out
-}
-
-// echoMarker answers a turn with the assistant-side tag for the same number.
-func echoMarker(req provider.Request) string {
-	for _, m := range slices.Backward(req.Messages) {
-		found := markerPattern.FindAllString(m.Content, -1)
-		if len(found) == 0 {
-			continue
-		}
-		return "PROBE-ECHO-" + found[len(found)-1][len("PROBE-MARK-"):] + " reply. "
-	}
-	return ""
-}
-
-func hasTool(schemas []provider.ToolSchema, name string) bool {
-	for _, t := range schemas {
-		if t.Name == name {
-			return true
-		}
-	}
-	return false
-}
-
-func text(s string) []provider.Chunk {
-	return []provider.Chunk{{Type: provider.ChunkText, Text: s}}
-}
-
-func done() provider.Chunk { return provider.Chunk{Type: provider.ChunkDone} }
-
-// todoCall writes the list the resume phase looks for. The step ids are the
-// point: they are the host identity a completion has to cite. A first write
-// may not open with a completed item — the host rejects that shape, and a
-// rejected call would leave the row unmeasured rather than answered.
-func todoCall(todos []map[string]any) []provider.Chunk {
-	args, _ := json.Marshal(map[string]any{"todos": todos})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: "probe_todo_" + strconv.Itoa(len(todos)), Name: "todo_write", Arguments: string(args)},
-	}}
-}
-
-// retodoSentinel asks the scripted reply to replace the task list, so a probe
-// can move host step identity from one set of ids to another.
-const retodoSentinel = "PROBE-RETODO"
-
-// askSentinel asks for a question the host must put to a person, alongside the
-// write that question is holding back.
-const (
-	askSentinel      = "PROBE-ASK"
-	deferredEffect   = "probe-deferred-effect.txt"
-	deferredContents = "this write was not held back by the decision"
-)
-
-// askAndWrite opens a question and asks to write in the same round. The round
-// ends at the question, so the write must not run — not while the answer is
-// pending, and not because a later process reloaded the session.
-func askAndWrite() []provider.Chunk {
-	ask, _ := json.Marshal(map[string]any{"questions": []map[string]any{{
-		"header":   "Boundary",
-		"question": "Which side of the fold should the probe measure?",
-		"reason":   "user_decision",
-		"options": []map[string]any{
-			{"label": "Below the fold", "description": "removes folded history"},
-			{"label": "Above the fold", "description": "leaves coverage intact"},
-		},
-	}}})
-	write, _ := json.Marshal(map[string]string{"path": deferredEffect, "content": deferredContents})
-	return []provider.Chunk{
-		{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "probe_ask_1", Name: "ask", Arguments: string(ask)}},
-		{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "probe_write_1", Name: "write_file", Arguments: string(write)}},
-	}
-}
-
-func requestMentions(req provider.Request, needle string) bool {
-	for _, m := range slices.Backward(req.Messages) {
-		if strings.Contains(m.Content, needle) {
-			return true
-		}
-	}
-	return false
-}
-
-func firstTodos() []map[string]any {
-	return []map[string]any{
-		{"content": "Establish probe state", "status": "in_progress", "activeForm": "Establishing probe state", "step_id": "probe_step_01"},
-		{"content": "Cross the process boundary", "status": "pending", "step_id": "probe_step_02"},
-		{"content": "Report the resurrection matrix", "status": "pending", "step_id": "probe_step_03"},
-	}
-}
-
-// secondTodos replaces the pending work while the in_progress item stays: the
-// host refuses to drop an item that is in flight, so this is what a list
-// rewrite actually looks like. The two new ids are what a stale note misses.
-func secondTodos() []map[string]any {
-	return []map[string]any{
-		{"content": "Establish probe state", "status": "in_progress", "activeForm": "Establishing probe state", "step_id": "probe_step_01"},
-		{"content": "Split the frozen body", "status": "pending", "step_id": "probe_step_21"},
-		{"content": "Hold the ordering contract", "status": "pending", "step_id": "probe_step_22"},
-	}
-}
-
-// turnBody is fat on purpose: the fold keeps a recent verbatim tail, so thin
-// replies leave no region between the pinned head and that tail.
-var turnBody = "Probe state established. " + strings.Repeat("The probe records deterministic filler so the fold has material to work on. ", 60)
-
-var summaryBody = strings.Join([]string{
-	"## Standing facts & constraints",
-	"The runtime-resume probe is measuring what survives a process boundary.",
-	"",
-	"## Goal",
-	"Establish host state, exit, and observe what a new process can still prove.",
-}, "\n")
-
-func resolverFor(prov *scripted) *provider.StaticResolver {
-	return &provider.StaticResolver{
-		Descriptors: []provider.Descriptor{{
-			Ref: probeModelRef, DisplayName: "probe", Model: "scripted",
-			ContextWindow: 128_000, Tools: true,
-		}, {
-			Ref: probeAltModelRef, DisplayName: "probe-alt", Model: "scripted-alt",
-			ContextWindow: 128_000, Tools: true,
-		}},
-		Providers: map[string]provider.Provider{probeModelRef: prov, probeAltModelRef: prov},
-	}
-}
-
-func errUnexpected(what string, got any) error {
-	return fmt.Errorf("probe could not establish %s (got %v); the arm measures nothing and is reported invalid", what, got)
-}
@@ -1,38 +0,0 @@
-package main
-
-import (
-	"strings"
-	"testing"
-)
-
-// Sentinels are matched by substring, so one contained in another dispatches on
-// the same prompt: the identity arm silently ran the cancellation fan-out for as
-// long as PROBE-PARALLEL sat inside PROBE-PARALLEL-IDENTITY, and its rows read a
-// scenario nobody wrote. Care did not catch it — the child sentinels carry a
-// comment about exactly this hazard and the parent ones had no guard at all.
-func TestNoSentinelContainsAnother(t *testing.T) {
-	sentinels := map[string]string{
-		"fleetWarm": fleetWarmSentinel, "fleetMixed": fleetMixedSentinel,
-		"fleetPair": fleetPairSentinel, "fleetHolder": fleetHolderSentinel,
-		"fleetRefused": fleetRefusedSentinel, "fleetTerminal": fleetTerminalSentinel,
-		"fleetOutcomes": fleetOutcomesSentinel, "fleetDerive": fleetDeriveSentinel,
-		"fleetIdentity": fleetIdentitySentinel, "fleetActive": fleetActiveSentinel,
-		"fleetSettled": fleetSettledSentinel, "fleetLive": fleetLiveSentinel,
-		"parallel": parallelSentinel, "parallelIdentity": parallelIdentity,
-		"task": taskSentinel, "sleep": sleepSentinel,
-		"ask": askSentinel, "retodo": retodoSentinel,
-		"childDone": childDone, "childHang": childHang, "childFail": childFail,
-		"childHold": childHold, "childRelease": childRelease, "childFailLate": childFailLate,
-	}
-	for outerName, outer := range sentinels {
-		for innerName, inner := range sentinels {
-			if outerName == innerName {
-				continue
-			}
-			if strings.Contains(outer, inner) {
-				t.Errorf("%s (%q) contains %s (%q): a prompt naming the first dispatches the second too",
-					outerName, outer, innerName, inner)
-			}
-		}
-	}
-}
@@ -1,694 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"os"
-	"path/filepath"
-	"reasonix/internal/state/sessionstore"
-	"sort"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/ext/skill"
-	"reasonix/internal/session/control"
-)
-
-// The skill-runner classification arms. They ask what kind of execution a
-// runAs=subagent skill is, because the answer decides whether it belongs to the
-// delegation class the task and fan-out arms established or is an execution
-// class of its own. Nothing here is a fix: a classification that guessed would
-// be answered by whichever plumbing happened to exist.
-const (
-	// armSkillCompleted lets a skill finish and the turn close, then asks what
-	// each layer kept. It runs both entry points — the tool a model calls and
-	// the controller's own — because their lineage differs by design and a row
-	// that read one for the other would report intent as a defect.
-	armSkillCompleted = "skill-completed"
-	// armSkillRunning kills the process with the child mid-execution: the death
-	// every other delegation arm takes.
-	armSkillRunning = "skill-running"
-	// armSkillQueued fills the ceiling first. It is the arm that says whether a
-	// skill belongs to the same admission domain as everything else, which is a
-	// different question from whether it leaves the same records.
-	armSkillQueued = "skill-queued"
-	// armSkillCancel stops a running skill through the turn that admitted it, so
-	// the stop has the owner a person's Stop has.
-	armSkillCancel = "skill-cancel"
-	// The read-only entry point is a separate surface, not a lighter version of
-	// the one above: it prepares no transcript and refuses continuation, so
-	// asking it for durability would be demanding it break its own contract.
-	// These three ask what it is instead.
-	armReadOnlySkillCompleted = "readonly-skill-completed"
-	armReadOnlySkillRunning   = "readonly-skill-running"
-	armReadOnlySkillQueued    = "readonly-skill-queued"
-)
-
-func skillArm(name string) bool {
-	switch name {
-	case armSkillCompleted, armSkillRunning, armSkillQueued, armSkillCancel:
-		return true
-	}
-	return readOnlySkillArm(name)
-}
-
-func readOnlySkillArm(name string) bool {
-	switch name {
-	case armReadOnlySkillCompleted, armReadOnlySkillRunning, armReadOnlySkillQueued:
-		return true
-	}
-	return false
-}
-
-func queuedSkillArm(name string) bool {
-	return name == armSkillQueued || name == armReadOnlySkillQueued
-}
-
-// runningSkillArm names the arms that die or are stopped with the child inside
-// the provider, which is the only state a durable record could be missing from.
-func runningSkillArm(name string) bool {
-	return name == armSkillRunning || name == armSkillCancel || name == armReadOnlySkillRunning
-}
-
-const (
-	skillSentinel = "PROBE-SKILL"
-	// skillCallID is the tool call the model makes. It is the identity the
-	// store would join to, so the arm names it rather than letting the round
-	// pick one.
-	skillCallID = "probe_skill"
-	// probeSkillName is the skill on disk. Not a built-in: those carry their own
-	// read-only promises and review contracts, and the arm needs a plain writer
-	// so the scheduler sees a claim rather than a reader.
-	probeSkillName = "probe-worker"
-	// hostSkillTask is what the controller's own entry point is asked to do. It
-	// completes, because that arm compares two lineages rather than two deaths.
-	hostSkillTask = childDone + " host-initiated skill"
-)
-
-// writeProbeSkill puts a runAs=subagent skill in the workspace, where a project
-// skill lives. It is a writer: a read-only one would clear the writer and claim
-// checks by not being subject to them, and the arm could not say whether a skill
-// is admitted the way every other writer is.
-func writeProbeSkill(workspace string) error {
-	dir := filepath.Join(workspace, ".reasonix", skill.SkillsDirname, probeSkillName)
-	if err := os.MkdirAll(dir, 0o755); err != nil {
-		return err
-	}
-	body := "---\nname: " + probeSkillName + "\n" +
-		"description: The probe's own subagent, which answers whatever task it is handed.\n" +
-		"runAs: subagent\n---\n\nAnswer the task you are given and stop.\n"
-	return os.WriteFile(filepath.Join(dir, skill.SkillFile), []byte(body), 0o644)
-}
-
-// skillCall is the dispatch under measurement, made through the tool a model
-// calls. The read-only arms enter through the other tool, which is a different
-// runner rather than a flag on this one.
-func skillCall(arm string) []provider.Chunk {
-	name := "run_skill"
-	if readOnlySkillArm(arm) {
-		name = "read_only_skill"
-	}
-	task := skillChildSentinel(arm) + " skill body"
-	args, _ := json.Marshal(map[string]any{"name": probeSkillName, "arguments": task})
-	return []provider.Chunk{{
-		Type:     provider.ChunkToolCall,
-		ToolCall: &provider.ToolCall{ID: skillCallID, Name: name, Arguments: string(args)},
-	}}
-}
-
-// skillChildSentinel is how this arm's child behaves. The completed arms need one
-// that finishes; every other arm needs one that holds until the process dies.
-func skillChildSentinel(arm string) string {
-	if arm == armSkillCompleted || arm == armReadOnlySkillCompleted {
-		return childDone
-	}
-	return childHang
-}
-
-// skillSentinels is what this arm's one prompt carries. The queued arms name the
-// holder as well: its ceiling has to be occupied before the skill asks for a
-// slot, or the refusal they are built to record never happens.
-func skillSentinels(arm string) string {
-	if queuedSkillArm(arm) {
-		return fleetHolderSentinel + " " + skillSentinel
-	}
-	return skillSentinel
-}
-
-// runSkillClassConstruct drives one skill and dies where this arm is about. The
-// completed arms let the turn close and then run the controller's own entry
-// point, so one death holds both lineages.
-func runSkillClassConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, prov *scripted, turn int) error {
-	prompt := fanOutTurn(turn+1, skillSentinels(arm))
-	completed := arm == armSkillCompleted || arm == armReadOnlySkillCompleted
-	switch {
-	case completed:
-		if err := ctrl.Run(ctx, prompt); err != nil {
-			return fmt.Errorf("skill turn: %w", err)
-		}
-		if err := runHostInitiatedSkill(ctx, ctrl, arm); err != nil {
-			return err
-		}
-	case arm == armSkillCancel:
-		// A stop belongs to whoever admitted the turn, so the arm that is going
-		// to be stopped is admitted the way a person's input is.
-		ctrl.Send(prompt)
-		if err := waitForSkill(sink, prov, arm); err != nil {
-			return err
-		}
-	default:
-		go func() { _ = ctrl.Run(ctx, prompt) }()
-		if err := waitForSkill(sink, prov, arm); err != nil {
-			return err
-		}
-	}
-	stop := ""
-	if arm == armSkillCancel {
-		stop = stopSkill(ctrl, prov, arm)
-	}
-	obs := capture("construct", arm, bootSystem, ctrl, sink, root)
-	child := skillChildSentinel(arm)
-	obs.Progress[probeChildEntered] = []string{fmt.Sprint(prov.fleets.held(childEntered(child)))}
-	obs.Progress[probeSkillChildLeft] = []string{fmt.Sprint(prov.fleets.held(childLeft(child)))}
-	obs.Progress[probeSkillChildCtxDone] = []string{fmt.Sprint(prov.fleets.held(childCtxDone(child)))}
-	obs.Progress[probeSkillStanding] = []string{skillStanding(sink, prov)}
-	if stop != "" {
-		obs.Progress[probeStopReached] = []string{stop}
-	}
-	if err := writeObservation(root, obs); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// runHostInitiatedSkill runs the same skill through the controller's own entry
-// point, which is what a slash invocation reaches. Its lineage is deliberately
-// different — the runner drops the call id for a host-initiated run — so the
-// identity row has both to compare rather than one to judge.
-func runHostInitiatedSkill(ctx context.Context, ctrl *control.Controller, arm string) error {
-	if _, err := ctrl.RunSubagentProfile(ctx, probeSkillName, hostSkillTask, readOnlySkillArm(arm)); err != nil {
-		return fmt.Errorf("host-initiated skill: %w", err)
-	}
-	return nil
-}
-
-// The records a skill's child leaves. They ride the observation because only the
-// scripted provider knows them, and a store row read without them would be
-// asking about work nothing proves ran.
-const (
-	probeSkillChildLeft    = "probe:skill-child-left"
-	probeSkillChildCtxDone = "probe:skill-child-ctxdone"
-	// probeSkillStanding is where the call itself stood: dispatched, answered,
-	// and whether the ceiling was occupied. A refusal has no event of its own,
-	// so this is the only positive evidence the admission row has.
-	probeSkillStanding = "probe:skill-standing"
-)
-
-// waitForSkill blocks until the skill stands where this arm dies. A refusal is a
-// state nothing announces, so the queued arms establish it from both sides — the
-// call went out and has not answered, the ceiling is occupied, and the skill's
-// own child never arrived — then hold still and ask again, because a child that
-// was merely slow to start would have reached the provider by then.
-func waitForSkill(sink *graphSink, prov *scripted, arm string) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		if !queuedSkillArm(arm) {
-			if prov.fleets.held(childEntered(childHang)) {
-				return nil
-			}
-		} else if skillHeldBack(sink, prov) {
-			time.Sleep(2 * time.Second)
-			if !skillHeldBack(sink, prov) {
-				return errUnexpected("a skill the scheduler was still holding back",
-					"its child ran or its call answered")
-			}
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	return errUnexpected("a skill child "+skillWanted(arm), skillStanding(sink, prov))
-}
-
-// skillHeldBack is the refusal expressed as facts rather than as an absence: the
-// ceiling is occupied, the call is out, it has not answered, and no child of its
-// own has reached the provider.
-func skillHeldBack(sink *graphSink, prov *scripted) bool {
-	_, dispatched := sink.toolDispatched(skillCallID)
-	_, answered := sink.toolResult(skillCallID)
-	return prov.fleets.held(childEntered(childHold)) && dispatched && !answered &&
-		!prov.fleets.held(childEntered(childHang))
-}
-
-func skillStanding(sink *graphSink, prov *scripted) string {
-	_, dispatched := sink.toolDispatched(skillCallID)
-	_, answered := sink.toolResult(skillCallID)
-	return fmt.Sprintf("ceiling=%t dispatched=%t answered=%t child=%t",
-		prov.fleets.held(childEntered(childHold)), dispatched, answered,
-		prov.fleets.held(childEntered(childHang)))
-}
-
-func skillWanted(arm string) string {
-	if queuedSkillArm(arm) {
-		return "held back by a full ceiling"
-	}
-	return "inside the provider"
-}
-
-// stopSkill stops the work the way a person does — the turn's own cancel — and
-// waits to see what that reaches. Nothing is demanded: whether a stop arrives is
-// what the arm measures, and an arm that failed when it did not would report its
-// own finding as a broken premise.
-func stopSkill(ctrl *control.Controller, prov *scripted, arm string) string {
-	stopped := time.Now()
-	ctrl.Cancel()
-	deadline := time.Now().Add(20 * time.Second)
-	for time.Now().Before(deadline) && ctrl.Running() {
-		time.Sleep(50 * time.Millisecond)
-	}
-	live := "gone"
-	if prov.fleets.held(childEntered(childHang)) && !prov.fleets.held(childLeft(childHang)) {
-		live = "still running"
-	}
-	return fmt.Sprintf("turn=%s work=%s ctx=%t after=%s", turnStanding(ctrl, arm), live,
-		prov.fleets.held(childCtxDone(childHang)), time.Since(stopped).Round(time.Second))
-}
-
-// skillClassRows are the seven dimensions, one row each. They are separate
-// because a skill can share one domain and not another — the whole reason this
-// is a classification and not a comparison.
-func skillClassRows(arm string, before, after Observation) []row {
-	return []row{
-		skillAdmissionRow(arm, before),
-		skillGraphRow(before),
-		skillJournalRow(before, after),
-		skillStoreRow(before, after),
-		skillCancelRow(arm, before),
-		skillRestartRow(arm, before, after),
-		skillIdentityRow(arm, before, after),
-		skillClassVerdictRow(arm, before, after),
-	}
-}
-
-const skillAuthority = "the scheduler, read through what the child did"
-
-// skillAdmissionRow is the scheduler dimension. There is no refusal hook on a
-// skill's acquire, so the refusal is read the only way it can be: the ceiling was
-// proven occupied and the child never arrived, against the sibling arm where the
-// ceiling was free and it did.
-func skillAdmissionRow(arm string, before Observation) row {
-	entered := firstProgress(before, probeChildEntered) == "true"
-	answer := "admitted: the child reached the provider"
-	verdict := verdictHolds
-	switch {
-	case queuedSkillArm(arm) && !entered:
-		answer = "refused: the ceiling was full and the child never arrived"
-	case queuedSkillArm(arm):
-		answer = "not held back: the child ran with the ceiling full"
-		verdict = verdictViolated
-	case !entered:
-		answer = "the child never arrived with capacity free"
-		verdict = verdictNotMeasured
-	}
-	return row{
-		Semantic: "does it ask the session scheduler", Authority: skillAuthority,
-		Artifact: "none observed", Reconstruction: "in-process only",
-		Before: answer + " [" + firstProgress(before, probeSkillStanding) + "]",
-		After:  "—", Verdict: verdict,
-	}
-}
-
-// skillGraphRow is the live picture. A skill that never draws one is invisible
-// to every frontend while it runs, which is a product fact rather than a
-// durability one and is why it is its own row.
-func skillGraphRow(before Observation) row {
-	var drawn []string
-	for _, n := range before.Graph.Nodes {
-		if strings.HasPrefix(n.ID, skillCallID) {
-			drawn = append(drawn, n.ID+":"+string(n.State))
-		}
-	}
-	sort.Strings(drawn)
-	// The whole graph's size rides along: "the filter found nothing" and "there
-	// was nothing to find" are different readings, and only the second is what
-	// this row means to report.
-	return row{
-		Semantic: "does the live graph hold it", Authority: liveAuthority,
-		Artifact: "none observed", Reconstruction: "in-process only",
-		Before: fmt.Sprintf("%s (graph held %d node(s))", orNone(join(drawn)), len(before.Graph.Nodes)),
-		After:  "—", Verdict: presenceVerdict(len(drawn) > 0),
-	}
-}
-
-// presenceVerdict reports what was found without calling absence a defect. What
-// a skill owes each layer is the question; a row that answered it here would be
-// asserting the classification it exists to establish.
-func presenceVerdict(present bool) string {
-	if present {
-		return verdictHolds
-	}
-	return verdictAbsent
-}
-
-const verdictAbsent = "absent"
-
-// skillJournalRow reports what the journal holds. It does not go through the
-// ordinary value comparison: two empty columns are equal, and calling that
-// persisted would report an execution nothing recorded as durably recorded —
-// which is the substitution this whole probe exists to refuse.
-func skillJournalRow(before, after Observation) row {
-	return presenceRow("does the journal record it", journalAuthority, "<stem>.execution.jsonl",
-		"ExecutionHistory over the whole session", skillExecutions(before), skillExecutions(after))
-}
-
-// presenceRow answers "is anything here" before it answers "is it the same".
-// Absence on both sides is absence, never agreement.
-func presenceRow(semantic, authority, artifact, reconstruction string, before, after []string) row {
-	r := row{Semantic: semantic, Authority: authority, Artifact: artifact,
-		Reconstruction: reconstruction, Before: orNone(join(before)), After: orNone(join(after))}
-	switch {
-	case len(before) == 0 && len(after) == 0:
-		r.Verdict = verdictAbsent
-	case len(after) == 0:
-		r.Verdict = verdictLost
-	case r.Before == r.After:
-		r.Verdict = verdictPersisted
-	default:
-		r.Verdict = verdictLossy
-	}
-	return r
-}
-
-// skillExecutions are the journal entries this arm's skill could be recorded
-// under. The whole journal is read rather than a prefix: a skill records nothing
-// today, so a filter written around an id it never wrote would report absence by
-// construction.
-func skillExecutions(o Observation) []string {
-	var out []string
-	for _, e := range o.Executions {
-		if strings.HasPrefix(e.ID, skillCallID) || strings.Contains(e.Name, probeSkillName) {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// skillStoreRow is the durable half. The two entry points differ here by design:
-// one prepares a transcript, the other refuses to.
-func skillStoreRow(before, after Observation) row {
-	return presenceRow("what the store holds for it", storeAuthority, "subagents/<ref>.meta.json",
-		"ListSubagentsByParent over the transcript stem",
-		skillChildren(before), skillChildren(after))
-}
-
-// skillChildren are the store's records of skill children, by the kind the
-// runner declared rather than by a name match: kind is what the store was told
-// this execution is.
-func skillChildren(o Observation) []string {
-	var out []string
-	for _, f := range o.Children.Facts {
-		if f.Kind == "skill" {
-			out = append(out, f.Name+"/"+orDash(f.ParentToolCallID)+"="+f.Status)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// skillCancelRow is what a stop reached. Three facts, kept apart: a context that
-// closed and a call that returned are different events, and work whose caller
-// stopped waiting leaves the same absence behind as work that was cancelled.
-func skillCancelRow(arm string, before Observation) row {
-	if arm != armSkillCancel {
-		return row{
-			Semantic: "what a stop reaches", Authority: "the turn's own cancel",
-			Artifact: "none observed", Reconstruction: "in-process only",
-			Before: "not stopped in this arm", After: "—", Verdict: verdictNotMeasured,
-		}
-	}
-	return row{
-		Semantic: "what a stop reaches", Authority: "the turn's own cancel",
-		Artifact: "subagents/<ref>.meta.json", Reconstruction: "read at one instant, after the record closed",
-		Before: fmt.Sprintf("%s store=%s", firstProgress(before, probeStopReached),
-			orNone(join(skillChildren(before)))),
-		After: "—", Verdict: skillCancelVerdict(before),
-	}
-}
-
-// skillCancelVerdict asks whether the terminal the store kept is the one that
-// happened. The cause is observed, not read off a message: the child's own
-// context closed, which is what a cancellation does and what a failure does not,
-// so a record that files it as failed is naming the wrong ending.
-func skillCancelVerdict(before Observation) string {
-	cancelled := firstProgress(before, probeSkillChildCtxDone) == "true"
-	terminal := ""
-	for _, f := range before.Children.Facts {
-		if f.Kind == "skill" {
-			terminal = f.Status
-		}
-	}
-	switch {
-	case terminal == "":
-		return verdictNotMeasured
-	case cancelled && terminal != string(sessionstore.SubagentCancelled):
-		return verdictViolated
-	default:
-		return verdictHolds
-	}
-}
-
-// skillRestartRow is what the next process can say. It is separate from the
-// store row because "a record survives" and "the execution is explicable" are
-// different claims: a completed child on disk explains nothing about a run that
-// was cut.
-func skillRestartRow(arm string, before, after Observation) row {
-	stored, rebuilt, cut := skillChildren(after), skillRebuiltNodes(after), skillInterruptions(after)
-	answer := fmt.Sprintf("store=%s rebuilt=%s cut=%s",
-		orNone(join(stored)), orNone(join(rebuilt)), orNone(join(cut)))
-	return row{
-		Semantic: "what a restart can say about it", Authority: rebuildAuthority,
-		Artifact:       "<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		Reconstruction: "execgraph.Rebuild over the journal, and the store beside it",
-		Before:         skillAtDeath(arm, before), After: answer,
-		Verdict: skillRestartVerdict(stored, rebuilt, cut),
-	}
-}
-
-// skillInterruptions are the interruptions that name this skill. The whole list
-// is wider than the arm: a queued arm fills its ceiling with a fan-out, and a
-// fan-out records everything — reading its interruption as the skill's is how an
-// arm reports a gap as closed by evidence belonging to something else.
-func skillInterruptions(o Observation) []string {
-	var out []string
-	for _, e := range o.InterruptedExecutions {
-		if strings.HasPrefix(e.ID, skillCallID) || strings.Contains(e.Name, probeSkillName) {
-			out = append(out, e.ID)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// skillRestartVerdict keeps two claims apart. That the store remembers a child
-// finished is not that a restart can reconstruct the delegated execution: the
-// first is a record, the second is provenance, and an arm that folded them would
-// report a run nothing can place as recovered.
-func skillRestartVerdict(stored, rebuilt, cut []string) string {
-	switch {
-	case len(rebuilt) > 0 || len(cut) > 0:
-		return verdictHolds
-	case len(stored) > 0:
-		return verdictRecordOnly
-	default:
-		return verdictAbsent
-	}
-}
-
-// verdictRecordOnly is a child the store kept for an execution no reconstruction
-// names.
-const verdictRecordOnly = "record-only"
-
-// skillRebuiltNodes are the nodes a restart's rebuild draws for this skill.
-func skillRebuiltNodes(o Observation) []string {
-	var out []string
-	for _, n := range rebuiltGraph(o).Graph.Nodes {
-		if strings.HasPrefix(n.ID, skillCallID) || strings.Contains(n.Label, probeSkillName) {
-			out = append(out, n.ID+":"+string(n.State))
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-func skillAtDeath(arm string, before Observation) string {
-	switch {
-	case queuedSkillArm(arm):
-		return "held back, never started"
-	case firstProgress(before, probeChildEntered) != "true":
-		return "never reached the provider"
-	case skillFinished(before):
-		return "ran and returned"
-	default:
-		return "executing inside the provider"
-	}
-}
-
-// skillIdentityRow is the seventh dimension, and it reads the two entry points
-// apart. A model-invoked skill has a call to join to; the controller's own entry
-// point drops that id on purpose, so an empty one there is top-level provenance
-// rather than a lost join — reading them together is what stops a correct value
-// being reported as a defect.
-func skillIdentityRow(arm string, before, after Observation) row {
-	var got []string
-	for _, f := range after.Children.Facts {
-		if f.Kind != "skill" {
-			continue
-		}
-		got = append(got, fmt.Sprintf("%s from=%s model=%s effort=%s",
-			f.Name, orDash(f.ParentToolCallID), orDash(f.Model), orDash(f.Effort)))
-	}
-	sort.Strings(got)
-	return row{
-		Semantic: "what identity the record carries", Authority: storeAuthority,
-		Artifact:       "subagents/<ref>.meta.json",
-		Reconstruction: "the call the model made, against the entry point that made it",
-		Before:         skillLineageExpected(arm), After: orNone(join(got)),
-		Verdict: skillIdentityVerdict(after),
-	}
-}
-
-func skillLineageExpected(arm string) string {
-	if arm == armSkillCompleted || arm == armReadOnlySkillCompleted {
-		return "model-invoked joins the execution under " + skillCallID + "; host-initiated names none"
-	}
-	return "model-invoked joins the execution opened under " + skillCallID
-}
-
-// skillIdentityVerdict holds when the model-invoked record joins the execution
-// the journal opened under the call the model made. The join is checked, never
-// one spelling of it: a record's parent is the execution's own identity, so
-// comparing it to the bare call id would fail the chain that makes them
-// joinable. A host-started run has no call to hang under, so it is not held.
-func skillIdentityVerdict(after Observation) string {
-	opened := map[string]bool{}
-	for _, id := range skillExecutions(after) {
-		opened[id] = true
-	}
-	joined, records := false, 0
-	for _, f := range after.Children.Facts {
-		if f.Kind != "skill" {
-			continue
-		}
-		records++
-		if opened[f.ParentToolCallID] && strings.HasPrefix(f.ParentToolCallID, skillCallID) {
-			joined = true
-		}
-	}
-	switch {
-	case records == 0:
-		return verdictNotMeasured
-	case joined:
-		return verdictHolds
-	default:
-		return verdictViolated
-	}
-}
-
-// skillClassVerdictRow is the arm's answer in one line, and it is a conjunction
-// rather than a comparison: an execution the scheduler admitted, whose child
-// reached the provider, that no live picture, no journal and no active store
-// record names, is lost while it is happening — not merely unrecorded.
-func skillClassVerdictRow(arm string, before, after Observation) row {
-	entered := firstProgress(before, probeChildEntered) == "true"
-	graph := len(skillGraphNodes(before)) > 0
-	journal := len(skillExecutions(before)) > 0
-	stored := len(skillChildren(before)) > 0
-	answer, verdict := "not this arm's question", verdictNotMeasured
-	switch {
-	case queuedSkillArm(arm):
-		answer, verdict = skillQueuedNegative(before)
-	case readOnlySkillArm(arm) && runningSkillArm(arm) && entered:
-		answer, verdict = readOnlyClass(graph, journal, stored)
-	case runningSkillArm(arm) && entered:
-		if graph || journal || stored {
-			answer, verdict = fmt.Sprintf("named by graph=%t journal=%t store=%t", graph, journal, stored), verdictHolds
-		} else {
-			answer, verdict = "LOST-SILENT: executing, and named by no layer", verdictViolated
-		}
-	case runningSkillArm(arm):
-		answer = "the child never reached the provider"
-	}
-	return row{
-		Semantic: "an execution in flight, named by nothing", Authority: crossAuthority,
-		Artifact:       "none, which is the finding",
-		Reconstruction: "scheduler admitted + child entered + no graph, journal or store record",
-		Before:         answer, After: "—", Verdict: verdict,
-	}
-}
-
-// readOnlyClass classifies the entry point that promised no durable side
-// effects. Absence there is its contract kept, not a gap — but it still holds a
-// scheduler slot, so whether anything draws it while it does is a separate
-// answer, and a durable record would be the contract broken the other way.
-func readOnlyClass(graph, journal, stored bool) (string, string) {
-	switch {
-	case journal || stored:
-		return "an entry point that promised no durable side effects left one", verdictViolated
-	case graph:
-		return "ephemeral by contract, and drawn while it holds a slot", verdictEphemeralSeen
-	default:
-		return "ephemeral by contract, and unseen while it holds a slot", verdictEphemeralUnseen
-	}
-}
-
-// The two readings of an execution that keeps nothing. Neither is a defect: what
-// separates them is whether anyone can see the slot being held.
-const (
-	verdictEphemeralSeen   = "ephemeral-seen"
-	verdictEphemeralUnseen = "ephemeral-unseen"
-)
-
-// skillQueuedNegative is the control the running row needs, and it is about the
-// store alone. An opening is the orchestration saying work entered it, which a
-// refused item did; a store record is a child execution artifact, which it has
-// none of. Demanding silence from the journal too would forbid the very record
-// that tells a restart the item was held back rather than lost.
-func skillQueuedNegative(before Observation) (string, string) {
-	if len(skillChildren(before)) > 0 {
-		return "a child record for work the scheduler never admitted", verdictViolated
-	}
-	return "held back: opened and queued, with no child execution artifact", verdictHolds
-}
-
-func skillGraphNodes(o Observation) []string {
-	var out []string
-	for _, n := range o.Graph.Nodes {
-		if strings.HasPrefix(n.ID, skillCallID) {
-			out = append(out, n.ID)
-		}
-	}
-	return out
-}
-
-// skillArmInvalid reports the premise this arm could not establish. A skill that
-// never reached the provider measures nothing about executions in flight, and a
-// ceiling that was not occupied measures nothing about admission.
-func skillArmInvalid(arm string, before Observation) string {
-	entered := firstProgress(before, probeChildEntered) == "true"
-	switch {
-	case queuedSkillArm(arm) && entered:
-		return "the skill ran while the ceiling was meant to be full"
-	case runningSkillArm(arm) && !entered:
-		return "the skill's child never reached the provider, so nothing was executing at the death"
-	case (arm == armSkillCompleted || arm == armReadOnlySkillCompleted) && !skillFinished(before):
-		return "no skill child finished, so the arm compares nothing"
-	}
-	return ""
-}
-
-func skillFinished(before Observation) bool {
-	return firstProgress(before, probeSkillChildLeft) == "true"
-}
@@ -1,69 +0,0 @@
-package main
-
-import "fmt"
-
-// successorRows ask what happens to a recorded interruption once work carries
-// on. The derivation has no way out of "interrupted" today, so these rows are
-// there to say what that costs rather than to grade it.
-func successorRows(arm string, successor, final Observation) []row {
-	return []row{{
-		Semantic: "the interruption after work continued", Authority: "the adjudication log",
-		Artifact: "<stem>.adjudication.jsonl", Reconstruction: "still derived from an open record with no owner",
-		Before: interruptionState(successor), After: interruptionState(final), Verdict: stickiness(final),
-	}, {
-		Semantic: "what that turn's request carried", Authority: "control, projected per request",
-		Artifact: "none: derived, never stored", Reconstruction: "the interruption block, if one still stands",
-		Before: fmt.Sprintf("%d block(s)", successor.ModelSeesInterruption),
-		After:  fmt.Sprintf("%d block(s)", final.ModelSeesInterruption), Verdict: verdictStable,
-	}, {
-		Semantic: "how the barrier ended", Authority: "the adjudication log",
-		Artifact: "<stem>.adjudication.jsonl", Reconstruction: "an edge is written only for a barrier a turn received",
-		Before: journalSummary(successor), After: journalSummary(final), Verdict: verdictStable,
-	}, {
-		Semantic: "the effect the dead decision held back", Authority: "the workspace",
-		Artifact: final.Deferred.MarkerPath, Reconstruction: "no successor turn may release it",
-		Before: ranOrNot(successor), After: ranOrNot(final), Verdict: heldBack(successor, final),
-	}}
-}
-
-func interruptionState(o Observation) string {
-	if len(o.Interrupted) == 0 {
-		return "settled"
-	}
-	return fmt.Sprintf("%d still interrupted", len(o.Interrupted))
-}
-
-// stickiness is the question these arms were built to ask: does an
-// interruption outlive the work that replaced it? A successor that took it over
-// leaves nothing behind for later requests to carry.
-func stickiness(final Observation) string {
-	if len(final.Interrupted) == 0 && final.ModelSeesInterruption == 0 {
-		return "inherited"
-	}
-	if len(final.Interrupted) == 0 {
-		return "sticky context without an owner"
-	}
-	return "sticky"
-}
-
-func journalSummary(o Observation) string {
-	var out []string
-	for _, e := range o.Journal {
-		switch {
-		case e.Open():
-			out = append(out, e.ID+":open")
-		case e.SupersededBy != "":
-			out = append(out, e.ID+":"+e.Disposition+" by turn "+e.SupersededBy)
-		default:
-			out = append(out, e.ID+":"+e.Disposition)
-		}
-	}
-	return orNone(join(out))
-}
-
-func heldBack(successor, final Observation) string {
-	if successor.Deferred.Executed || final.Deferred.Executed {
-		return verdictViolated
-	}
-	return verdictHolds
-}
@@ -1,533 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"sort"
-	"strings"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/session/control"
-)
-
-// The terminal-disposition arms. SETTLED says an opening is no longer active
-// and nothing more, so these ask what that costs: two items can share every
-// durable fact and mean opposite things, and a restart that cannot tell them
-// apart has lost the answer a dependent needs, not a label a picture wanted.
-const (
-	// armTerminalAdopted reuses an earlier child's answer. It never runs and
-	// still holds a final answer, which is the one terminal that is answered
-	// without having executed.
-	armTerminalAdopted = "terminal-adopted"
-	// armTerminalSkippedDep is the same lifecycle with the opposite meaning: an
-	// item whose dependency failed, which never ran and holds no answer.
-	armTerminalSkippedDep = "terminal-skipped-dep"
-	// armTerminalCancelled is the other way an item ends without an answer. It
-	// was built to reach parallel_tasks' skipped branch and found it
-	// unreachable: running[i] is set when an item is dispatched, not when it is
-	// admitted, so a cancellation marks every unfinished item cancelled.
-	armTerminalCancelled = "terminal-cancelled"
-	// armTerminalContext asks whether a context edge has to be recorded at all,
-	// or already follows from the ordering and what the upstream answered.
-	armTerminalContext = "terminal-context"
-	// armChildTerminal compares the graph's terminal for an executed item
-	// against what the sub-agent store recorded for the same child. The store
-	// is the owner of "what happened to this child", so a distinction the graph
-	// draws and the store does not keep is one no reconstruction can recover.
-	armChildTerminal = "child-terminal"
-	// armIdentitySemantics asks what a node's model and effort mean. Two
-	// producers resolve them through different chains and the store resolves
-	// them again, so before a rebuild can be held to them, one of those has to
-	// be the fact the field carries.
-	armIdentitySemantics = "identity-semantics"
-)
-
-func terminalArm(name string) bool {
-	switch name {
-	case armTerminalAdopted, armTerminalSkippedDep, armTerminalCancelled, armTerminalContext, armChildTerminal, armIdentitySemantics:
-		return true
-	}
-	return false
-}
-
-// terminalSentinels is what this arm's one prompt carries. The adopted arm
-// needs a completed reference first, so it names the warm-up fleet too.
-func terminalSentinels(arm string) string {
-	switch arm {
-	case armTerminalAdopted:
-		return fleetWarmSentinel + " " + fleetMixedSentinel
-	case armTerminalCancelled:
-		return parallelSentinel
-	case armChildTerminal:
-		return fleetOutcomesSentinel
-	case armIdentitySemantics:
-		return fleetWarmSentinel + " " + fleetIdentitySentinel + " " + parallelIdentity
-	}
-	return fleetTerminalSentinel
-}
-
-// runTerminalConstruct lets the fan-out settle, then dies before the turn ends.
-// The settling is the point: a fan-out publishes an item's skip only in its
-// closing delta, so an arm that died mid-flight would never see one.
-func runTerminalConstruct(ctx context.Context, root armRoot, arm, bootSystem string, ctrl *control.Controller, sink *graphSink, turn int) error {
-	runCtx, cancelRun := context.WithCancel(ctx)
-	defer cancelRun()
-	go func() { _ = ctrl.Run(runCtx, fanOutTurn(turn+1, terminalSentinels(arm))) }()
-	if arm == armTerminalCancelled {
-		if err := cancelBeforeStart(sink, cancelRun); err != nil {
-			return err
-		}
-	}
-	if arm == armChildTerminal {
-		if err := cancelAfterOutcomes(sink, cancelRun); err != nil {
-			return err
-		}
-	}
-	if err := waitForTerminal(sink, arm); err != nil {
-		return err
-	}
-	if err := writeObservation(root, capture("construct", arm, bootSystem, ctrl, sink, root)); err != nil {
-		return err
-	}
-	os.Exit(0)
-	return nil
-}
-
-// cancelBeforeStart interrupts the group once one item is running and another
-// is still waiting for a slot, which is the state parallel_tasks splits on. It
-// cancels the turn's own context rather than calling Controller.Cancel: the
-// synchronous headless Run does not pass through turn admission, so the gate
-// that Cancel operates has nothing registered for it.
-func cancelBeforeStart(sink *graphSink, cancelRun context.CancelFunc) error {
-	deadline := time.Now().Add(60 * time.Second)
-	for time.Now().Before(deadline) {
-		graph, _ := sink.snapshot()
-		if runningCount(graph) > 0 && len(queuedWorkers(graph)) > 0 {
-			cancelRun()
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected("one item running and one still queued", waitStates(graph))
-}
-
-// cancelAfterOutcomes interrupts once one item has completed and another has
-// failed, with a third still running. All three terminals then come from one
-// group, so the store's record of them is compared under identical conditions.
-func cancelAfterOutcomes(sink *graphSink, cancelRun context.CancelFunc) error {
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		graph, _ := sink.snapshot()
-		if len(nodesInState(graph, agentgraph.StateCompleted)) > 0 &&
-			len(nodesInState(graph, agentgraph.StateFailed)) > 0 &&
-			runningCount(graph) > 0 {
-			cancelRun()
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected("one completed, one failed and one still running", waitStates(graph))
-}
-
-func runningCount(g agentgraph.Graph) int {
-	n := 0
-	for _, node := range g.Nodes {
-		if node.Kind != agentgraph.KindGroup && node.State == agentgraph.StateRunning {
-			n++
-		}
-	}
-	return n
-}
-
-// wantedTerminal is the disposition this arm exists to produce.
-func wantedTerminal(arm string) agentgraph.NodeState {
-	if arm == armTerminalAdopted {
-		return agentgraph.StateAdopted
-	}
-	if arm == armTerminalContext {
-		return agentgraph.StateCompleted
-	}
-	if arm == armTerminalCancelled || arm == armChildTerminal {
-		return agentgraph.StateCancelled
-	}
-	if arm == armIdentitySemantics {
-		return agentgraph.StateAdopted
-	}
-	return agentgraph.StateSkipped
-}
-
-// waitForTerminal blocks until the graph shows this arm's disposition. The
-// identity arm needs more than a state: it compares three producers, so it
-// waits until the last of them has settled something rather than the first.
-func waitForTerminal(sink *graphSink, arm string) error {
-	want := wantedTerminal(arm)
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		graph, _ := sink.snapshot()
-		if len(nodesInState(graph, want)) > 0 && identityReady(graph, arm) {
-			return nil
-		}
-		time.Sleep(50 * time.Millisecond)
-	}
-	graph, _ := sink.snapshot()
-	return errUnexpected("an item in state "+string(want), waitStates(graph))
-}
-
-// identityReady reports whether every producer the identity arm compares has
-// published a settled item. Waiting on one of them leaves the others unread.
-func identityReady(g agentgraph.Graph, arm string) bool {
-	if arm != armIdentitySemantics {
-		return true
-	}
-	settled := map[string]int{}
-	for _, n := range g.Nodes {
-		if n.Kind == agentgraph.KindGroup || n.State == "" || !n.State.Terminal() {
-			continue
-		}
-		settled[n.ParentID]++
-	}
-	return settled["probe_fleet_identity"] >= 4 && settled["probe_parallel_identity"] >= 2
-}
-
-func nodesInState(g agentgraph.Graph, want agentgraph.NodeState) []string {
-	var out []string
-	for _, n := range g.Nodes {
-		if n.Kind != agentgraph.KindGroup && n.State == want {
-			out = append(out, n.ID)
-		}
-	}
-	sort.Strings(out)
-	return out
-}
-
-// terminalRows put the two lifecycles side by side. Adopted and skipped reach
-// the same durable shape by different routes and mean opposite things to a
-// dependent, so every row here compares what the graph knew against what the
-// journal can still be asked.
-func terminalRows(arm string, before, after Observation) []row {
-	rows := []row{
-		terminalStateRow(arm, before, after),
-		terminalLifecycleRow(before, after),
-		terminalAnswerRow(before, after),
-	}
-	if arm == armTerminalAdopted {
-		rows = append(rows, adoptSourceRow(before, after))
-	}
-	if arm == armTerminalContext || arm == armTerminalSkippedDep {
-		rows = append(rows, contextDerivabilityRow(before, after))
-	}
-	if arm == armChildTerminal {
-		rows = append(rows, childOutcomeRow(before, after))
-	}
-	if arm == armIdentitySemantics {
-		rows = append(rows, identitySemanticsRow(before), identityRebuildRow(before, after))
-	}
-	return rows
-}
-
-// terminalStateRow is the arm's own control and the loss in one line: which
-// disposition the graph ended on, and whether anything still says so.
-func terminalStateRow(arm string, before, after Observation) row {
-	want := wantedTerminal(arm)
-	verdict := verdictNotMeasured
-	if len(nodesInState(graphOf(before), want)) > 0 {
-		verdict = verdictHolds
-	}
-	return row{
-		Semantic: "the disposition this arm reached", Authority: "the fan-out's closing delta",
-		Artifact: "none observed", Reconstruction: "published once, when the group ends",
-		Before: terminalSummary(before), After: terminalSummary(after), Verdict: verdict,
-	}
-}
-
-// terminalLifecycleRow is the comparison that decides whether a disposition has
-// to be recorded. Two items with the same durable lifecycle and opposite
-// meanings cannot be told apart by anything a restart reads.
-func terminalLifecycleRow(before, after Observation) row {
-	return valueRow("the durable lifecycle each item left", "the execution journal",
-		"<stem>.execution.jsonl", "opened / queued / started / settled, per item",
-		lifecycleSummary(before), lifecycleSummary(after), true)
-}
-
-// terminalAnswerRow is what a dependent actually needs. Completed and adopted
-// carry a final answer a dependent may start from; skipped and cancelled do
-// not, and the difference is the whole reason the distinction exists.
-func terminalAnswerRow(before, after Observation) row {
-	return valueRow("items holding an answer a dependent may use", "agentgraph.NodeState.Answered",
-		"none observed", "derived from the disposition, which is not recorded",
-		answeredSummary(before), answeredSummary(after), false)
-}
-
-// adoptSourceRow is the provenance an adoption carries: whose answer stood in.
-// Reuse is only visible if the source is nameable — otherwise the picture shows
-// work that cost nothing with no way to say what paid for it. The comparison is
-// the graph's edge against the journal's opening, so a record that named a
-// different source would read as a loss rather than as agreement.
-func adoptSourceRow(before, after Observation) row {
-	return valueRow("whose answer an adoption reused", "the adopt edge, recorded with the opening",
-		"<stem>.execution.jsonl", "read back with the openings, never re-derived",
-		adoptEdges(before), journalAdoptions(after), true)
-}
-
-// journalAdoptions renders the openings' recorded sources the way the graph
-// renders its edges, so the two sides of the row are the same measurement.
-func journalAdoptions(o Observation) string {
-	var out []string
-	for _, e := range o.Executions {
-		if e.AdoptedFrom != "" {
-			out = append(out, e.ID+"<-"+e.AdoptedFrom)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// contextDerivabilityRow asks whether the delivery edge follows from facts that
-// are already durable. A context edge that always accompanies an ordering edge
-// whose upstream answered is derivable; one that does not is a fact of its own.
-func contextDerivabilityRow(before, after Observation) row {
-	implied, actual := impliedContext(before), contextEdges(before)
-	verdict := verdictHolds
-	if implied != actual {
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic:       "whether a context edge follows from the durable facts",
-		Authority:      "the ordering edges and what each upstream answered",
-		Artifact:       "<stem>.execution.jsonl for the ordering; the disposition is not recorded",
-		Reconstruction: "an ordering edge whose upstream answered delivered its answer",
-		Before:         "implied " + orNone(implied) + " / actual " + orNone(actual),
-		After:          "implied " + orNone(impliedContext(after)) + " / actual: not observable",
-		Verdict:        verdict,
-	}
-}
-
-// executedTerminals are the graph's terminals for items that actually ran.
-// Adopted and skipped are excluded: nothing executed under them, so the store
-// owes them no record and counting them would read a correct store as short.
-func executedTerminals(o Observation) map[string]int {
-	out := map[string]int{}
-	for _, n := range o.Graph.Nodes {
-		if n.Kind == agentgraph.KindGroup || n.State == "" || !n.State.Terminal() {
-			continue
-		}
-		if n.State == agentgraph.StateAdopted || n.State == agentgraph.StateSkipped {
-			continue
-		}
-		out[string(n.State)]++
-	}
-	return out
-}
-
-// childOutcomeRow compares the two owners of the same fact. The graph draws
-// completed, failed and cancelled apart; the store is what a restart reads. A
-// terminal the graph distinguishes and the store folds into another is gone for
-// good — no journal record could recover it, because the owner never kept it.
-func childOutcomeRow(before, after Observation) row {
-	graphTerminals := executedTerminals(before)
-	verdict := verdictNotMeasured
-	if len(graphTerminals) > 0 {
-		verdict = verdictHolds
-		for state := range graphTerminals {
-			if !storeKnows(after, state) {
-				verdict = verdictViolated
-			}
-		}
-	}
-	return row{
-		Semantic:       "terminal outcome: what the graph drew, what the store kept",
-		Authority:      "the graph's closing delta against SubagentStore.Meta.Status",
-		Artifact:       "subagents/<ref>.meta.json",
-		Reconstruction: "the store owns what happened to a child; every terminal it draws it must keep",
-		Before:         "graph " + countSummary(graphTerminals) + " / store " + orNone(childStatusCounts(before)),
-		After:          "store " + orNone(childStatusCounts(after)), Verdict: verdict,
-	}
-}
-
-// storeKnows reports whether any child carries this terminal as its status.
-func storeKnows(o Observation, state string) bool {
-	for _, f := range o.Children.Facts {
-		if f.Status == state {
-			return true
-		}
-	}
-	return false
-}
-
-func childStatusCounts(o Observation) string {
-	counts := map[string]int{}
-	for _, f := range o.Children.Facts {
-		counts[f.Status]++
-	}
-	if len(counts) == 0 {
-		return ""
-	}
-	return countSummary(counts)
-}
-
-// identitySemanticsRow lays the three answers beside each other: what the
-// caller named, what the graph shows, and what the store resolved. They are not
-// expected to agree — the point is which of them a node's field is reporting,
-// because a rebuild cannot be held to a field whose meaning is unsettled.
-func identitySemanticsRow(before Observation) row {
-	store := map[string]string{}
-	for _, f := range before.Children.Facts {
-		store[f.ParentToolCallID] = orNone(f.Model) + "/" + orNone(f.Effort)
-	}
-	var out []string
-	for _, n := range before.Graph.Nodes {
-		if n.Kind != agentgraph.KindWorker {
-			continue
-		}
-		out = append(out, fmt.Sprintf("%s graph=%s/%s store=%s", n.ID,
-			orNone(n.Model), orNone(n.Effort), orNone(store[n.ID])))
-	}
-	sort.Strings(out)
-	return row{
-		Semantic: "worker identity: what each layer reports", Authority: "the fan-out's spec against the store's resolution",
-		Artifact: "subagents/<ref>.meta.json", Reconstruction: "the graph resolves a subagent chain; the store resolves the final identity",
-		Before: join(out), After: "not applicable: read before the boundary", Verdict: verdictStable,
-	}
-}
-
-// identityRebuildRow is the same question asked of the rebuild: whichever of
-// those layers the field is meant to carry, does a restart get it back?
-func identityRebuildRow(before, after Observation) row {
-	return valueRow("worker identity after a restart", "execgraph.Rebuild over journal + store",
-		"<stem>.execution.jsonl + subagents/<ref>.meta.json",
-		"the journal records no worker identity; the store records the one it resolved",
-		identitySummary(graphOf(before)), identitySummary(rebuiltGraph(after).Graph), false)
-}
-
-func graphOf(o Observation) agentgraph.Graph {
-	return agentgraph.Graph{Nodes: o.Graph.Nodes, Edges: o.Graph.Edges}
-}
-
-func terminalSummary(o Observation) string {
-	if len(o.Graph.Nodes) == 0 {
-		return ""
-	}
-	counts := map[string]int{}
-	for _, n := range o.Graph.Nodes {
-		if n.Kind != agentgraph.KindGroup && n.State != "" {
-			counts[string(n.State)]++
-		}
-	}
-	return countSummary(counts)
-}
-
-// lifecycleSummary renders each item's durable transitions, which is what a
-// restart actually has. Two items printing the same line here are two items a
-// restart cannot tell apart.
-func lifecycleSummary(o Observation) string {
-	if len(o.Executions) == 0 {
-		return ""
-	}
-	out := make([]string, 0, len(o.Executions))
-	for _, e := range o.Executions {
-		marks := []string{"open"}
-		if e.Disposition != "" && e.Disposition != "pending" {
-			marks = append(marks, e.Disposition)
-		}
-		if e.Queued() {
-			marks = append(marks, "queued("+e.Cause+")")
-		}
-		if e.Started() {
-			marks = append(marks, "started")
-		}
-		if !e.SettledAt.IsZero() {
-			marks = append(marks, "settled")
-		}
-		out = append(out, e.ID+":"+strings.Join(marks, "+"))
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func answeredSummary(o Observation) string {
-	var out []string
-	for _, n := range o.Graph.Nodes {
-		if n.Kind != agentgraph.KindGroup && n.State.Answered() {
-			out = append(out, n.ID)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func adoptEdges(o Observation) string {
-	var out []string
-	for _, e := range o.Graph.Edges {
-		if e.Kind == agentgraph.Adopt {
-			out = append(out, e.To+"<-"+e.From)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func contextEdges(o Observation) string {
-	var out []string
-	for _, e := range o.Graph.Edges {
-		if e.Kind == agentgraph.Context {
-			out = append(out, e.To+"<-"+e.From)
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-// impliedContext is what the durable facts alone would predict: every ordering
-// edge whose upstream ended up answered.
-func impliedContext(o Observation) string {
-	answered := map[string]bool{}
-	for _, n := range o.Graph.Nodes {
-		if n.State.Answered() {
-			answered[n.ID] = true
-		}
-	}
-	var out []string
-	for _, e := range o.Executions {
-		for _, up := range e.DependsOn {
-			if answered[up] {
-				out = append(out, e.ID+"<-"+up)
-			}
-		}
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func countSummary(counts map[string]int) string {
-	keys := make([]string, 0, len(counts))
-	for k := range counts {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
-	parts := make([]string, 0, len(keys))
-	for _, k := range keys {
-		parts = append(parts, fmt.Sprintf("%s=%d", k, counts[k]))
-	}
-	return strings.Join(parts, " ")
-}
-
-// terminalArmInvalid refuses an arm that never reached its disposition. Every
-// row below it compares against a shape that was not established, and reporting
-// those as evidence is how a probe answers a question nobody asked.
-func terminalArmInvalid(arm string, before Observation) string {
-	want := wantedTerminal(arm)
-	if len(nodesInState(graphOf(before), want)) == 0 {
-		return "no item reached " + string(want) + "; the fan-out ended as " + orNone(terminalSummary(before))
-	}
-	if arm == armTerminalContext && contextEdges(before) == "" {
-		return "the fan-out delivered no upstream answer, so there is no context edge to derive"
-	}
-	if arm == armTerminalCancelled && len(nodesInState(graphOf(before), agentgraph.StateSkipped)) > 0 {
-		return "an item was skipped rather than cancelled; parallel_tasks' skipped branch was thought unreachable"
-	}
-	return ""
-}
@@ -1,111 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"slices"
-)
-
-// Contract verdicts for an arm that asserts a fact rather than comparing two
-// readings: the note either satisfies what todoIdentityProjection claims about
-// itself, or it does not.
-const (
-	verdictHolds    = "holds"
-	verdictViolated = "VIOLATED"
-)
-
-// todoRows judge the host's step-identity note against the three things it
-// claims: there is one of it, it rides the turn tail, and it names the identity
-// the host holds now. Each is checked on both sides of the process boundary.
-func todoRows(before, after Observation) []row {
-	return []row{
-		todoRow("host identity notes in the view", "one note when the history lost the ids, none when it kept them",
-			before, after, noteCount, exactlyOnce),
-		todoRow("where the identity note sits", "the note rides the live tail, as a turn-tail note",
-			before, after, notePlacement, ridesTheTail),
-		todoRow("the identity the note carries", "the note names the host's current step ids",
-			before, after, noteIdentity, matchesHost),
-	}
-}
-
-func todoRow(semantic, contract string, before, after Observation,
-	render func(Observation) string, holds func(Observation) bool) row {
-	verdict := verdictHolds
-	switch {
-	case before.TodoNotes.ViewLen == 0 || after.TodoNotes.ViewLen == 0:
-		verdict = verdictNotMeasured
-	case !holds(before) || !holds(after):
-		verdict = verdictViolated
-	}
-	return row{
-		Semantic: semantic, Authority: "agent todoState (host-owned)",
-		Artifact:       "none: derived for each request, stored nowhere",
-		Reconstruction: contract,
-		Before:         render(before), After: render(after), Verdict: verdict,
-	}
-}
-
-func noteCount(o Observation) string {
-	owed := "history lost the ids"
-	if o.TodoNotes.ReadableInHistory {
-		owed = "history still shows them"
-	}
-	return fmt.Sprintf("%d (%s)", o.TodoNotes.Count, owed)
-}
-
-// exactlyOnce is conditional on need: a request whose history still shows the
-// ids is owed nothing, and demanding a note there would turn the migration into
-// an unconditional append.
-func exactlyOnce(o Observation) bool {
-	if o.TodoNotes.ReadableInHistory {
-		return o.TodoNotes.Count == 0
-	}
-	return o.TodoNotes.Count == 1
-}
-
-// notePlacement reports the note's index against the end of the frozen body.
-// An index below it means the note was frozen ahead of messages the splice adds
-// after it, which is the opposite of riding the tail.
-func notePlacement(o Observation) string {
-	n := o.TodoNotes
-	if n.Count == 0 {
-		return "no note"
-	}
-	where := "after the live tail"
-	if n.Indexes[len(n.Indexes)-1] < n.BodyLen {
-		where = "inside the frozen body"
-	}
-	return fmt.Sprintf("index %v of %d, body ends at %d — %s", n.Indexes, n.ViewLen, n.BodyLen, where)
-}
-
-// ridesTheTail passes vacuously when no note is owed: placement is a claim
-// about a note that exists.
-func ridesTheTail(o Observation) bool {
-	n := o.TodoNotes
-	if n.Count == 0 {
-		return n.ReadableInHistory
-	}
-	return n.Indexes[len(n.Indexes)-1] >= n.BodyLen
-}
-
-func noteIdentity(o Observation) string {
-	n := o.TodoNotes
-	return fmt.Sprintf("note %s / host %s", join(dedupe(n.IDs)), join(n.HostIDs))
-}
-
-func matchesHost(o Observation) bool {
-	n := o.TodoNotes
-	if n.Count == 0 {
-		return n.ReadableInHistory
-	}
-	return len(n.HostIDs) > 0 && slices.Equal(dedupe(n.IDs), n.HostIDs)
-}
-
-func dedupe(in []string) []string {
-	var out []string
-	for _, s := range in {
-		if !slices.Contains(out, s) {
-			out = append(out, s)
-		}
-	}
-	return out
-}
@@ -1,356 +0,0 @@
-package main
-
-import (
-	"encoding/json"
-	"fmt"
-	"net/http"
-	"os"
-	"os/exec"
-	"path/filepath"
-	"sort"
-	"strings"
-	"sync"
-	"time"
-
-	"reasonix/internal/contract/agentgraph"
-	"reasonix/internal/contract/config"
-	"reasonix/internal/contract/event"
-	"reasonix/internal/contract/provider"
-	"reasonix/internal/frontend/serve"
-	"reasonix/internal/session/control"
-	"reasonix/internal/state/execjournal"
-)
-
-// armUIGraphMixed is the only arm that reads the frontend rather than the host.
-// Everything below it has been shown to rebuild the graph exactly; this asks
-// the question a value comparison cannot: through which door did each fact
-// reach the view? A history replayed as live deltas settles on the same state
-// and tells the user work is starting that ended in a process now dead.
-const armUIGraphMixed = "ui-graph-mixed"
-
-func uiArm(name string) bool { return name == armUIGraphMixed }
-
-// The two extra fleets this arm's turn opens. The settled one ends before the
-// death, which is the only way a skip exists at all — a fan-out publishes one in
-// its closing delta — and the live one is dispatched by the resumed process.
-const (
-	fleetSettledSentinel = "PROBE-FLEET-SETTLED"
-	fleetLiveSentinel    = "PROBE-FLEET-LIVE"
-)
-
-// envProbeSabotage names a deliberate defect to run under. The arm must go red
-// with it and green without: a probe nobody has watched fail is a decoration.
-const envProbeSabotage = "REASONIX_PROBE_SABOTAGE"
-
-// envProbeRepo carries the checkout to the child that has to build the observer.
-// A child's working directory is its own workspace, so nothing else in the
-// process can find the frontend.
-const envProbeRepo = "REASONIX_PROBE_REPO"
-
-const (
-	sabotagePublish    = "publish"
-	sabotageTrajectory = "trajectory"
-)
-
-// settledFleet runs to completion with its dependent skipped, so the death
-// inherits a terminal the mid-flight fleet cannot produce.
-func settledFleet() []provider.Chunk {
-	return fleetCall("probe_fleet_settled", []map[string]any{
-		{"id": "up", "prompt": childFail + " settled upstream", "description": "fails", "read_only": true},
-		{"id": "down", "prompt": childDone + " settled downstream", "description": "skipped by up",
-			"read_only": true, "depends_on": []string{"up"}},
-	})
-}
-
-// uiMixedFleet is the density fleet plus one item the session ceiling refuses,
-// so the death also holds a node that was queued and never admitted.
-func uiMixedFleet(adoptRef string) []provider.Chunk {
-	return fleetCall("probe_fleet_mixed", []map[string]any{
-		{"id": "m1", "prompt": childDone + " mixed one", "description": "completes", "read_only": true},
-		{"id": "m2", "prompt": childFail + " mixed two", "description": "fails", "read_only": true},
-		{"id": "m3", "prompt": childDone + " mixed three", "description": "blocked by m2",
-			"read_only": true, "depends_on": []string{"m2"}},
-		{"id": "m4", "adopt_ref": adoptRef, "description": "adopted"},
-		{"id": "m5", "prompt": childHang + " mixed five", "description": "running at death",
-			"write_paths": []string{"probe-fanout-out"}},
-		{"id": "m6", "prompt": childHang + " mixed six", "description": "queued at death", "read_only": true},
-	})
-}
-
-// liveFleet is what the resumed process dispatches while the view is watching.
-// Without it a frontend that had stopped folding deltas entirely would pass
-// every other row here.
-func liveFleet() []provider.Chunk {
-	return fleetCall("probe_fleet_live", []map[string]any{
-		{"id": "l1", "prompt": childDone + " live one", "description": "starts after the resume", "read_only": true},
-		{"id": "l2", "prompt": childDone + " live two", "description": "starts after the resume", "read_only": true},
-	})
-}
-
-// livePrompt opens that fleet from the frontend, through /submit.
-func livePrompt() string { return fanOutTurn(200, fleetLiveSentinel) }
-
-// uiBus is the broadcaster the one arm with a frontend needs beside the probe's
-// own fold: in the dying process it is what records the wire log a window would
-// have recorded, and in the resumed one it is what the client reads.
-func uiBus(arm string, sink *graphSink) (*serve.Broadcaster, event.Sink) {
-	if !uiArm(arm) {
-		return nil, sink
-	}
-	bc := serve.NewBroadcaster()
-	return bc, teeSink{to: []event.Sink{sink, bc}}
-}
-
-// recordLikeAWindow attaches the wire log the trajectory is read from. The
-// server is never served: the constructor is what subscribes the recorder, and
-// the dying process needs the record, not a socket.
-func recordLikeAWindow(ctrl control.SessionAPI, bc *serve.Broadcaster) {
-	if bc != nil {
-		_ = serve.New(ctrl, bc, config.ServeConfig{})
-	}
-}
-
-// teeSink lets one runtime's events reach both the probe's fold and the server
-// the frontend reads. Nothing production does this; the arm needs the same
-// stream on two surfaces.
-type teeSink struct{ to []event.Sink }
-
-func (t teeSink) Emit(e event.Event) {
-	for _, s := range t.to {
-		s.Emit(e)
-	}
-}
-
-// UIEntry is one state the view passed through, as the frontend's own store
-// reports it. States is what the graph showed, keyed by node.
-type UIEntry struct {
-	Phase           string            `json:"phase"`
-	Origin          string            `json:"origin"`
-	SessionID       string            `json:"sessionId"`
-	States          map[string]string `json:"states"`
-	Interruptions   []UIInterruption  `json:"interruptions"`
-	IdentityUnknown []string          `json:"identityUnknown"`
-}
-
-type UIInterruption struct {
-	Execution string `json:"execution"`
-	Kind      string `json:"kind"`
-}
-
-// UIObs is the whole observation sequence, not a final value: the defect this
-// arm exists to catch agrees with the host on every final value.
-type UIObs struct {
-	Trace []UIEntry `json:"trace,omitempty"`
-	// Deltas is what each publication the view folded named. A republished
-	// history moves no visible state and is still the kernel saying that work
-	// which ended in a dead process is happening now.
-	Deltas [][]string `json:"deltas,omitempty"`
-	Err    string     `json:"err,omitempty"`
-}
-
-// observeUI serves the resumed session on a loopback listener and runs the real
-// client against it. The server is the production one and the client is the
-// production one; only the shell between them is the probe's, which is what
-// lets the trace be read as what a user could have seen.
-func observeUI(root armRoot, ctrl control.SessionAPI, bc *serve.Broadcaster, obs Observation) UIObs {
-	ln, err := serve.ListenLoopback()
-	if err != nil {
-		return UIObs{Err: "listen: " + err.Error()}
-	}
-	defer ln.Close()
-	srv := serve.New(ctrl, bc, config.ServeConfig{})
-	handler := srv.Handler()
-	if os.Getenv(envProbeSabotage) == sabotagePublish {
-		handler = republishOnce(handler, func() { publishRebuiltGraph(bc, obs) })
-	}
-	go func() { _ = http.Serve(ln, handler) }()
-
-	bundle, err := bundleObserver(root)
-	if err != nil {
-		return UIObs{Err: err.Error()}
-	}
-	out, runErr := runObserver(bundle, serve.LoopbackOrigin(ln))
-	seen := parseTrace(out)
-	if runErr != nil && seen.Err == "" {
-		seen.Err = runErr.Error()
-	}
-	return seen
-}
-
-// bundleObserver builds the observer with the frontend's own toolchain. It is
-// the client's code, imported the way the app imports it — a hand-written stand
-// -in would be measuring the probe.
-func bundleObserver(root armRoot) (string, error) {
-	repo := strings.TrimSpace(os.Getenv(envProbeRepo))
-	if repo == "" {
-		return "", fmt.Errorf("no checkout named in %s, so the frontend could not be built", envProbeRepo)
-	}
-	dir := filepath.Join(repo, "desktop", "frontend-next")
-	esbuild := filepath.Join(dir, "node_modules", ".bin", "esbuild")
-	if _, err := os.Stat(esbuild); err != nil {
-		return "", fmt.Errorf("the frontend's toolchain is not installed: %w", err)
-	}
-	out := filepath.Join(root.Dir, "graph-observer.mjs")
-	cmd := exec.Command(esbuild, filepath.Join("src", "e2e", "graph_observer.ts"),
-		"--bundle", "--platform=node", "--format=esm", "--outfile="+out)
-	cmd.Dir = dir
-	if msg, err := cmd.CombinedOutput(); err != nil {
-		return "", fmt.Errorf("bundle the observer: %w: %s", err, strings.TrimSpace(string(msg)))
-	}
-	return out, nil
-}
-
-func runObserver(bundle, base string) ([]byte, error) {
-	cmd := exec.Command("node", bundle, "--base", base, "--prompt", livePrompt(),
-		"--sabotage", os.Getenv(envProbeSabotage))
-	cmd.Stderr = os.Stderr
-	return cmd.Output()
-}
-
-// parseTrace reads the observer's last line. Anything it printed before that is
-// the frontend's own noise, which belongs on the terminal and not in the row.
-func parseTrace(out []byte) UIObs {
-	lines := strings.Split(strings.TrimSpace(string(out)), "\n")
-	var obs UIObs
-	if len(lines) == 0 {
-		return UIObs{Err: "the observer printed nothing"}
-	}
-	if err := json.Unmarshal([]byte(lines[len(lines)-1]), &obs); err != nil {
-		return UIObs{Err: "the observer's answer did not parse: " + err.Error()}
-	}
-	return obs
-}
-
-// republishOnce fires the sabotage after a client has read the snapshot, which
-// is the only window where it can be seen at all: what a resume publishes before
-// that is numbered below the watermark the client then resumes past, and the
-// bootstrap absorbs it. The defect this models is the Studio path — a pane that
-// is already attached when a session is resumed under it.
-func republishOnce(next http.Handler, fire func()) http.Handler {
-	var once sync.Once
-	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		next.ServeHTTP(w, r)
-		if r.URL.Path == "/execution-graph" {
-			once.Do(func() { go fire() })
-		}
-	})
-}
-
-// publishRebuiltGraph is the sabotage itself: a resume that helpfully republishes
-// what it rebuilt, which is how a view gets told that finished work is starting
-// now. Production must never do this, and the arm's own rows are what says so.
-func publishRebuiltGraph(sink event.Sink, obs Observation) {
-	rebuilt := rebuiltGraph(obs)
-	for _, n := range rebuilt.Graph.Nodes {
-		sink.Emit(event.Event{Kind: event.GraphDelta, Graph: &agentgraph.Delta{Nodes: []agentgraph.Node{n}}})
-	}
-	if len(rebuilt.Graph.Edges) > 0 {
-		sink.Emit(event.Event{Kind: event.GraphDelta, Graph: &agentgraph.Delta{Edges: rebuilt.Graph.Edges}})
-	}
-	// The publication has to reach the socket before the client attaches, or the
-	// sabotage would be measuring a race instead of a defect.
-	time.Sleep(200 * time.Millisecond)
-}
-
-// historicalIDs is every node the dying process had drawn, groups included: a
-// group redrawn as running is the same lie about the same dead process.
-func historicalIDs(o Observation) map[string]bool {
-	out := map[string]bool{}
-	for _, n := range o.Graph.Nodes {
-		out[n.ID] = true
-	}
-	return out
-}
-
-// firstSight is the origin each node carried the first time the view held it.
-func firstSight(trace []UIEntry) map[string]string {
-	out := map[string]string{}
-	for _, e := range trace {
-		for id := range e.States {
-			if _, seen := out[id]; !seen {
-				out[id] = e.Origin
-			}
-		}
-	}
-	return out
-}
-
-var liveStates = map[string]bool{"pending": true, "queued": true, "running": true}
-
-// ghostSightings are the historical nodes the view ever drew as someone's work
-// in progress. One render tick is enough: the user saw it.
-func ghostSightings(trace []UIEntry, historical map[string]bool) []string {
-	seen := map[string]bool{}
-	for _, e := range trace {
-		for id, state := range e.States {
-			if historical[id] && liveStates[state] {
-				seen[id+":"+state] = true
-			}
-		}
-	}
-	return sortedKeys(seen)
-}
-
-// rebirths are nodes the view lost and introduced again. A snapshot replace is
-// not a rebirth; leaving a state and coming back inside one session is.
-func rebirths(trace []UIEntry) []string {
-	gone, twice := map[string]bool{}, map[string]bool{}
-	held := map[string]bool{}
-	for _, e := range trace {
-		if e.Phase == "loading" {
-			gone, held = map[string]bool{}, map[string]bool{}
-			continue
-		}
-		for id := range held {
-			if _, still := e.States[id]; !still {
-				gone[id] = true
-			}
-		}
-		for id := range e.States {
-			if gone[id] {
-				twice[id] = true
-			}
-		}
-		held = map[string]bool{}
-		for id := range e.States {
-			held[id] = true
-		}
-	}
-	return sortedKeys(twice)
-}
-
-func sortedKeys(m map[string]bool) []string {
-	out := make([]string, 0, len(m))
-	for k := range m {
-		out = append(out, k)
-	}
-	sort.Strings(out)
-	return out
-}
-
-// wantedInterruptions is what the host says, formatted the way the view reports
-// it. The row compares the two rather than a hand-written expectation.
-func wantedInterruptions(o Observation) string {
-	var out []string
-	for _, i := range rebuiltGraph(o).Interrupted {
-		kind := execjournal.InterruptedBeforeStart
-		if i.Started {
-			kind = execjournal.InterruptedDuringExecution
-		}
-		out = append(out, i.Execution+":"+kind)
-	}
-	sort.Strings(out)
-	return join(out)
-}
-
-func shownInterruptions(trace []UIEntry) string {
-	if len(trace) == 0 {
-		return ""
-	}
-	var out []string
-	for _, i := range trace[len(trace)-1].Interruptions {
-		out = append(out, i.Execution+":"+i.Kind)
-	}
-	sort.Strings(out)
-	return join(out)
-}
@@ -1,184 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"sort"
-)
-
-// The rows this arm exists for. Every other graph row compares values; these
-// compare doors. A frontend that rebuilt the same graph by replaying history as
-// live transitions would pass all of those and fail all of these.
-const uiAuthority = "the frontend's own store, read through the real transport"
-
-func uiRows(before, after Observation) []row {
-	ui := after.UI
-	if ui == nil || len(ui.Trace) == 0 {
-		return []row{uiRow("the view's observation sequence", "a sequence of visible states",
-			orNone(uiFailure(ui)), verdictNotMeasured)}
-	}
-	historical := historicalIDs(before)
-	first := firstSight(ui.Trace)
-	return []row{
-		uiFirstPictureRow(ui.Trace),
-		uiHistoryOriginRow(historical, first),
-		uiGhostRow(ui.Trace, historical),
-		uiInterruptionRow(ui.Trace, after),
-		uiLiveRow(historical, first),
-		uiRebirthRow(ui.Trace),
-		uiRepublishRow(historical, ui.Deltas),
-	}
-}
-
-func uiFailure(ui *UIObs) string {
-	if ui == nil {
-		return ""
-	}
-	return ui.Err
-}
-
-func uiRow(semantic, want, got, verdict string) row {
-	return row{
-		Semantic: semantic, Authority: uiAuthority, Artifact: "none — a sequence, not a value",
-		Reconstruction: want, Before: want, After: got, Verdict: verdict,
-	}
-}
-
-func held(ok bool) string {
-	if ok {
-		return verdictHolds
-	}
-	return verdictViolated
-}
-
-// uiFirstPictureRow reads the handover itself: nothing, then the authority. A
-// graph that appears from a delta was assembled out of transitions, whatever it
-// settles on afterwards.
-func uiFirstPictureRow(trace []UIEntry) row {
-	got := trace[0].Phase
-	for _, e := range trace {
-		if len(e.States) > 0 {
-			got = trace[0].Phase + " → " + e.Origin
-			break
-		}
-	}
-	return uiRow("the first graph a resumed view draws", "loading → snapshot", got, held(got == "loading → snapshot"))
-}
-
-// uiHistoryOriginRow is the arm's whole point: an execution the dead process
-// opened may enter the view only as part of an answer, never as news.
-func uiHistoryOriginRow(historical map[string]bool, first map[string]string) row {
-	var wrong []string
-	seen := 0
-	for id, origin := range first {
-		if !historical[id] {
-			continue
-		}
-		seen++
-		if origin != "snapshot" {
-			wrong = append(wrong, id+" via "+origin)
-		}
-	}
-	sort.Strings(wrong)
-	got := fmt.Sprintf("%d from the snapshot", seen)
-	if len(wrong) > 0 {
-		got = join(wrong)
-	}
-	return uiRow("how work from the dead process entered the view",
-		fmt.Sprintf("%d executions, each from the snapshot", len(historical)), got, held(len(wrong) == 0))
-}
-
-// uiGhostRow watches every frame, not the last one: a node shown as running for
-// a single render and corrected afterwards was still shown as running.
-func uiGhostRow(trace []UIEntry, historical map[string]bool) row {
-	ghosts := ghostSightings(trace, historical)
-	return uiRow("a dead run shown as work in progress",
-		"never pending, queued or running", orNone(join(ghosts)), held(len(ghosts) == 0))
-}
-
-// uiInterruptionRow compares the two lists that describe the same executions:
-// what the host derived, and what the view told the user.
-func uiInterruptionRow(trace []UIEntry, after Observation) row {
-	want, got := wantedInterruptions(after), shownInterruptions(trace)
-	return uiRow("interruptions the view names", orNone(want), orNone(got), held(want == got))
-}
-
-// uiLiveRow is the positive control. Without it a view that had stopped folding
-// deltas altogether would satisfy every row above.
-func uiLiveRow(historical map[string]bool, first map[string]string) row {
-	var wrong, fresh []string
-	for id, origin := range first {
-		if historical[id] {
-			continue
-		}
-		fresh = append(fresh, id)
-		if origin != "delta" {
-			wrong = append(wrong, id+" via "+origin)
-		}
-	}
-	sort.Strings(fresh)
-	sort.Strings(wrong)
-	switch {
-	case len(fresh) == 0:
-		return uiRow("work started after the resume", "at least one, from a delta",
-			"nothing new was ever drawn", verdictViolated)
-	case len(wrong) > 0:
-		return uiRow("work started after the resume", "each from a delta", join(wrong), verdictViolated)
-	}
-	return uiRow("work started after the resume", "each from a delta",
-		fmt.Sprintf("%d from a delta: %s", len(fresh), join(fresh)), verdictHolds)
-}
-
-// uiRebirthRow: a node the view dropped and drew again is a second birth to
-// whatever the interface hangs off an introduction, however idempotent the fold
-// underneath it was.
-func uiRebirthRow(trace []UIEntry) row {
-	twice := rebirths(trace)
-	return uiRow("a node introduced twice in one session",
-		"each node introduced once", orNone(join(twice)), held(len(twice) == 0))
-}
-
-// uiRepublishRow is the rule underneath the first-appearance one, and the only
-// one that catches a republication at a view that already holds the snapshot:
-// after a resume, no delta may name work the dead process opened. Nothing in
-// production produces one — an execution belongs to the process that opened it.
-func uiRepublishRow(historical map[string]bool, deltas [][]string) row {
-	named := map[string]bool{}
-	for _, ids := range deltas {
-		for _, id := range ids {
-			if historical[id] {
-				named[id] = true
-			}
-		}
-	}
-	old := sortedKeys(named)
-	return uiRow("deltas naming work the dead process opened",
-		"none: history is answered, never published", orNone(join(old)), held(len(old) == 0))
-}
-
-// uiArmInvalid guards the premise. The states the death has to hold are the
-// whole reason this arm can say anything about ghosts, and an observer that
-// never ran leaves every row below it describing nothing.
-func uiArmInvalid(before, after Observation) string {
-	if after.UI == nil || len(after.UI.Trace) == 0 {
-		return "the frontend was never observed: " + orNone(uiFailure(after.UI))
-	}
-	var missing []string
-	for _, state := range wantedStates(armUIGraphMixed) {
-		if !holdsState(before, string(state)) {
-			missing = append(missing, string(state))
-		}
-	}
-	if len(missing) > 0 {
-		return "the death did not hold " + join(missing) + ", so the view had nothing to get wrong about them"
-	}
-	return ""
-}
-
-func holdsState(o Observation, want string) bool {
-	for _, n := range o.Graph.Nodes {
-		if string(n.State) == want {
-			return true
-		}
-	}
-	return false
-}
@@ -1,119 +0,0 @@
-package main
-
-import "testing"
-
-// The shapes the probe exists to tell apart, written out rather than run: a
-// full arm takes three processes and a live kernel, which is the wrong place to
-// find out that a verdict was computed backwards.
-
-func historyOf(ids ...string) map[string]bool {
-	out := map[string]bool{}
-	for _, id := range ids {
-		out[id] = true
-	}
-	return out
-}
-
-func entry(phase, origin string, states map[string]string) UIEntry {
-	return UIEntry{Phase: phase, Origin: origin, States: states}
-}
-
-func TestAHistoryReadBackHoldsEveryRow(t *testing.T) {
-	historical := historyOf("g", "g/1", "g/2")
-	trace := []UIEntry{
-		entry("loading", "reset", nil),
-		entry("live", "snapshot", map[string]string{"g": "", "g/1": "completed", "g/2": "failed"}),
-		entry("live", "delta", map[string]string{"g": "", "g/1": "completed", "g/2": "failed", "n/1": "running"}),
-		entry("live", "delta", map[string]string{"g": "", "g/1": "completed", "g/2": "failed", "n/1": "completed"}),
-	}
-	first := firstSight(trace)
-	for _, r := range []row{
-		uiFirstPictureRow(trace),
-		uiHistoryOriginRow(historical, first),
-		uiGhostRow(trace, historical),
-		uiLiveRow(historical, first),
-		uiRebirthRow(trace),
-		uiRepublishRow(historical, [][]string{{"n/1"}}),
-	} {
-		if r.Verdict != verdictHolds {
-			t.Errorf("%s = %s (%s), want %s", r.Semantic, r.Verdict, r.After, verdictHolds)
-		}
-	}
-}
-
-// The defect the arm was built for: the same final state, assembled out of
-// transitions. Every value row elsewhere in this probe would still be green.
-func TestHistoryArrivingAsNewsIsCaught(t *testing.T) {
-	historical := historyOf("g/1")
-	trace := []UIEntry{
-		entry("loading", "reset", nil),
-		entry("live", "delta", map[string]string{"g/1": "running"}),
-		entry("live", "delta", map[string]string{"g/1": "completed"}),
-	}
-	first := firstSight(trace)
-	if got := uiFirstPictureRow(trace); got.Verdict != verdictViolated {
-		t.Errorf("first picture = %s, want %s", got.Verdict, verdictViolated)
-	}
-	if got := uiHistoryOriginRow(historical, first); got.Verdict != verdictViolated {
-		t.Errorf("history origin = %s, want %s", got.Verdict, verdictViolated)
-	}
-	if got := uiGhostRow(trace, historical); got.Verdict != verdictViolated {
-		t.Errorf("ghost = %s, want %s", got.Verdict, verdictViolated)
-	}
-}
-
-// A snapshot the view holds and a delta that names the same work again: the
-// state never moves, and the kernel still said finished work is happening.
-func TestRepublishedHistoryIsCaughtWithoutMovingTheState(t *testing.T) {
-	historical := historyOf("g/1")
-	trace := []UIEntry{
-		entry("loading", "reset", nil),
-		entry("live", "snapshot", map[string]string{"g/1": "completed"}),
-		entry("live", "delta", map[string]string{"g/1": "completed"}),
-	}
-	if got := uiHistoryOriginRow(historical, firstSight(trace)); got.Verdict != verdictHolds {
-		t.Fatalf("history origin = %s, want %s: the snapshot did arrive first", got.Verdict, verdictHolds)
-	}
-	if got := uiRepublishRow(historical, [][]string{{"g/1"}}); got.Verdict != verdictViolated {
-		t.Errorf("republication = %s, want %s", got.Verdict, verdictViolated)
-	}
-}
-
-// A view that stopped folding deltas would satisfy every rule above, which is
-// what the positive control is for.
-func TestAViewThatNeverMovesAgainIsCaught(t *testing.T) {
-	historical := historyOf("g/1")
-	trace := []UIEntry{
-		entry("loading", "reset", nil),
-		entry("live", "snapshot", map[string]string{"g/1": "completed"}),
-	}
-	if got := uiLiveRow(historical, firstSight(trace)); got.Verdict != verdictViolated {
-		t.Errorf("live control = %s (%s), want %s", got.Verdict, got.After, verdictViolated)
-	}
-}
-
-// A node the view dropped and drew again, which is a second introduction to
-// anything the interface hangs off one.
-func TestANodeIntroducedTwiceIsCaught(t *testing.T) {
-	trace := []UIEntry{
-		entry("live", "snapshot", map[string]string{"g/1": "completed"}),
-		entry("live", "delta", map[string]string{}),
-		entry("live", "delta", map[string]string{"g/1": "completed"}),
-	}
-	if got := uiRebirthRow(trace); got.Verdict != verdictViolated {
-		t.Errorf("rebirth = %s (%s), want %s", got.Verdict, got.After, verdictViolated)
-	}
-}
-
-// A session switch is not a rebirth: the view is told to forget, and what comes
-// back is the next conversation's first picture.
-func TestASwitchIsNotARebirth(t *testing.T) {
-	trace := []UIEntry{
-		entry("live", "snapshot", map[string]string{"g/1": "completed"}),
-		entry("loading", "reset", map[string]string{}),
-		entry("live", "snapshot", map[string]string{"g/1": "completed"}),
-	}
-	if got := uiRebirthRow(trace); got.Verdict != verdictHolds {
-		t.Errorf("rebirth across a switch = %s (%s), want %s", got.Verdict, got.After, verdictHolds)
-	}
-}
@@ -7,7 +7,7 @@ import "fmt"
 // arm costs more: did the extra agents produce verified work, or just tokens.
 func renderDelegation(results []result) string {
 	var runs, nested, childCalls, parentCalls, mutations, dupes int
-	var reports, prose, falseDone, downgrades, violations int
+	var violations int
 	var scopeHints, namedFiles, evidencePaths, discoveredPaths int
 	solved, total, childTokens := 0, 0, 0
 	for _, r := range results {
@@ -24,10 +24,6 @@ func renderDelegation(results []result) string {
 		parentCalls += r.ToolCalls - r.SubagentToolCalls
 		mutations += r.SubagentMutations
 		dupes += r.DuplicateWorkPaths
-		reports += r.CompletionReports
-		prose += r.CompletionsProsedOnly
-		falseDone += r.FalseCompletions
-		downgrades += r.CriterionDowngrades
 		violations += r.WriteScopeViolations
 		scopeHints += r.ParentScopeHints
 		namedFiles += r.ParentNamedFiles
@@ -70,11 +66,6 @@ func renderDelegation(results []result) string {
 	if dupes > 0 {
 		b += fmt.Sprintf("- **duplicate work**: %d file(s) mutated by more than one child\n", dupes)
 	}
-	closed := reports + prose
-	b += fmt.Sprintf("- closed with a checkable claim: **%d/%d** (%s)\n", reports, closed, pct(reports, closed))
-	if falseDone > 0 {
-		b += fmt.Sprintf("- **false completions**: %d run(s), %d criterion claim(s) the host refused\n", falseDone, downgrades)
-	}
 	if violations > 0 {
 		b += fmt.Sprintf("- **write-scope violations**: %d\n", violations)
 	}
@@ -27,9 +27,7 @@ func TestRenderDelegationExposesWhatDelegationCost(t *testing.T) {
 		runMetrics: runMetrics{
 			ToolCalls: 30, SubagentToolCalls: 24,
 			SubagentRuns: 3, SubagentNestedRuns: 1, SubagentMutations: 5,
-			DuplicateWorkPaths: 2,
-			CompletionReports:  2, CompletionsProsedOnly: 1,
-			FalseCompletions: 1, CriterionDowngrades: 3,
+			DuplicateWorkPaths:   2,
 			WriteScopeViolations: 1,
 		},
 	}})
@@ -38,8 +36,6 @@ func TestRenderDelegationExposesWhatDelegationCost(t *testing.T) {
 		"parent **6** tool calls",
 		"children **24**",
 		"duplicate work**: 2",
-		"checkable claim: **2/3**",
-		"false completions**: 1 run(s), 3 criterion",
 		"write-scope violations**: 1",
 	} {
 		if !strings.Contains(got, want) {
@@ -92,14 +88,11 @@ func TestRenderDelegationSaysWhenEvidenceOriginIsUnscored(t *testing.T) {
 func TestRenderDelegationStaysQuietWhenNothingWentWrong(t *testing.T) {
 	got := renderDelegation([]result{{
 		Passed:     true,
-		runMetrics: runMetrics{ToolCalls: 12, SubagentToolCalls: 8, SubagentRuns: 2, CompletionReports: 2},
+		runMetrics: runMetrics{ToolCalls: 12, SubagentToolCalls: 8, SubagentRuns: 2},
 	}})
 	for _, unwanted := range []string{"duplicate work", "false completions", "write-scope violations"} {
 		if strings.Contains(got, unwanted) {
 			t.Errorf("clean run reported %q:\n%s", unwanted, got)
 		}
 	}
-	if !strings.Contains(got, "checkable claim: **2/2**") {
-		t.Errorf("clean run lost its completion coverage:\n%s", got)
-	}
 }
@@ -243,11 +243,11 @@ func renderDiff(r diffReport) string {
 	fmt.Fprintf(&b, "| Tokens (prompt / completion) | %s / %s |\n", comma(r.m.PromptTokens), comma(r.m.CompletionTokens))
 	fmt.Fprintf(&b, "| Model calls | %d |\n", r.m.Steps)
 	fmt.Fprintf(&b, "| Cost | %s%.4f |\n", currencySym(r.m.Currency), r.m.Cost)
-	if r.m.CapabilityRoutes > 0 || r.m.CapabilitySkillInvocations > 0 || r.m.CapabilityMCPCall > 0 || r.m.ReadinessChecks > 0 {
+	if r.m.CapabilityRoutes > 0 || r.m.CapabilitySkillInvocations > 0 || r.m.CapabilityMCPCall > 0 {
 		fmt.Fprintf(&b, "| Capability routes (semantic) | %d (%d) |\n", r.m.CapabilityRoutes, r.m.CapabilitySemanticRoutes)
 		fmt.Fprintf(&b, "| Routed candidates (require / prefer / suggest / declined) | %d (%d / %d / %d / %d) |\n", r.m.CapabilityRoutedCandidates, r.m.CapabilityRoutedRequire, r.m.CapabilityRoutedPrefer, r.m.CapabilityRoutedSuggest, r.m.CapabilityDeclines)
 		fmt.Fprintf(&b, "| Skill invocations / MCP proxy calls | %d / %d |\n", r.m.CapabilitySkillInvocations, r.m.CapabilityMCPCall)
-		fmt.Fprintf(&b, "| Review blocks / readiness recoveries | %d / %d |\n", r.m.CapabilityReviewBlocks, r.m.ReadinessRecoveries)
+		fmt.Fprintf(&b, "| Review blocks | %d |\n", r.m.CapabilityReviewBlocks)
 		if r.m.CapabilityRouterCost > 0 || r.m.CapabilityRouterLatencyMs > 0 {
 			fmt.Fprintf(&b, "| Capability-router cost / latency | %s%.4f / %dms |\n", currencySym(r.m.Currency), r.m.CapabilityRouterCost, r.m.CapabilityRouterLatencyMs)
 		}
@@ -66,15 +66,11 @@ type runMetrics struct {

 	// Delegation counters mirror internal/frontend/cli.RunMetrics. They are what makes a
 	// single-agent arm comparable against a delegated one for the same model.
-	SubagentRuns          int `json:"subagent_runs,omitempty"`
-	SubagentNestedRuns    int `json:"subagent_nested_runs,omitempty"`
-	SubagentMutations     int `json:"subagent_mutations,omitempty"`
-	CompletionReports     int `json:"completion_reports,omitempty"`
-	CompletionsProsedOnly int `json:"completions_prose_only,omitempty"`
-	FalseCompletions      int `json:"false_completions,omitempty"`
-	CriterionDowngrades   int `json:"criterion_downgrades,omitempty"`
-	WriteScopeViolations  int `json:"write_scope_violations,omitempty"`
-	DuplicateWorkPaths    int `json:"duplicate_work_paths,omitempty"`
+	SubagentRuns         int `json:"subagent_runs,omitempty"`
+	SubagentNestedRuns   int `json:"subagent_nested_runs,omitempty"`
+	SubagentMutations    int `json:"subagent_mutations,omitempty"`
+	WriteScopeViolations int `json:"write_scope_violations,omitempty"`
+	DuplicateWorkPaths   int `json:"duplicate_work_paths,omitempty"`
 	// Evidence origin: what the parent's own delegation text scoped and named,
 	// and how much of what the children looked at they had to find themselves.
 	ParentScopeHints     int `json:"parent_scope_hints,omitempty"`
@@ -85,8 +81,6 @@ type runMetrics struct {
 	// graph, which is where a scheduling wait is recorded and nowhere else.
 	Fanout *fanoutMetrics `json:"fanout,omitempty"`
 	// Optional Delivery capability counters (omitempty for baseline/old metrics).
-	ReadinessChecks            int     `json:"readiness_checks,omitempty"`
-	ReadinessRecoveries        int     `json:"readiness_recoveries,omitempty"`
 	CapabilityRoutes           int     `json:"capability_routes,omitempty"`
 	CapabilityRoutedCandidates int     `json:"capability_routed_candidates,omitempty"`
 	CapabilityRoutedRequire    int     `json:"capability_routed_require,omitempty"`
@@ -129,7 +129,7 @@
 !macro rxScript
  FileOpen $0 "$PLUGINSDIR\rxclose.ps1" w
  FileWrite $0 "param([string]$$Dir, [string]$$Mode, [int]$$Self)$\r$\n"
-  FileWrite $0 "$$names = @('${APP_EXECUTABLE_FILENAME}', 'reasonix-studio-host.exe', 'cua-driver.exe')$\r$\n"
+  FileWrite $0 "$$names = @('${APP_EXECUTABLE_FILENAME}', 'reasonix-studio-host.exe')$\r$\n"
  FileWrite $0 "$$hit = @(Get-CimInstance -ClassName Win32_Process | ? { $$_.ProcessId -ne $$Self -and (($$_.Path -and $$_.Path.StartsWith($$Dir, 'CurrentCultureIgnoreCase')) -or (-not $$_.Path -and $$names -contains $$_.Name)) })$\r$\n"
  FileWrite $0 "if ($$Mode -eq 'find') { if ($$hit.Count -gt 0) { exit 0 } else { exit 1 } }$\r$\n"
  FileWrite $0 "$$hit | % { Stop-Process -Id $$_.ProcessId -Force -ErrorAction SilentlyContinue }$\r$\n"
@@ -33,7 +33,7 @@
    "launcher-live": "node test/launcher_live.mjs"
  },
  "devDependencies": {
-    "electron": "44.7.2",
+    "electron": "45.1.0",
    "electron-builder": "^26.18.0",
    "playwright": "1.64.1"
  },
@@ -0,0 +1,35 @@
+"use strict";
+const fs = require("node:fs");
+const path = require("node:path");
+
+// Exit 0 when the artifact is missing or any file under the given paths is
+// newer than it, 1 when it is up to date. run-studio.sh asks this before
+// rebuilding; the shell has no portable `find` on every platform it runs on.
+const [artifact, ...inputs] = process.argv.slice(2);
+if (!artifact) process.exit(2);
+
+let built;
+try {
+  built = fs.statSync(artifact).mtimeMs;
+} catch {
+  process.exit(0);
+}
+
+const SKIP = new Set(["node_modules", ".git", "dist"]);
+
+function newer(entry) {
+  let info;
+  try {
+    info = fs.statSync(entry);
+  } catch {
+    return false;
+  }
+  if (!info.isDirectory()) return info.mtimeMs > built;
+  for (const name of fs.readdirSync(entry)) {
+    if (SKIP.has(name)) continue;
+    if (newer(path.join(entry, name))) return true;
+  }
+  return false;
+}
+
+process.exit(inputs.some(newer) ? 0 : 1);
@@ -9,8 +9,8 @@ importers:
  .:
    devDependencies:
      electron:
-        specifier: 44.7.2
-        version: 44.7.2([email protected])
+        specifier: 45.1.0
+        version: 45.1.0([email protected])
      electron-builder:
        specifier: ^26.18.0
        version: 26.18.0([email protected])([email protected])
@@ -374,8 +374,8 @@ packages:
    resolution: {integrity: sha512-bO3y10YikuUwUuDUQRM4KfwNkKhnpVO7IPdbsrejwN9/AABJzzTQ4GeHwyzNSrVO+tEH3/Np255a3sVZpZDjvg==}
    engines: {node: '>=8.0.0'}

-  electron@44.7.2:
-    resolution: {integrity: sha512-u3Pbz0Hi1sSSwBMz2v8LiagsfUG4QURPUe/qphM+g4THKP8o4BYdnuboyhlLO8HIC/KfV5rvltyacbP4jEq9yw==}
+  electron@45.1.0:
+    resolution: {integrity: sha512-i7qKVAh/KR5v1pf56NhGGX2QYHcdjvLgS7koL6ui50c+zUDGLRunR95nXcTpHPURKOQ4GDBG8BmXiDgOQ8DD0A==}
    engines: {node: '>= 22.12.0'}
    hasBin: true

@@ -1551,7 +1551,7 @@ snapshots:
    transitivePeerDependencies:
      - supports-color

-  electron@44.7.2([email protected]):
+  electron@45.1.0([email protected]):
    dependencies:
      '@electron-internal/extract-zip': 1.0.5
      '@electron/get': 5.1.0([email protected])
@@ -12,11 +12,29 @@ if [ ! -d "$here/node_modules" ]; then
  (cd "$here" && pnpm install --frozen-lockfile)
 fi

-echo "==> frontend"
-(cd "$here/../frontend-next" && pnpm install --frozen-lockfile >/dev/null && pnpm build >/dev/null)
+# A step is skipped when nothing it reads is newer than what it produced;
+# STUDIO_REBUILD=1 forces every step.
+stale() {
+  [ -n "${STUDIO_REBUILD:-}" ] || node "$here/packaging/stale.js" "$@"
+}

-echo "==> kernel"
-(cd "$here" && pnpm build:host)
+root="$here/../.."
+web="$here/../frontend-next"
+host=$(ls "$here"/bin/reasonix-studio-host* 2>/dev/null | head -n 1 || true)
+
+if stale "$web/dist/index.html" "$web/src" "$web/public" "$web/index.html" "$web/package.json" "$web/pnpm-lock.yaml" "$web/vite.config.ts" "$web/tsconfig.json"; then
+  echo "==> frontend"
+  (cd "$web" && pnpm install --frozen-lockfile >/dev/null && pnpm build >/dev/null)
+else
+  echo "==> frontend (up to date)"
+fi
+
+if [ -z "$host" ] || stale "$host" "$root/internal" "$root/cmd" "$root/go.mod" "$root/go.sum" "$here/packaging"; then
+  echo "==> kernel"
+  (cd "$here" && pnpm build:host)
+else
+  echo "==> kernel (up to date)"
+fi

 echo "==> studio"
 cd "$here" && exec pnpm start
@@ -0,0 +1,22 @@
+const net = require("node:net");
+
+function probeHost(origin, timeoutMs = 4000) {
+  return new Promise((resolve, reject) => {
+    const url = new URL(origin);
+    const socket = net.connect({ host: url.hostname, port: Number(url.port) });
+    const fail = (cause) => {
+      socket.destroy();
+      const err = new Error(`the host at ${origin} did not accept a connection from the shell (${cause})`);
+      err.code = "unreachable";
+      reject(err);
+    };
+    socket.setTimeout(timeoutMs, () => fail(`no answer in ${timeoutMs / 1000}s`));
+    socket.once("error", (e) => fail(e.code || e.message));
+    socket.once("connect", () => {
+      socket.destroy();
+      resolve();
+    });
+  });
+}
+
+module.exports = { probeHost };
@@ -39,6 +39,7 @@ const { processMemoryHandler } = require("./procmem");
 const { captureHandler } = require("./capture");
 const { createPowerGuard, keepsAwake } = require("./powerguard");
 const { openLogs, redactArgv, failStartup } = require("./shelllog");
+const { probeHost } = require("./hostprobe");
 const { raise } = require("./raise");
 const { keepCredential, restoreIfStale } = require("./credential");
 const { actHandler } = require("./acts");
@@ -223,6 +224,7 @@ cua = await startEmbedded(computerDriver, contract, text => logs.host.raw(text))
  tray = installTray(client, { onOpen: showWindow, onQuit: () => app.quit() });
  win.on("close", onWindowClose);
  hostAgentBrowser();
+  await probeHost(origin);
  await win.loadURL(origin + PAGE_PATH);
  await tray?.refresh();
  // After the window, deliberately. This asks about an install left behind by
@@ -172,6 +172,10 @@ const HINTS = {
    "The kernel program could not be run. Security software blocking or locking it is a common cause.",
    "无法运行内核程序。安全软件拦截或占用它是常见原因。",
  ],
+  unreachable: [
+    "The kernel is running, but this program cannot connect to it. A firewall or security product that blocks inbound connections for programs without an allow rule is the usual cause on Windows: allow Reasonix Studio and its reasonix-studio-host program, then start again.",
+    "内核已在运行，但本程序连不上它。在 Windows 上，通常是防火墙或安全软件对没有放行规则的程序拦截了入站连接：请放行 Reasonix Studio 和它的 reasonix-studio-host 程序，然后重新启动。",
+  ],
 };

 function startupFailure(locale, reason, logDir, code) {
@@ -718,7 +718,7 @@ test("a kernel that never answers is started once and not retried", { skip: proc
 test("each failure cause gets its own explanation in both languages", () => {
  const { startupFailure } = require("../src/shelllog.js");
  const seen = new Set();
-  for (const code of ["timeout", "exited", "spawn"]) {
+  for (const code of ["timeout", "exited", "spawn", "unreachable"]) {
    for (const locale of ["en", "zh-CN"]) {
      const { detail } = startupFailure(locale, "r", "/logs", code);
      assert.ok(!seen.has(detail), `${code}/${locale} repeated another cause's text`);
@@ -728,6 +728,17 @@ test("each failure cause gets its own explanation in both languages", () => {
  assert.equal(startupFailure("en", "r", "/logs", "overflow").detail, "r\n\nLogs are in:\n/logs");
 });

+test("the shell finds out at once that the host it started cannot be reached", async () => {
+  const net = require("node:net");
+  const { probeHost } = require("../src/hostprobe.js");
+  const server = net.createServer((socket) => socket.end());
+  await new Promise((resolve) => server.listen(0, "127.0.0.1", resolve));
+  const { port } = server.address();
+  await probeHost(`http://127.0.0.1:${port}`, 2000);
+  await new Promise((resolve) => server.close(resolve));
+  await assert.rejects(probeHost(`http://127.0.0.1:${port}`, 2000), (err) => err.code === "unreachable");
+});
+
 // A fake child: only the stdout half readActs touches, and a way to push bytes
 // through it in whatever chunks the test wants -- which is the point, since the
 // bug this guards is a line split across two of them.
@@ -3921,3 +3932,11 @@ test("a pick arms the page, answers what it reports, and a page that cannot answ
    }
  }
 });
+
+test("the installer's by-name fallback for a process it cannot read the path of never names the shared computer-use driver", () => {
+  const nsh = fs.readFileSync(new URL("../assets/installer.nsh", import.meta.url), "utf8");
+  const listed = /\$\$names = @\(([^)]*)\)/.exec(nsh)?.[1] ?? "";
+  assert.ok(listed.includes("${APP_EXECUTABLE_FILENAME}"), "the app's own executable is matched by name");
+  assert.ok(listed.includes("reasonix-studio-host.exe"), "the shell's own host is matched by name");
+  assert.ok(!listed.includes("cua-driver.exe"), "other computer-use apps ship the same driver name, so a path-less one is not ours");
+});
@@ -36,6 +36,9 @@
    "@mantine/core": "9.7.4",
    "@mantine/hooks": "9.7.4",
    "@noble/curves": "2.4.0",
+    "@xterm/addon-fit": "^0.11.0",
+    "@xterm/addon-web-links": "^0.12.0",
+    "@xterm/xterm": "^6.0.0",
    "codemirror": "^6.0.2",
    "docx-preview": "0.4.1",
    "dompurify": "^3.4.17",
@@ -43,10 +46,10 @@
    "katex": "^0.19.1",
    "lucide-react": "^1.64.0",
    "maplibre-gl": "6.16.0",
-    "react": "^19.3.1",
-    "react-dom": "^19.3.1",
    "papaparse": "^5.7.0",
    "pdfjs-dist": "^6.4.299",
+    "react": "^19.3.1",
+    "react-dom": "^19.3.1",
    "react-map-gl": "8.1.3",
    "react-markdown": "^10.1.0",
    "recharts": "3.10.1",
@@ -43,6 +43,7 @@ node perf/rail-toggle-panes.mjs # Rail-fold frame cost with 1/3/6 panes; existin
 node perf/verify.mjs     # 行为验证：跟随、块的装卸、切页往返、大会话树
 node perf/panels.mjs     # 两侧栏：收起补间、拖动改宽、上下限、窄屏退场
 node perf/look.mjs       # 外观验证：字号、界面缩放、自定义字体、壁纸上传与调节（含焦点）
+node perf/versions-row.mjs # Version rows: download/rollback, notes, restart-ready and the 600px breakpoint in English and Chinese
 node perf/measure.mjs    # Custom conversation percentage, preset retention, narrow composer controls and populated phone layout
 node perf/theme-schedule.mjs # Scheduled appearance: daily boundaries, midnight, reload/resume, modes, active packs and DST; THEME_SCHEDULE_SHOTS writes wide/phone light/dark screenshots
 node perf/pick.mjs       # 补全菜单：/ 和 @ 的选中态，深浅两色下各走一遍
@@ -19,11 +19,11 @@ for (const scheme of ["light", "dark"]) {
  const page = await ctx.newPage();
  page.on("pageerror", (e) => fails.push("page error: " + e.message));
  await page.goto(PAGE, { waitUntil: "domcontentloaded" });
-  await page.waitForSelector('.projects-heading [data-action="workspace.add"]');
+  await page.waitForSelector('.projects-header [data-action="workspace.add"]');
  const current = () => page.evaluate(() => [...document.querySelectorAll(".wsnode[data-current]")].map((el) => el.textContent ?? ""));
  const was = await current();
  if (SHOTS) await page.screenshot({ path: join(SHOTS, `${scheme}-before.png`) });
-  await page.locator('.projects-heading [data-action="workspace.add"]').click();
+  await page.locator('.projects-header [data-action="workspace.add"]').click();
  await page.getByRole("dialog", { name: "打开项目", exact: true }).getByRole("button", { name: "选择本机文件夹", exact: true }).click();
  await page.waitForFunction(() => [...document.querySelectorAll(".wsnode[data-current]")].some((el) => el.textContent?.includes("mock-workspace")), null, { timeout: 5000 }).catch(() => {});
  const now = await current();
@@ -16,11 +16,11 @@ try {
      const page = await browser.newPage({ viewport: { width, height: 800 }, colorScheme: theme, reducedMotion: "reduce" });
      page.on("pageerror", (e) => fails.push("page error: " + e.message));
      await page.goto(BASE, { waitUntil: "networkidle" });
-      await page.waitForSelector(".projects-heading");
+      await page.waitForSelector(".projects-header");
      await page.mouse.move(width - 2, 790);
      await page.waitForTimeout(300);
      const g = await page.evaluate(() => {
-        const row = document.querySelector(".projects-heading");
+        const row = document.querySelector(".projects-header");
        const add = row?.querySelector('[data-action="workspace.add"]');
        if (!add) return null;
        const r = add.getBoundingClientRect();
@@ -43,11 +43,11 @@ for (const scheme of ["light", "dark"]) {
  });
  await page.waitForSelector(".studio-deckchip");
  await page.locator('[data-action="deck.agents"]').hover();
-  await page.locator('.studio-deck-pop .ag-open[data-target]').first().waitFor();
+  await page.locator('.studio-deck-pop .ag[data-action="agent.open"][data-target]').first().waitFor();
  const rows = await page.locator(".studio-deck-pop .ag").count();
  check(`${scheme}: the deck lists both delegations`, rows === 2, `rows=${rows}`);

-  await page.locator('.studio-deck-pop .ag', { hasText: "security-review" }).locator('[data-action="agent.open"]').click();
+  await page.locator('.studio-deck-pop .ag', { hasText: "security-review" }).click();
  await page.locator('.workbench-tab-pick', { hasText: "security-review" }).waitFor();
  await page.locator(".agentrun .txt", { hasText: "rotated on every refresh" }).waitFor();
  await page.waitForFunction(() => !document.querySelector(".agentrun")?.textContent.includes("## Finding"));
@@ -68,7 +68,7 @@ for (const scheme of ["light", "dark"]) {
  check(`${scheme}: the record stays inside the workbench column`, box.flowRight <= box.wbRight + 1 && box.flowLeft >= box.wbLeft - 1, JSON.stringify(box));

  await page.locator('[data-action="deck.agents"]').hover();
-  await page.locator('.studio-deck-pop .ag', { hasText: "other" }).locator('[data-action="agent.open"]').click();
+  await page.locator('.studio-deck-pop .ag', { hasText: "other" }).click();
  await page.locator(".agent-tx").waitFor();
  check(`${scheme}: a call with no run keeps the full-record dialog`, true);
  await page.close();
@@ -12,6 +12,8 @@ try {
      const page = await browser.newPage({ viewport: { width: 390, height: 844 }, reducedMotion: "reduce" });
      page.on("pageerror", (error) => fails.push(error.message));
      await page.goto(`${base}?pref=${language}&turns=2&git=none&zoom=${zoom}`, { waitUntil: "networkidle" });
+      const more = page.locator(".compose .turn-more");
+      if (await more.isVisible()) await more.click();
      const chip = page.locator(".studio-branch[data-norepo]");
      await chip.waitFor();
      await page.evaluate(() => document.fonts.ready);
@@ -35,7 +35,7 @@ for (const scheme of ["light", "dark"]) {
    });
    const tag = `${scheme} ${width}px`;
    if (width >= 760) {
-      check(`${tag}: the badge is on the wordmark's row, vertically centred on it`, m.shown && m.dy <= 1.5 && m.after, JSON.stringify(m));
+      check(`${tag}: the badge is on the wordmark's row, within the 3px optical drop the sidebar board draws`, m.shown && m.dy <= 3.5 && m.after, JSON.stringify(m));
      check(`${tag}: it is an outlined uppercase badge in both themes and the text is untouched`, m.fill === "rgba(0, 0, 0, 0)" && m.border !== "0px" && m.radius !== "0px" && m.upper === "uppercase" && m.text === "studio", JSON.stringify(m));
      if (!seen) seen = { left: m.left, top: m.top };
      check(`${tag}: it sits in the same place whichever theme is on`, Math.abs(m.left - seen.left) <= 0.5 && Math.abs(m.top - seen.top) <= 0.5, `${m.left},${m.top} vs ${seen.left},${seen.top}`);
@@ -14,6 +14,7 @@ const browser = await chromium.launch();
 try {
  for (const scheme of ["light", "dark"]) {
    const page = await browser.newPage({ colorScheme: scheme, viewport: { width: 1400, height: 900 }, reducedMotion: "reduce" });
+    await page.addInitScript(() => localStorage.setItem("rx-nav-rail", "on"));
    await page.goto(BASE, { waitUntil: "networkidle" });
    await page.waitForSelector(".app");
    await page.locator('.navbtn[aria-label="文件"]').first().click();
@@ -34,21 +34,23 @@ for (const scheme of ["light", "dark"]) {
  });
  await page.waitForTimeout(800);
  await page.mouse.move(995, 5);
-  const rows = await page.evaluate(() =>
-    [...document.querySelectorAll(".flow .call")]
+  const rows = await page.evaluate(() => {
+    const read = (kind, shell, hl) => {
+      const cs = getComputedStyle(shell);
+      return {
+        kind,
+        radius: cs.borderTopLeftRadius,
+        border: cs.borderTopColor,
+        bg: cs.backgroundColor,
+        head: hl ? getComputedStyle(hl).backgroundColor : "rgba(0, 0, 0, 0)",
+      };
+    };
+    const calls = [...document.querySelectorAll(".flow .call")]
      .filter((c) => ["mem", "host", "extension"].includes(c.getAttribute("data-k")) || c.hasAttribute("data-prompt"))
-      .map((c) => {
-        const cs = getComputedStyle(c.querySelector(":scope > .c"));
-        const hl = c.querySelector(":scope > .c > .hl");
-        return {
-          kind: c.getAttribute("data-k"),
-          radius: cs.borderTopLeftRadius,
-          border: cs.borderTopColor,
-          bg: cs.backgroundColor,
-          head: hl ? getComputedStyle(hl).backgroundColor : "rgba(0, 0, 0, 0)",
-        };
-      }));
-  check(`${scheme}: four kinds of card found`, rows.length === 4, rows.map((r) => r.kind).join(","));
+      .map((c) => read(c.getAttribute("data-k"), c.querySelector(":scope > .c"), c.querySelector(":scope > .c > .hl")));
+    return calls;
+  });
+  check(`${scheme}: three kinds of card found`, rows.length === 3, rows.map((r) => r.kind).join(","));
  const same = (key) => new Set(rows.map((r) => r[key])).size === 1;
  check(`${scheme}: one radius`, rows.length > 0 && same("radius"), [...new Set(rows.map((r) => r.radius))].join(" / "));
  check(`${scheme}: one border colour, whatever the state`, rows.length > 0 && same("border"), [...new Set(rows.map((r) => r.border))].join(" / "));
@@ -18,6 +18,7 @@ try {
    const ctx = await browser.newContext({ locale: "zh-CN", colorScheme: scheme, viewport: { width: 1400, height: 900 }, reducedMotion: "reduce" });
    const page = await ctx.newPage();
    page.on("pageerror", (e) => fails.push("page error: " + e.message));
+    await ctx.addInitScript(() => localStorage.setItem("rx-nav-rail", "on"));
    await page.goto(`${BASE}?pref=zh&ws=2&sess=3&turns=2&changes=${CHANGES}`, { waitUntil: "networkidle" });
    await page.waitForSelector(".app");
    await page.locator('.navbtn[aria-label="文件"]').first().click();
@@ -26,7 +27,7 @@ try {
    const rows = await page.locator(".workbench-change-row").count();
    check(`${scheme}: ${CHANGES} changes leave the page under ${NODES_MAX} nodes`, nodes < NODES_MAX, `${nodes} nodes, ${rows} rows`);
    check(`${scheme}: the list says how many it left out`, (await page.locator(".workbench-change-more").textContent())?.includes("没有显示") === true);
-    check(`${scheme}: the count in the head is every change`, (await page.locator(".workbench-changes .workbench-explorer-head small").first().textContent()) === String(CHANGES));
+    check(`${scheme}: the count in the head is every change`, (await page.locator(".workbench-changes .workbench-explorer-head small").first().textContent())?.replace(/\D/g, "") === String(CHANGES));
    await page.locator(".workbench-change-more").scrollIntoViewIfNeeded();
    if (shotDir) await page.screenshot({ path: `${shotDir}/${scheme}-changes-cap.png` });
    await page.locator('[data-action="workbench.more-changes"]').click();
@@ -13,6 +13,7 @@ const browser = await chromium.launch();
 try {
  for (const scheme of ["light", "dark"]) {
    const page = await browser.newPage({ colorScheme: scheme, viewport: { width: 1400, height: 900 } });
+    await page.addInitScript(() => localStorage.setItem("rx-nav-rail", "on"));
    await page.goto(BASE, { waitUntil: "networkidle" });
    await page.waitForSelector(".app");
    await page.locator('.navbtn[aria-label="文件"]').first().click();
@@ -13,6 +13,7 @@ const browser = await chromium.launch();
 try {
  for (const scheme of ["light", "dark"]) {
    const page = await browser.newPage({ colorScheme: scheme, viewport: { width: 1400, height: 900 } });
+    await page.addInitScript(() => localStorage.setItem("rx-nav-rail", "on"));
    await page.goto(BASE, { waitUntil: "networkidle" });
    await page.waitForSelector(".app");
    await page.locator('.navbtn[aria-label="文件"]').first().click();
@@ -11,7 +11,7 @@ const check = (name, ok, detail = "") => {

 const browser = await chromium.launch();
 for (const scheme of ["light", "dark"]) {
-  for (const [pref, label] of [["zh", "询问来自外部的文字"], ["en", "Ask about text from outside"]]) {
+  for (const [pref, label] of [["zh", "外部文字"], ["en", "Outside text"]]) {
    for (const width of [1280, 420]) {
      const tag = `${scheme} ${pref} ${width}px`;
      const page = await browser.newPage({ viewport: { width, height: 1000 }, locale: pref === "zh" ? "zh-CN" : "en-US", colorScheme: scheme, reducedMotion: "reduce" });
@@ -19,13 +19,13 @@ for (const scheme of ["light", "dark"]) {
      await page.goto(`${BASE}?pref=${pref}&turns=2`, { waitUntil: "networkidle" });
      await page.waitForSelector(".app");
      await page.evaluate(() => document.querySelector('[data-action="chrome.settings"]')?.click());
-      await page.waitForSelector(".prefs-nav button");
-      await page.evaluate(() => { const b = [...document.querySelectorAll(".prefs-nav button")].find((x) => /工具与权限|Tools/.test(x.textContent)); b?.click(); });
+      await page.waitForSelector('[data-action="settings.section"][data-value="tools"]');
+      await page.evaluate(() => document.querySelector('[data-action="settings.section"][data-value="tools"]')?.click());
      const sw = page.getByRole("switch", { name: label });
      await sw.waitFor();
      check(`${tag}: it starts off`, (await sw.getAttribute("aria-checked")) === "false");
      await sw.scrollIntoViewIfNeeded();
-      const said = await page.evaluate(() => document.body.textContent);
+      const said = await page.evaluate(() => document.body.textContent + [...document.querySelectorAll("#set-external-origin [title]")].map((el) => el.title).join(" "));
      check(`${tag}: it says what it does not cover and that a headless run refuses`, pref === "zh" ? said.includes("改写过或编码过的文字不在其内") && said.includes("会被拒绝而不是执行") : said.includes("reworded or encoded") && said.includes("refused instead of run"));
      if (shotDir) { await page.mouse.move(width - 3, 995); await page.waitForTimeout(250); await page.locator("#set-external-origin").first().screenshot({ path: `${shotDir}/${scheme}-${pref}-${width}-off.png` }).catch(() => page.screenshot({ path: `${shotDir}/${scheme}-${pref}-${width}-off.png` })); }
      await sw.click();
@@ -39,6 +39,7 @@ const browser = await chromium.launch();
 for (const sc of SCENES) {
  const page = await browser.newPage({ viewport: { width: sc.w ?? 1400, height: 900 }, locale: "zh-CN", reducedMotion: "reduce" });
  page.on("pageerror", (e) => fails.push("页面异常: " + e.message));
+  await page.addInitScript(() => localStorage.setItem("rx-nav-rail", "on"));
  await page.goto(BASE + sc.q, { waitUntil: "networkidle" });
  await page.waitForSelector(".app");
  await page.evaluate(() => document.fonts.ready);
@@ -19,6 +19,7 @@ try {
      const open = async (query) => {
        const page = await browser.newPage({ colorScheme: scheme, viewport: { width, height: 1000 }, locale: "zh-CN", reducedMotion: "reduce" });
        page.on("pageerror", (e) => fails.push("页面异常: " + e.message));
+        await page.addInitScript(() => localStorage.setItem("rx-nav-rail", "on"));
        await page.goto(`${BASE}?pref=zh&ws=2&sess=3&turns=2${query}`, { waitUntil: "networkidle" });
        await page.waitForSelector(".app");
        await page.locator('.navbtn[aria-label="文件"]').first().click();
@@ -59,7 +60,7 @@ try {
      await one.waitForSelector('.push-card[data-phase="done"]');
      check(`${tag}: a confirmed push reports it`, (await one.locator(".push-card").textContent()).includes("已把 2 个提交推送到 github.com"));
      await shot(one, "pushed");
-      await one.getByText("提交历史").click();
+      if (!(await one.locator(".workbench-history").evaluate((d) => d.open))) await one.getByText("提交历史", { exact: true }).click();
      await one.getByRole("button", { name: /show the commit graph/ }).click();
      await one.getByRole("button", { name: /retry\.go/ }).last().click();
      await one.waitForSelector(".gh-diff");
@@ -1,54 +1,11 @@
+import "../src/styles/legacy.css";
+import "../src/design/design.css";
 import { createRoot } from "react-dom/client";
 import { boot as bootLang } from "../src/i18n";
 import { track as trackWidth } from "../src/ui/viewport";
 import { pushToast } from "../src/state/toasts";
 import { apply as applyPack } from "../src/ui/theme";
 import { mount as mountEditor } from "./syntax-mount";
-import "../src/styles/tokens.css";
-import "../src/styles/app.css";
-import "../src/styles/dock.css";
-import "../src/styles/studio.css";
-import "../src/styles/changedpill.css";
-import "../src/styles/askskip.css";
-import "../src/styles/pastekeep.css";
-import "../src/styles/ctxitems.css";
-import "../src/styles/sidepanel.css";
-import "../src/styles/contentmatches.css";
-import "../src/styles/scrolljump.css";
-import "../src/styles/chart.css";
-import "../src/styles/community.css";
-import "../src/styles/providerpresets.css";
-import "../src/styles/sessionnotes.css";
-import "../src/styles/sessionquick.css";
-import "../src/styles/sidebardrop.css";
-import "../src/styles/sponsor.css";
-import "../src/styles/versions.css";
-import "../src/styles/phonecomposer.css";
-import "../src/styles/stophint.css";
-import "../src/styles/dialogcenter.css";
-import "../src/styles/treerows.css";
-import "../src/styles/workbench-changes.css";
-import "../src/styles/agentrun.css";
-import "../src/styles/gitgraph.css";
-import "../src/styles/tips.css";
-import "../src/styles/waitnote.css";
-import "../src/styles/bring.css";
-import "../src/styles/peerthreads.css";
-import "../src/styles/line.css";
-import "../src/styles/notice.css";
-import "../src/styles/peers.css";
-import "../src/styles/jsontext.css";
-import "../src/styles/skillpin.css";
-import "../src/styles/narrow.css";
-import "../src/styles/workfold.css";
-import "../src/styles/workbench-launcher.css";
-import "../src/styles/feedback.css";
-import "../src/styles/toasts.css";
-import "../src/styles/modelmenu.css";
-import "../src/styles/composer-menu.css";
-import "../src/styles/runsignals.css";
-import "../src/styles/overview.css";
-import "../src/styles/themes.css";
 import { App } from "../src/ui/App";
 import { MockHub } from "../src/port/mock_hub";
 import { MockPort } from "../src/port/mock";
@@ -56,7 +13,7 @@ import { fromHistory } from "../src/state/session";
 import { HttpError } from "../src/port/port";
 import type { AgentPort, Appearance, ChipCall, HistoryMessage, McpEntry } from "../src/port/port";
 import type { HistoryPage } from "../src/port/session";
-import type { UpdateProgress } from "../src/port/version";
+import type { UpdateProgress, VersionHub } from "../src/port/version";
 import type { ThemePack } from "../src/port/look";
 import type { RuntimeView, TreeWindow, TreeWorkspace } from "../src/port/hub";
 import type { WireEvent } from "../src/port/wire";
@@ -218,8 +175,17 @@ class BenchPort extends MockPort {
    return query.has("update") ? "notify" : super.updateMode();
  }

-  async versions() {
+  async versions(): Promise<VersionHub> {
    if (query.has("update")) return { current: "2.42.0", pinned: "", stalePin: false, latest: "2.43.0", newer: true, versions: [] };
+    if (query.has("versionlayout")) {
+      return {
+        current: "2.56.0", pinned: "", stalePin: false, latest: "2.57.0", newer: true,
+        versions: ["2.57.0", "2.56.1", "2.56.0", "2.55.0", "2.54.0", "2.53.1"].map((version, i) => ({
+          version, tag: `studio-v${version}`, publishedAt: new Date(Date.now() - i * 86400000).toISOString(),
+          hasNotes: i !== 4, current: i === 2, older: i > 2,
+        })),
+      };
+    }
    if (!query.has("versions")) return super.versions();
    const names = ["2.38.0", "2.37.0", "2.36.0", "2.35.0", "2.34.0", "2.33.0", "2.32.0", "2.31.0", "2.30.0", "2.29.0"];
    return {
@@ -307,12 +273,6 @@ class BenchPort extends MockPort {
  }

  private settingCardSent = false;
-  private receiptSent = false;
-
-  async receiptSettings() {
-    const read = await super.receiptSettings();
-    return query.has("receipt") ? { ...read, show: true } : read;
-  }
  private meterSent = false;

  subscribe(on: (e: WireEvent) => void) {
@@ -328,14 +288,6 @@ class BenchPort extends MockPort {
        }
      }, 500);
    }
-    if (query.has("receipt") && !this.receiptSent) {
-      this.receiptSent = true;
-      const script = "cd \"C:/Users/me/work\" && cat > \"$TMPDIR/check/verify_006.py\" <<'PYEOF'\n" + "import os, re, sys, unittest\nclass T(unittest.TestCase):\n    def test_a(self):\n        self.assertEqual(1, 1)\n".repeat(18) + "PYEOF\npython -m unittest verify_006 2>&1 | tail -12";
-      setTimeout(() => {
-        this.feed({ kind: "turn_started", msgIndex: 9 } as unknown as WireEvent);
-        this.feed({ kind: "turn_done", receipt: { saysSomething: true, severity: "warn", gaps: [{ kind: "failed_verification", detail: script }, { kind: "inconclusive_verification", detail: "go test ./internal/... 2>&1 | tail -5" }] } } as unknown as WireEvent);
-      }, 500);
-    }
    const card = query.get("settingcard");
    if (card && !this.settingCardSent) {
      this.settingCardSent = true;
@@ -447,10 +399,22 @@ class BenchPort extends MockPort {
    return super.switchBranch(name);
  }

+  async versionNotes(version: string) {
+    if (!query.has("versionlayout")) return super.versionNotes(version);
+    return { version, markdown: "# Release notes\n\nDownload and rollback improvements.", cached: false };
+  }
+
  async goToVersion() {}

  onUpdateProgress(cb?: (p: UpdateProgress) => void) {
    if (cb && query.has("update")) queueMicrotask(() => cb({ version: "2.43.0", phase: "ready", received: 0, total: 0 }));
+    const phase = query.get("versionlayout");
+    if (cb && (phase === "downloading" || phase === "verifying" || phase === "ready")) {
+      queueMicrotask(() => cb({
+        version: query.has("rollback") ? "2.55.0" : "2.57.0", phase,
+        received: 178_900_000, total: 183_900_000,
+      }));
+    }
    return () => {};
  }

@@ -524,10 +488,13 @@ class BenchHub extends MockHub {
  // ?search=1 answers every query with two conversations that say it.
  async searchHistory(q: string) {
    if (!query.has("search")) return super.searchHistory(q);
-    return [
-      { path: "/sessions/mock.jsonl", id: "mock", snippet: `…上周讨论过要不要把 ${q} 的轮换放进发布流程，结论是每月一次…`, role: "assistant", messageIndex: 4, score: 2 },
-      { path: "/sessions/older.jsonl", id: "older", snippet: `怎么把 ${q} 配到预发环境里？`, role: "user", messageIndex: 0, score: 1 },
-    ];
+    return {
+      hits: [
+        { path: "/sessions/mock.jsonl", id: "mock", snippet: `…上周讨论过要不要把 ${q} 的轮换放进发布流程，结论是每月一次…`, role: "assistant", messageIndex: 4, score: 2 },
+        { path: "/sessions/older.jsonl", id: "older", snippet: `怎么把 ${q} 配到预发环境里？`, role: "user", messageIndex: 0, score: 1 },
+      ],
+      index: { state: "ready", indexed: 2, total: 2, failed: 0, partial: false },
+    };
  }

  private readonly settingsPort = new BenchPort();
@@ -672,7 +639,7 @@ function settingCardFrame(kind: string): WireEvent {
    return { kind: "notice", level: "info", code: "setting_open", text: "助手需要改动 permissions.mode，只有你本人能改。", detail: "permissions.mode" } as WireEvent;
  }
  const elevated = kind === "elevated";
-  const id = elevated ? "peers.dispatch" : "ui.show_receipt";
+  const id = elevated ? "peers.dispatch" : "ui.show_changed_files";
  return {
    kind: "approval_request",
    approval: {
@@ -38,6 +38,8 @@ try {
    check(`${context}: one hover ground across the row`, grounds.size === 1, [...grounds].join(" / "));
    await page.setViewportSize({ width: 390, height: 844 });
    await page.goto(url.href, { waitUntil: "networkidle" });
+    const more = page.locator(".compose .turn-more");
+    if (await more.isVisible()) await more.click();
    const chip = page.locator(turns ? ".compose .studio-branch" : ".compose .worktree-trigger").first();
    await chip.click();
    check(`${context}: a narrow-pane pointer click opens the workspace menu`, await chip.getAttribute("aria-expanded") === "true");
@@ -6,9 +6,11 @@ import { chromium } from "playwright";
 const BASE = process.env.PERF_URL ?? "http://localhost:4399/perf.html";
 // Names the demo data itself carries, and names that are written in Chinese on purpose.
 const FIXTURE = [
-  /并行会话演示/, /上一次的会话/, /站点改版/, /数据管线/, /跑一遍评测/, /^中文$/, /^做游戏的小鱼$/, /^DeepSeek-Reasonix官方群$/,
+  /并行会话演示/, /上一次的会话/, /站点改版/, /数据管线/, /跑一遍评测/, /^中文$/, /^简体中文$/,/^做游戏的小鱼$/, /^DeepSeek-Reasonix官方群$/,
  /缓存命中率为什么会掉/, /Co-Authored-By/, /Generated with/, /^回复用中文$/, /^跟着用户每条消息的语言走$/, /v2 是从零重写的 Go 内核/,
  /没有 web，桌面端重做/, /问题里提到了 main-v2/, /构建用 -tags legacy/, /^迁移前的构建方式$/,
+  /^发布流程$/, /^工作流$/, /^架构$/, /^用户偏好$/, /^发版前先确认合并队列$/, /^队列里有严重修复就先合进同一版$/, /^发版前先扫一遍待合并的修复/,
+  /^每天最多发一个版本$/, /^热修只在完全不可用时例外$/, /^北京日历日最多发一个版本/, /^测试都在 Windows 机器上跑$/, /^Mac 只改代码$/, /^所有 Go\/前端测试和 lint/,
 ];
 const fails = [];
 const browser = await chromium.launch();
@@ -18,12 +18,12 @@ for (const scheme of ["light", "dark"]) {
    await page.goto(`${BASE}?jobs=3&jobnames=1`, { waitUntil: "networkidle" });
    await page.waitForSelector(".app");
    await page.waitForTimeout(800);
-    await page.evaluate(() => [...document.querySelectorAll("button")].find((b) => /后台任务/.test(b.textContent))?.click());
-    await page.waitForSelector(".jobs .job .cmd");
+    await page.locator('[data-action="deck.jobs"]').click();
+    await page.waitForSelector(".job-deck-name");
    const tag = `${scheme} ${width}px`;
-    const shown = await page.evaluate(() => [...document.querySelectorAll(".jobs .job .cmd")].map((e) => e.textContent));
-    check(`${tag}: the list shows the names, not the commands`, shown.length === 3 && shown.every((t) => !/go test|npm run/.test(t)), shown.join(" | "));
-    await page.locator(".jobs .job .cmd").first().hover();
+    const shown = await page.evaluate(() => [...document.querySelectorAll(".job-deck-name")].map((e) => e.textContent));
+    check(`${tag}: the list shows the names, not the commands`, shown.length >= 3 && shown.every((t) => !/go test|npm run/.test(t)), shown.join(" | "));
+    await page.locator(".job-deck-name").first().hover();
    await page.waitForSelector(".job-tip");
    const tip = await page.evaluate(() => {
      const el = document.querySelector(".job-tip");
@@ -35,12 +35,12 @@ for (const scheme of ["light", "dark"]) {
    await page.mouse.move(width - 5, 5);
    await page.waitForTimeout(200);
    check(`${tag}: it closes when the pointer leaves`, (await page.locator(".job-tip").count()) === 0);
-    await page.locator(".jobs .job .cmd").nth(1).focus();
+    await page.locator(".job-deck-name").nth(1).focus();
    await page.waitForSelector(".job-tip");
    check(`${tag}: focus from the keyboard opens it too`, /内核测试/.test(await page.locator(".job-tip .jt-name").textContent()));
    await page.keyboard.press("Escape");
-    await page.waitForTimeout(150);
-    check(`${tag}: Escape closes it`, (await page.locator(".job-tip").count()) === 0);
+    const closed = await page.locator(".job-tip").waitFor({ state: "detached", timeout: 2000 }).then(() => true, () => false);
+    check(`${tag}: Escape closes it`, closed);
    await ctx.close();
  }
 }
@@ -1,5 +1,5 @@
-// An allowed guardian review at low risk and a remembered fact are Lines, not cards: one row over a hairline with no box.
-// A refusal, or an allow that carried real risk, keeps the card.
+// An allowed guardian review at low risk is a Line, not a card: one row over a hairline with no box.
+// A refusal, or an allow that carried real risk, keeps its note, and a remembered fact is a note that carries its take-back.
 import { chromium } from "playwright";

 const BASE = process.env.PERF_URL ?? "http://localhost:4399/perf.html";
@@ -36,15 +36,15 @@ for (const scheme of ["light", "dark"]) {
      lines: rows.length,
      boxes: rows.map(box),
      radius: rows.map((r) => getComputedStyle(r).borderTopLeftRadius),
-      cards: [...document.querySelectorAll('.flow .call[data-k="host"]')].length,
-      forget: !!document.querySelector('.flow .line-row [data-action="memory.forget"]'),
+      cards: [...document.querySelectorAll(".flow .note")].filter((n) => n.textContent.includes("守卫复核")).length,
+      forget: !!document.querySelector('.flow .note [data-action="memory.forget"]'),
      oldMem: document.querySelectorAll('.flow .call[data-k="mem"]').length,
    };
  });
-  check(`${scheme}: the remembered fact and the low-risk allow are rows`, m.lines === 2, `${m.lines} rows`);
+  check(`${scheme}: the low-risk allow is a row`, m.lines === 1, `${m.lines} rows`);
  check(`${scheme}: rows carry no box`, m.lines > 0 && m.boxes.every((b) => b === "0px,0px,0px") && m.radius.every((r) => r === "0px"), JSON.stringify(m));
-  check(`${scheme}: the high-risk allow and the refusal keep their cards`, m.cards === 2, `${m.cards} cards`);
-  check(`${scheme}: the take-back sits in the row, and no memory card remains`, m.forget && m.oldMem === 0);
+  check(`${scheme}: the high-risk allow and the refusal keep their notes`, m.cards === 2, `${m.cards} cards`);
+  check(`${scheme}: the remembered fact keeps its take-back, and no memory card remains`, m.forget && m.oldMem === 0);
  await ctx.close();
 }
 await browser.close();
@@ -94,7 +94,8 @@ for (const scheme of ["light", "dark"]) {
    }

    if (width === 1280) {
-      await page.locator('[data-action="extensions.add"]').first().click();
+      await page.locator('.ext-add-btn[data-action="extensions.add"]').first().click();
+      await page.locator('[role="menuitem"][data-value="source"]').click();
      const source = page.locator("textarea, input").filter({ hasNot: page.locator("[aria-label]") }).last();
      await page.locator('textarea[placeholder*="github.com"], input[placeholder*="github.com"]').first().fill("https://github.com/acme/claude-code mod git-guard");
      await page.locator('[data-action="extensions.inspect"]').first().click();
@@ -0,0 +1,28 @@
+import assert from "node:assert/strict";
+import { chromium } from "playwright";
+
+const PAGE = process.env.PERF_URL ?? "http://localhost:4399/perf.html?pref=zh";
+const browser = await chromium.launch();
+try {
+  const page = await browser.newPage({ viewport: { width: 1300, height: 900 } });
+  const errors = [];
+  page.on("pageerror", (error) => errors.push(error.message));
+  await page.goto(PAGE, { waitUntil: "networkidle" });
+  await page.waitForSelector(".model-picker");
+  await page.locator(".model-picker").first().click();
+  await page.waitForSelector(".studio-model-menu .studio-item-line");
+  const sizes = await page.evaluate(() => {
+    const lines = [...document.querySelectorAll(".studio-model-menu .studio-item-line")].filter((line) => line.querySelector(".studio-item-badge"));
+    for (const line of lines) line.querySelector(".lb").textContent = "workbuddy/deepseek-v4.1-flash-preview-extended";
+    return lines.map((line) => {
+      const badge = line.querySelector(".studio-item-badge").getBoundingClientRect();
+      return { width: Math.round(badge.width), height: Math.round(badge.height) };
+    });
+  });
+  assert.ok(sizes.length > 0, "the menu has a model that reads images");
+  for (const size of sizes) assert.ok(size.height <= 20, `a long model name leaves the badge on one line (${JSON.stringify(size)})`);
+  assert.deepEqual(errors, []);
+  console.log(`PASS model menu badge stays on one line beside a long name (${sizes.length} rows)`);
+} finally {
+  await browser.close();
+}
@@ -18,7 +18,8 @@ try {
    await page.goto(BASE, { waitUntil: "networkidle" });
    await page.waitForSelector(".app", { timeout: 20000 });
    await page.getByRole("button", { name: "设置", exact: true }).click();
-    await page.getByText("模型服务", { exact: true }).first().click();
+    await page.locator('[role="tab"][title="模型"]').click();
+    await page.locator('[data-action="settings.open-page"][data-target="providers"]').click();
    await page.getByText("myrelay", { exact: true }).first().click();
    await page.getByText("从服务商读取可用模型").first().click();
    await page.waitForSelector(".mrows .mline");
@@ -23,7 +23,8 @@ for (const scheme of ["light", "dark"]) {
    await page.waitForTimeout(500);
    await page.evaluate(() => [...document.querySelectorAll("button")].find((x) => /备份原文件并修复/.test(x.textContent))?.click());
    await page.waitForTimeout(400);
-    await page.evaluate(() => [...document.querySelectorAll('[data-action="settings.section"]')].find((x) => x.textContent.includes("模型服务"))?.click());
+    await page.evaluate(() => document.querySelector('[role="tab"][title="模型"]')?.click());
+    await page.evaluate(() => document.querySelector('[data-action="settings.open-page"][data-target="providers"]')?.click());
    await page.waitForSelector(".pdetail", { state: "attached" });
    await page.evaluate(() => [...document.querySelectorAll("button,li,a")].find((x) => /myrelay-work/.test(x.textContent) && x.textContent.length < 80)?.click());
    await page.waitForSelector(".pdetail .mlhead", { state: "attached" });
@@ -139,6 +139,11 @@ for (const scheme of ["light", "dark"]) {
    await page.waitForSelector(".app");
    await page.evaluate(() => document.fonts.ready);
    await page.waitForTimeout(500);
+    const more = page.locator(".compose .turn-more");
+    if (await more.isVisible()) {
+      await more.click();
+      await page.waitForTimeout(300);
+    }
    const tag = `${scheme} ${width}px zoom ${zoom}`;
    const m = await probe(page);
    check(`${tag}: title bar controls do not overlap`, m.bar.overlaps.length === 0, m.bar.overlaps.join(", "));