docs: draft code signing policy page (not in force)

2026-11-07 10:45:59 -10:00 · 2026-09-30 01:46:32 -07:00
6938 changed files with 35638 additions and 564216 deletions
@@ -5,7 +5,6 @@
 *.bat text eol=crlf
 *.css text eol=lf
 *.go text eol=lf
-*.golden text eol=lf
 *.html text eol=lf
 *.js text eol=lf
 *.json text eol=lf
@@ -26,8 +25,6 @@
 *.xml text eol=lf
 *.yaml text eol=lf
 *.yml text eol=lf
-Makefile text eol=lf
 .githooks/** text eol=lf
 benchmarks/**/*.sh text eol=lf
 internal/runtime/agent/testdata/perseveration/** -text
-desktop/frontend-next/perf/media/delivery.pdf binary
@@ -10,7 +10,6 @@

 # User-facing guides and references.
 /docs/                                    @SivanCola @esengine
-/docs/THEME_AUTHOR_GUIDE.md                @esengine @SivanCola
 /sdk/                                     @SivanCola @esengine
 /README.md                                @SivanCola @esengine
 /README.zh-CN.md                          @SivanCola @esengine
@@ -23,6 +22,7 @@
 /docs/*_CONTRACT.md                       @esengine @SivanCola
 /docs/TOOL_APPROVAL_MODES.md              @esengine @SivanCola
 /docs/CAPABILITY_DIAGNOSTICS.md           @esengine @SivanCola
+/docs/CODE_SIGNING_POLICY.md            @esengine @SivanCola
 /docs/CHECKPOINTS.md                      @esengine @SivanCola
 /docs/EXTENSION_*.md                      @esengine @SivanCola
 /docs/SUBAGENT_*.md                       @esengine @SivanCola
@@ -37,7 +37,6 @@
 /.github/workflows/studio-certum-signing-smoke.yml @esengine @SivanCola
 /.github/actions/setup-certum/            @esengine @SivanCola
 /cmd/signpath-contract/                   @esengine @SivanCola
-/scripts/install-smoke.ps1                @esengine @SivanCola
 /scripts/sign-certum.ps1                  @esengine @SivanCola
 /scripts/verify-windows-authenticode.ps1  @esengine @SivanCola
 /scripts/verify-windows-package.ps1       @esengine @SivanCola
@@ -6,9 +6,3 @@ contact_links:
  - name: Questions & discussions
    url: https://github.com/esengine/DeepSeek-Reasonix/discussions
    about: For usage questions, ideas, and general discussion — please use Discussions instead of opening an issue.
-  - name: Discord
-    url: https://discord.gg/XF78rEME2D
-    about: Bilingual community chat for setup help, workflow showcases, and feature ideas.
-  - name: QQ group (DeepSeek-Reasonix官方群, 1093562660)
-    url: https://qm.qq.com/q/i59b0z2R8s
-    about: Chinese-language community chat for setup help and everyday questions.
@@ -8,34 +8,6 @@ only, on `main-v2`.

 -

-## Cause
-
-
-
-## Blast radius
-
-
-
-## Neighbouring behaviours tested
-
-
-
-## Why this layer
-
-
-
-## UI changes
-
-<!--
-Delete this section if the PR touches nothing under `desktop/frontend-next/src`
-or the website pages.
-->
-
- Screenshots, before and after: light and dark, a wide window and a phone-width one (about 390 px). Attach them to this description.
- Entries moved, hidden or removed: none / list each one.
- Hiding or removing a visible control needs a linked issue and a release-notes line naming it. Do not retire UI with `display: none !important` or a feature flag.
- If a row of `docs/STUDIO_PARITY.md` changes, update it here.
-
 ## Issues

 <!--
@@ -54,7 +26,3 @@ the source-to-target contract here: easing syntax, time units, callbacks,
 cancellation, reduced-motion behavior, and failure fallback. Verification must
 assert that the target API was actually called; a mock that silently skips it
 does not count.
-
-## Disclosure
-
- [ ] AI assistance was used for this change (optional; for transparency only, it does not change how the change is reviewed)
@@ -1,19 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg id="_图层_2" data-name="图层 2" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 145.55 113.29">
-  <defs>
-    <style>
-      .cls-1 {
-        fill: #ffffff;
-        stroke: #ffffff;
-        stroke-miterlimit: 10;
-      }
-    </style>
-  </defs>
-  <g id="_图层_1-2" data-name="图层 1">
-    <g>
-      <path class="cls-1" d="M144.68,38.49l-.06-.23c-.88-3.28-2.5-5.94-4.58-8.06.14,5.65-2.96,11.02-6.22,16.66l-.39.68c-2.26,3.91-4.66,7.94-6.57,11.1l-.86,1.38c-3.36,5.44-6.27,10.14-12.18,14.18-8.34,5.87-18.2,5.81-26.92,5.76-2.81-.02-5.48-.03-7.95.14l-.35.02c-3.22.06-5.96,1.57-8.17,4.49l-.14.18c-.86,1.06-1.7,2.26-2.58,3.54-3.43,4.92-7.69,11.04-16.17,12.42-4.37.86-9.98.84-14.94.83h-1.95c-.52,0-1.06,0-1.61,0-6.95,0-16.08-.89-21.94-6.55-.15-.15-.3-.3-.44-.45.5,4.73,2.33,8.64,5.44,11.65,5.86,5.66,14.98,6.55,21.94,6.55.55,0,1.09,0,1.61-.01h1.93c4.96.02,10.58.04,14.96-.83,8.48-1.37,12.74-7.5,16.17-12.42.88-1.27,1.72-2.48,2.58-3.54l.14-.18c2.21-2.92,4.95-4.43,8.17-4.49l.35-.02c2.47-.17,5.13-.16,7.95-.14,8.72.05,18.58.11,26.91-5.76,5.92-4.03,8.82-8.73,12.19-14.17l.86-1.39c1.91-3.15,4.3-7.18,6.57-11.09l.39-.69c3.81-6.6,7.41-12.84,5.86-19.57ZM120.9,23.02c-.28,0-.56,0-.83,0-9.68-.19-24.03-.09-35.57-.01l-2.04.02c-.93.01-1.82.02-2.67.03-8.36.08-14.4.13-23.33,3.82l-.27.12c-10.76,4.68-16.91,12.16-22.83,21.95-5.53,8.76-12.62,20.57-16.32,26.79-.41.69-.77,1.3-1.09,1.84l-.49.84c-1.49,2.53-3.23,5.49-4.21,8.95,3.98,1.9,8.52,2.65,12.71,2.91.27-.77.64-1.56,1.07-2.38.48-.94,1.05-1.9,1.63-2.89l.49-.84c.96-1.62,2.38-3.99,4.05-6.78,3.82-6.36,8.98-14.88,13.19-21.55l.07-.11c5.02-8.31,9.2-13.45,16.91-16.81l.11-.05c6.57-2.7,10.54-2.74,18.43-2.81.87-.01,1.78-.02,2.69-.03l1.99-.02c9.17-.06,20.13-.14,29.01-.07,2.26.01,4.39.04,6.32.08h.12s.11,0,.11,0c1.25-.01,2.91.11,4.61.46,1.03.22,2.08.52,3.05.93.21-.35.41-.71.62-1.06l.39-.68c1.92-3.33,3.79-6.57,4.97-9.82-4.08-1.92-8.7-2.76-12.89-2.82Z"/>
-      <path class="cls-1" d="M115.07,11.81c-9.7-.19-24.07-.09-35.62-.01h-1.99c-.93.03-1.82.04-2.67.04-8.36.08-14.4.14-23.33,3.83l-.27.12c-10.76,4.67-16.91,12.16-22.83,21.95-6.13,9.71-14.19,23.21-17.41,28.63l-.5.84c-2.08,3.55-4.68,7.97-4.94,13.39v.2s0,.21,0,.21c.01.81.06,1.61.15,2.38.14.15.29.3.44.45,1.53,1.47,3.28,2.63,5.15,3.52,3.98,1.9,8.52,2.65,12.71,2.91,1.41.09,2.78.12,4.08.12.55,0,1.09-.01,1.61-.01h1.95c4.95.01,10.57.03,14.94-.83,8.48-1.38,12.74-7.5,16.17-12.42.88-1.28,1.72-2.48,2.58-3.54l.14-.18c2.21-2.92,4.95-4.44,8.17-4.49l.35-.02c2.47-.17,5.14-.16,7.95-.14,8.71.05,18.58.1,26.92-5.76,5.91-4.04,8.82-8.74,12.18-14.18l.86-1.39c1.74-2.86,3.87-6.45,5.95-10.03.21-.35.41-.71.62-1.06l.39-.68c1.92-3.33,3.79-6.57,4.97-9.82.82-2.26,1.31-4.53,1.25-6.84-5.18-5.29-13.22-7.25-19.97-7.19ZM122.56,40.37l-.39.67c-2.21,3.81-4.55,7.77-6.39,10.79l-.84,1.36c-3.01,4.87-4.83,7.81-8.48,10.29l-.1.07c-4.94,3.49-11.95,3.45-19.38,3.41-2.88-.02-5.87-.04-8.79.16-7.1.19-13.51,3.58-18.07,9.57-1.13,1.4-2.12,2.83-3.08,4.21-2.92,4.18-4.71,6.57-7.64,7.02l-.31.06c-3.09.63-8.28.61-12.45.6h-2.16c-3.85.07-7.01-.16-9.45-.69-2.24-.48-3.87-1.22-4.89-2.2-.66-.64-1.54-1.81-1.63-4.65.11-1.35.71-2.82,1.52-4.35.48-.94,1.05-1.9,1.63-2.89l.49-.84c3.17-5.33,11.19-18.75,17.24-28.33l.07-.11c5.02-8.32,9.2-13.45,16.92-16.81l.1-.05c6.57-2.7,10.54-2.74,18.43-2.82.87,0,1.78,0,2.69-.03h1.94c11.52-.09,25.86-.19,35.38,0h.23c1.25-.01,2.92.11,4.61.46,3.15.66,6.4,2.12,7.27,5.02.2,1.2-.89,3.62-2.27,6.18-.7,1.31-1.48,2.65-2.2,3.9ZM120.07,23.01c-9.68-.19-24.03-.09-35.57-.01l-2.04.02c-.93.01-1.82.02-2.67.03-8.36.08-14.4.13-23.33,3.82l-.27.12c-10.76,4.68-16.91,12.16-22.83,21.95-5.53,8.76-12.62,20.57-16.32,26.79-.78-.35-1.41-.77-1.9-1.24-.66-.64-1.54-1.81-1.63-4.65.18-2.18,1.62-4.64,3.15-7.25l.49-.83c3.17-5.32,11.18-18.73,17.24-28.33l.07-.12c5.02-8.31,9.2-13.45,16.92-16.81l.1-.04c6.57-2.7,10.54-2.74,18.43-2.82.87-.01,1.78-.01,2.69-.03h1.99c11.5-.09,25.82-.19,35.33,0h.23c3.57-.04,10.55,1.02,11.88,5.48.14.84-.35,2.27-1.13,3.93-.28,0-.56,0-.83,0Z"/>
-      <path class="cls-1" d="M134.68,16.09l-.06-.23c-3.07-11.45-15.08-15.33-24.55-15.25-9.68-.19-24.03-.09-35.57-.01h-2.04c-.93.03-1.82.04-2.67.04-8.36.08-14.4.14-23.33,3.83l-.27.12c-10.76,4.67-16.91,12.16-22.83,21.95-6.14,9.73-14.2,23.22-17.41,28.62l-.49.85c-2.09,3.55-4.69,7.96-4.95,13.39v.2s0,.21,0,.21c.09,5.57,1.82,10.13,5.15,13.58.14.15.29.3.44.45,1.53,1.47,3.28,2.63,5.15,3.52,3.98,1.9,8.52,2.65,12.71,2.91,1.41.09,2.78.12,4.08.12.55,0,1.09-.01,1.61-.01h1.95c4.95.01,10.57.03,14.94-.83,8.48-1.38,12.74-7.5,16.17-12.42.88-1.28,1.72-2.48,2.58-3.54l.14-.18c2.21-2.92,4.95-4.44,8.17-4.49l.35-.02c2.47-.17,5.14-.16,7.95-.14,8.71.05,18.58.1,26.92-5.76,5.91-4.04,8.82-8.74,12.18-14.18l.86-1.39c1.74-2.86,3.87-6.45,5.95-10.03.21-.35.41-.71.62-1.06l.39-.68c1.92-3.33,3.79-6.57,4.97-9.82.82-2.26,1.31-4.53,1.25-6.84-.02-.96-.14-1.93-.36-2.91ZM119.76,25.27c-.7,1.31-1.48,2.65-2.2,3.9l-.39.67c-1.19,2.04-2.41,4.13-3.57,6.09-1.01,1.7-1.97,3.31-2.82,4.7l-.84,1.36c-3.01,4.87-4.83,7.81-8.48,10.29l-.1.07c-4.94,3.49-11.96,3.45-19.38,3.41-2.89-.02-5.87-.04-8.79.16-7.1.18-13.51,3.58-18.07,9.57-1.13,1.4-2.12,2.83-3.08,4.21-2.92,4.18-4.71,6.57-7.64,7.02l-.31.06c-3.09.63-8.27.61-12.45.6h-2.16c-3.85.07-7.01-.16-9.45-.69-1.16-.25-2.16-.57-2.99-.96-.78-.35-1.41-.77-1.9-1.24-.66-.64-1.54-1.81-1.63-4.65.18-2.18,1.62-4.64,3.15-7.25l.49-.83c3.17-5.32,11.18-18.73,17.24-28.33l.07-.12c5.02-8.31,9.2-13.45,16.92-16.81l.1-.04c6.57-2.7,10.54-2.74,18.43-2.82.87-.01,1.78-.01,2.69-.03h1.99c11.5-.09,25.82-.19,35.33,0h.23c3.57-.04,10.55,1.02,11.88,5.48.14.84-.35,2.27-1.13,3.93-.34.72-.73,1.48-1.14,2.25Z"/>
-    </g>
-  </g>
-</svg>
@@ -1,19 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg id="_图层_2" data-name="图层 2" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 145.55 113.29">
-  <defs>
-    <style>
-      .cls-1 {
-        fill: #111111;
-        stroke: #111111;
-        stroke-miterlimit: 10;
-      }
-    </style>
-  </defs>
-  <g id="_图层_1-2" data-name="图层 1">
-    <g>
-      <path class="cls-1" d="M144.68,38.49l-.06-.23c-.88-3.28-2.5-5.94-4.58-8.06.14,5.65-2.96,11.02-6.22,16.66l-.39.68c-2.26,3.91-4.66,7.94-6.57,11.1l-.86,1.38c-3.36,5.44-6.27,10.14-12.18,14.18-8.34,5.87-18.2,5.81-26.92,5.76-2.81-.02-5.48-.03-7.95.14l-.35.02c-3.22.06-5.96,1.57-8.17,4.49l-.14.18c-.86,1.06-1.7,2.26-2.58,3.54-3.43,4.92-7.69,11.04-16.17,12.42-4.37.86-9.98.84-14.94.83h-1.95c-.52,0-1.06,0-1.61,0-6.95,0-16.08-.89-21.94-6.55-.15-.15-.3-.3-.44-.45.5,4.73,2.33,8.64,5.44,11.65,5.86,5.66,14.98,6.55,21.94,6.55.55,0,1.09,0,1.61-.01h1.93c4.96.02,10.58.04,14.96-.83,8.48-1.37,12.74-7.5,16.17-12.42.88-1.27,1.72-2.48,2.58-3.54l.14-.18c2.21-2.92,4.95-4.43,8.17-4.49l.35-.02c2.47-.17,5.13-.16,7.95-.14,8.72.05,18.58.11,26.91-5.76,5.92-4.03,8.82-8.73,12.19-14.17l.86-1.39c1.91-3.15,4.3-7.18,6.57-11.09l.39-.69c3.81-6.6,7.41-12.84,5.86-19.57ZM120.9,23.02c-.28,0-.56,0-.83,0-9.68-.19-24.03-.09-35.57-.01l-2.04.02c-.93.01-1.82.02-2.67.03-8.36.08-14.4.13-23.33,3.82l-.27.12c-10.76,4.68-16.91,12.16-22.83,21.95-5.53,8.76-12.62,20.57-16.32,26.79-.41.69-.77,1.3-1.09,1.84l-.49.84c-1.49,2.53-3.23,5.49-4.21,8.95,3.98,1.9,8.52,2.65,12.71,2.91.27-.77.64-1.56,1.07-2.38.48-.94,1.05-1.9,1.63-2.89l.49-.84c.96-1.62,2.38-3.99,4.05-6.78,3.82-6.36,8.98-14.88,13.19-21.55l.07-.11c5.02-8.31,9.2-13.45,16.91-16.81l.11-.05c6.57-2.7,10.54-2.74,18.43-2.81.87-.01,1.78-.02,2.69-.03l1.99-.02c9.17-.06,20.13-.14,29.01-.07,2.26.01,4.39.04,6.32.08h.12s.11,0,.11,0c1.25-.01,2.91.11,4.61.46,1.03.22,2.08.52,3.05.93.21-.35.41-.71.62-1.06l.39-.68c1.92-3.33,3.79-6.57,4.97-9.82-4.08-1.92-8.7-2.76-12.89-2.82Z"/>
-      <path class="cls-1" d="M115.07,11.81c-9.7-.19-24.07-.09-35.62-.01h-1.99c-.93.03-1.82.04-2.67.04-8.36.08-14.4.14-23.33,3.83l-.27.12c-10.76,4.67-16.91,12.16-22.83,21.95-6.13,9.71-14.19,23.21-17.41,28.63l-.5.84c-2.08,3.55-4.68,7.97-4.94,13.39v.2s0,.21,0,.21c.01.81.06,1.61.15,2.38.14.15.29.3.44.45,1.53,1.47,3.28,2.63,5.15,3.52,3.98,1.9,8.52,2.65,12.71,2.91,1.41.09,2.78.12,4.08.12.55,0,1.09-.01,1.61-.01h1.95c4.95.01,10.57.03,14.94-.83,8.48-1.38,12.74-7.5,16.17-12.42.88-1.28,1.72-2.48,2.58-3.54l.14-.18c2.21-2.92,4.95-4.44,8.17-4.49l.35-.02c2.47-.17,5.14-.16,7.95-.14,8.71.05,18.58.1,26.92-5.76,5.91-4.04,8.82-8.74,12.18-14.18l.86-1.39c1.74-2.86,3.87-6.45,5.95-10.03.21-.35.41-.71.62-1.06l.39-.68c1.92-3.33,3.79-6.57,4.97-9.82.82-2.26,1.31-4.53,1.25-6.84-5.18-5.29-13.22-7.25-19.97-7.19ZM122.56,40.37l-.39.67c-2.21,3.81-4.55,7.77-6.39,10.79l-.84,1.36c-3.01,4.87-4.83,7.81-8.48,10.29l-.1.07c-4.94,3.49-11.95,3.45-19.38,3.41-2.88-.02-5.87-.04-8.79.16-7.1.19-13.51,3.58-18.07,9.57-1.13,1.4-2.12,2.83-3.08,4.21-2.92,4.18-4.71,6.57-7.64,7.02l-.31.06c-3.09.63-8.28.61-12.45.6h-2.16c-3.85.07-7.01-.16-9.45-.69-2.24-.48-3.87-1.22-4.89-2.2-.66-.64-1.54-1.81-1.63-4.65.11-1.35.71-2.82,1.52-4.35.48-.94,1.05-1.9,1.63-2.89l.49-.84c3.17-5.33,11.19-18.75,17.24-28.33l.07-.11c5.02-8.32,9.2-13.45,16.92-16.81l.1-.05c6.57-2.7,10.54-2.74,18.43-2.82.87,0,1.78,0,2.69-.03h1.94c11.52-.09,25.86-.19,35.38,0h.23c1.25-.01,2.92.11,4.61.46,3.15.66,6.4,2.12,7.27,5.02.2,1.2-.89,3.62-2.27,6.18-.7,1.31-1.48,2.65-2.2,3.9ZM120.07,23.01c-9.68-.19-24.03-.09-35.57-.01l-2.04.02c-.93.01-1.82.02-2.67.03-8.36.08-14.4.13-23.33,3.82l-.27.12c-10.76,4.68-16.91,12.16-22.83,21.95-5.53,8.76-12.62,20.57-16.32,26.79-.78-.35-1.41-.77-1.9-1.24-.66-.64-1.54-1.81-1.63-4.65.18-2.18,1.62-4.64,3.15-7.25l.49-.83c3.17-5.32,11.18-18.73,17.24-28.33l.07-.12c5.02-8.31,9.2-13.45,16.92-16.81l.1-.04c6.57-2.7,10.54-2.74,18.43-2.82.87-.01,1.78-.01,2.69-.03h1.99c11.5-.09,25.82-.19,35.33,0h.23c3.57-.04,10.55,1.02,11.88,5.48.14.84-.35,2.27-1.13,3.93-.28,0-.56,0-.83,0Z"/>
-      <path class="cls-1" d="M134.68,16.09l-.06-.23c-3.07-11.45-15.08-15.33-24.55-15.25-9.68-.19-24.03-.09-35.57-.01h-2.04c-.93.03-1.82.04-2.67.04-8.36.08-14.4.14-23.33,3.83l-.27.12c-10.76,4.67-16.91,12.16-22.83,21.95-6.14,9.73-14.2,23.22-17.41,28.62l-.49.85c-2.09,3.55-4.69,7.96-4.95,13.39v.2s0,.21,0,.21c.09,5.57,1.82,10.13,5.15,13.58.14.15.29.3.44.45,1.53,1.47,3.28,2.63,5.15,3.52,3.98,1.9,8.52,2.65,12.71,2.91,1.41.09,2.78.12,4.08.12.55,0,1.09-.01,1.61-.01h1.95c4.95.01,10.57.03,14.94-.83,8.48-1.38,12.74-7.5,16.17-12.42.88-1.28,1.72-2.48,2.58-3.54l.14-.18c2.21-2.92,4.95-4.44,8.17-4.49l.35-.02c2.47-.17,5.14-.16,7.95-.14,8.71.05,18.58.1,26.92-5.76,5.91-4.04,8.82-8.74,12.18-14.18l.86-1.39c1.74-2.86,3.87-6.45,5.95-10.03.21-.35.41-.71.62-1.06l.39-.68c1.92-3.33,3.79-6.57,4.97-9.82.82-2.26,1.31-4.53,1.25-6.84-.02-.96-.14-1.93-.36-2.91ZM119.76,25.27c-.7,1.31-1.48,2.65-2.2,3.9l-.39.67c-1.19,2.04-2.41,4.13-3.57,6.09-1.01,1.7-1.97,3.31-2.82,4.7l-.84,1.36c-3.01,4.87-4.83,7.81-8.48,10.29l-.1.07c-4.94,3.49-11.96,3.45-19.38,3.41-2.89-.02-5.87-.04-8.79.16-7.1.18-13.51,3.58-18.07,9.57-1.13,1.4-2.12,2.83-3.08,4.21-2.92,4.18-4.71,6.57-7.64,7.02l-.31.06c-3.09.63-8.27.61-12.45.6h-2.16c-3.85.07-7.01-.16-9.45-.69-1.16-.25-2.16-.57-2.99-.96-.78-.35-1.41-.77-1.9-1.24-.66-.64-1.54-1.81-1.63-4.65.18-2.18,1.62-4.64,3.15-7.25l.49-.83c3.17-5.32,11.18-18.73,17.24-28.33l.07-.12c5.02-8.31,9.2-13.45,16.92-16.81l.1-.04c6.57-2.7,10.54-2.74,18.43-2.82.87-.01,1.78-.01,2.69-.03h1.99c11.5-.09,25.82-.19,35.33,0h.23c3.57-.04,10.55,1.02,11.88,5.48.14.84-.35,2.27-1.13,3.93-.34.72-.73,1.48-1.14,2.25Z"/>
-    </g>
-  </g>
-</svg>
@@ -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"
-    ]
-  }
-]
@@ -1,37 +0,0 @@
-name: acp-conformance
-
-# The agent is driven over stdio by the official ACP TypeScript SDK, and every
-# message it writes is held to the schema that SDK ships. Not a required check:
-# it runs when the ACP frontend or the harness changes, and on request.
-on:
-  pull_request:
-    branches: [main-v2, studio]
-    paths:
-      - "internal/frontend/acp/**"
-      - "tools/acp-conformance/**"
-      - ".github/workflows/acp-conformance.yml"
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-concurrency:
-  group: acp-conformance-${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  conformance:
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          persist-credentials: false
-      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
-        with:
-          go-version-file: go.mod
-          cache: true
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
-        with:
-          node-version: "22"
-      - run: make acp-conformance
@@ -1,293 +0,0 @@
-name: Android
-
-# The Android shell under `android/` ships from the same tag the desktop line
-# publishes, and lands on the same release page.
-#
-# It is a separate workflow rather than a job inside release-studio.yml for two
-# reasons. The toolchains share nothing — Go, the Android SDK, a Gradle wrapper —
-# so a runner carrying both would install three toolchains to build one artifact.
-# And `signing-contract` pins which workflows may reach the SignPath credentials;
-# this one never touches them, and keeping it out of that graph is what keeps the
-# fingerprint stable.
-#
-# A failure here does not hold back the desktop release: the two are independent
-# by construction, and an APK is worth less than the desktop builds it sits
-# beside.
-on:
-  push:
-    tags: ['studio-v*']
-  workflow_dispatch:
-    inputs:
-      tag:
-        description: "Existing studio tag to build for and attach to (for example studio-v2.51.0)"
-        required: false
-        type: string
-      ref:
-        description: "A branch to build as a trial run of this workflow; nothing is attached to any release"
-        required: false
-        type: string
-
-permissions:
-  contents: write # upload the APK to the release the tag belongs to
-
-concurrency:
-  group: android-${{ inputs.tag || inputs.ref || github.ref_name }}
-  cancel-in-progress: false
-
-jobs:
-  resolve:
-    name: resolve the tag
-    runs-on: ubuntu-latest
-    permissions:
-      contents: read
-    outputs:
-      tag: ${{ steps.release.outputs.tag }}
-      version: ${{ steps.release.outputs.version }}
-      sha: ${{ steps.release.outputs.sha }}
-      candidate: ${{ steps.release.outputs.candidate }}
-      publish: ${{ steps.release.outputs.publish }}
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          fetch-depth: 0
-          persist-credentials: false
-
-      # The same tag grammar release-studio.yml enforces, spelled the same way:
-      # two workflows that disagree about which tags name a release would publish
-      # one artifact for a tag the other refused.
-      - name: Resolve tag and version
-        id: release
-        env:
-          IN_TAG: ${{ inputs.tag }}
-          IN_REF: ${{ inputs.ref }}
-          REF_NAME: ${{ github.ref_name }}
-        run: |
-          # A trial run builds a branch as a throwaway version and publishes nothing.
-          if [ -n "$IN_REF" ] && [ -z "$IN_TAG" ]; then
-            sha="$(git rev-parse --verify "origin/$IN_REF^{commit}")"
-            { echo "tag="; echo "version=v0.0.0-dev.${sha:0:7}"; echo "sha=$sha"
-              echo "candidate=false"; echo "publish=false"; } >> "$GITHUB_OUTPUT"
-            echo "trial build of $IN_REF at $sha"
-            exit 0
-          fi
-          tag="${IN_TAG:-$REF_NAME}"
-          semver='^(0|[1-9][0-9]*)\.(0|[1-9][0-9]*)\.(0|[1-9][0-9]*)(-([0-9A-Za-z-]+)(\.[0-9A-Za-z-]+)*)?$'
-          if [[ ! "$tag" =~ ^studio-v(.+)$ ]] || [[ ! "${BASH_REMATCH[1]}" =~ $semver ]]; then
-            echo "::error::studio release tag must be studio-vMAJOR.MINOR.PATCH[-PRERELEASE], got: $tag"
-            exit 1
-          fi
-          sha="$(git rev-parse --verify "refs/tags/$tag^{commit}")"
-          if ! git merge-base --is-ancestor "$sha" origin/studio; then
-            echo "::error::$tag is not on studio history"
-            exit 1
-          fi
-          version="${tag#studio-}"
-          candidate=false
-          if [[ "${version#v}" =~ $semver ]] && [ -n "${BASH_REMATCH[4]}" ]; then
-            candidate=true
-          fi
-          { echo "tag=$tag"; echo "version=$version"; echo "sha=$sha"
-            echo "candidate=$candidate"; echo "publish=true"; } >> "$GITHUB_OUTPUT"
-          echo "resolved $tag -> $version at $sha (candidate=$candidate)"
-
-  build:
-    name: build the APK
-    needs: resolve
-    runs-on: ubuntu-latest
-    permissions:
-      contents: read
-    # `publish` decides whether to attach anything, and the only honest answer to
-    # "will this APK install" is whether a key was restored.
-    outputs:
-      signed: ${{ steps.key.outputs.signed }}
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          ref: ${{ needs.resolve.outputs.sha }}
-          fetch-depth: 0
-          persist-credentials: false
-
-      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
-        with:
-          go-version-file: go.mod
-          cache: true
-          cache-dependency-path: go.sum
-
-      # pnpm first: setup-node's pnpm cache needs the executable to exist.
-      - uses: pnpm/action-setup@ea17c68df8912ef543352723c149a84f56e3d413 # v4
-        with:
-          version: 10
-
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
-        with:
-          node-version: "24"
-          cache: pnpm
-          cache-dependency-path: desktop/frontend-next/pnpm-lock.yaml
-
-      - uses: actions/setup-java@b843ebf9b913146b38da2eaa7aeddf76074ae1ad # v6.1.0
-        with:
-          distribution: temurin
-          java-version: "21"
-          cache: gradle
-
-      # The SDK is the one thing nothing sets for you, per android/README.md. The
-      # runner image carries one; this asserts the two pieces the build names
-      # rather than re-installing them, so a missing package fails here by name
-      # instead of surfacing later as an unresolved compileSdk.
-      - name: Provide the Android SDK
-        run: |
-          if [ -z "${ANDROID_HOME:-}" ]; then
-            echo "::error::ANDROID_HOME is unset; the runner image is expected to provide the Android SDK"
-            exit 1
-          fi
-          for piece in "platforms/android-35" "build-tools/35.0.0"; do
-            if [ ! -d "$ANDROID_HOME/$piece" ]; then
-              echo "::error::the Android SDK at $ANDROID_HOME has no $piece; android/README.md names it as a requirement"
-              exit 1
-            fi
-          done
-          echo "ANDROID_SDK_ROOT=$ANDROID_HOME" >> "$GITHUB_ENV"
-
-      # A key that exists signs; a key that does not leaves the artifact
-      # unsigned, which is what a release with no key has always produced. The
-      # decision is the repository's, not this workflow's: it reads the four
-      # secrets android/README.md already names.
-      - name: Restore the signing key
-        id: key
-        env:
-          KEYSTORE_BASE64: ${{ secrets.REASONIX_KEYSTORE_BASE64 }}
-          STORE_PASSWORD: ${{ secrets.REASONIX_KEYSTORE_PASSWORD }}
-          KEY_PASSWORD: ${{ secrets.REASONIX_KEY_PASSWORD }}
-          KEY_ALIAS: ${{ secrets.REASONIX_KEY_ALIAS }}
-        run: |
-          if [ -z "$KEYSTORE_BASE64" ] || [ -z "$STORE_PASSWORD" ] || [ -z "$KEY_ALIAS" ]; then
-            echo "signed=false" >> "$GITHUB_OUTPUT"
-            echo "no signing key configured: the APK will be unsigned"
-            exit 0
-          fi
-          mkdir -p android/keystore
-          printf '%s' "$KEYSTORE_BASE64" | base64 --decode > android/keystore/release.jks
-          {
-            echo "storeFile=keystore/release.jks"
-            echo "storePassword=${STORE_PASSWORD}"
-            echo "keyAlias=${KEY_ALIAS}"
-            # PKCS#12 keystores use one password for both; only write the second
-            # when it is genuinely a different fact.
-            [ -n "$KEY_PASSWORD" ] && echo "keyPassword=${KEY_PASSWORD}"
-          } > android/keystore/keystore.properties
-          chmod 600 android/keystore/*
-          echo "signed=true" >> "$GITHUB_OUTPUT"
-          echo "signing key restored"
-
-      - name: Build the kernel and the page
-        working-directory: android
-        env:
-          VERSION: ${{ needs.resolve.outputs.version }}
-        run: |
-          export REASONIX_APP_VERSION="${VERSION#v}"
-          bash ./scripts/build-kernel.sh
-          bash ./scripts/sync-page.sh
-
-      # The one end-to-end check that needs no phone: the kernel the shell starts,
-      # serving the page the APK carries. Running it here means an APK that would
-      # open blank fails the build instead of the install.
-      - name: Verify the packaged page
-        working-directory: android
-        run: |
-          (cd .. && go build -o reasonix-verify ./cmd/reasonix)
-          KERNEL_BIN="$PWD/../reasonix-verify" bash ./scripts/verify-page.sh
-
-      - name: Assemble
-        working-directory: android
-        env:
-          VERSION: ${{ needs.resolve.outputs.version }}
-        run: |
-          # The version comes from the tag, so the About screen and the update
-          # check name this release. versionCode keeps its existing shape, which
-          # is why the build digit stays 1: it is a single decimal place, and a
-          # number that grows past it would collide with the next patch.
-          ./gradlew assembleRelease --no-daemon "-PreasonixVersion=${VERSION#v}"
-
-      - name: Name the artifact
-        id: apk
-        working-directory: android/app/build/outputs/apk/release
-        run: |
-          src="$(ls ./*.apk | head -n 1)"
-          if [ -z "$src" ]; then
-            echo "::error::assembleRelease produced no APK"
-            exit 1
-          fi
-          case "$src" in
-            *-unsigned.apk) name="ReasonixMobile-${{ needs.resolve.outputs.version }}-unsigned.apk" ;;
-            *)              name="ReasonixMobile-${{ needs.resolve.outputs.version }}.apk" ;;
-          esac
-          mv "$src" "$name"
-          sha256sum "$name" > "$name.sha256"
-          {
-            echo "path=$PWD/$name"
-            echo "name=$name"
-          } >> "$GITHUB_OUTPUT"
-          ls -l "$name" "$name.sha256"
-
-      # An unsigned package is not a lesser APK, it is an unusable one: Android's
-      # package manager rejects it outright rather than reporting a signature
-      # mismatch. Say so where the run summary can be read, instead of letting a
-      # green build imply a shippable artifact.
-      - name: Say what an unsigned build means
-        if: steps.key.outputs.signed != 'true'
-        run: |
-          echo "::warning::no signing key configured, so this APK is unsigned. Android refuses to install an unsigned package, so the release will carry no APK; the file stays a workflow artifact for whoever signs it."
-          {
-            echo "### No APK on this release"
-            echo ""
-            echo "The build produced an unsigned APK, which Android refuses to install."
-            echo "It is attached to this run as a workflow artifact only. Add the four"
-            echo "signing secrets to publish one."
-          } >> "$GITHUB_STEP_SUMMARY"
-
-      - uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v8
-        with:
-          name: android-apk
-          path: |
-            android/app/build/outputs/apk/release/*.apk
-            android/app/build/outputs/apk/release/*.sha256
-          if-no-files-found: error
-
-  publish:
-    name: attach the APK
-    needs: [resolve, build]
-    # Nothing to attach unless the build was signed. Skipping is the honest
-    # outcome: a release page with an APK nobody can install is worse than a
-    # release page without one.
-    if: needs.build.outputs.signed == 'true' && needs.resolve.outputs.publish == 'true'
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write
-    steps:
-      - uses: actions/download-artifact@9000827ccba6bdab643e8b6fd33ac0654aef8333 # v8.0.2
-        with:
-          name: android-apk
-          path: dist
-
-      # The release itself belongs to release-studio.yml, which creates it and
-      # owns its notes. This waits for it rather than creating a second one: two
-      # jobs racing to define a release is how a page ends up with the wrong body.
-      - name: Attach to the release
-        env:
-          GH_TOKEN: ${{ github.token }}
-          TAG: ${{ needs.resolve.outputs.tag }}
-          REPO: ${{ github.repository }}
-        run: |
-          # 90 minutes: on the last Studio release the page appeared about 40
-          # minutes after the tag run started, and the APK build spends part of
-          # that before this wait begins.
-          for _ in $(seq 1 360); do
-            if gh release view "$TAG" --repo "$REPO" >/dev/null 2>&1; then
-              gh release upload "$TAG" dist/*.apk dist/*.sha256 --repo "$REPO" --clobber
-              echo "attached to $TAG"
-              exit 0
-            fi
-            sleep 15
-          done
-          echo "::error::release $TAG did not appear within 90 minutes; the APK was built but not published"
-          exit 1
@@ -43,20 +43,20 @@ jobs:
        if: matrix.branch != ''
        run: echo "sha=$(git rev-parse HEAD)" >> "$GITHUB_OUTPUT"
      - name: Initialize CodeQL
-        uses: github/codeql-action/init@9b3254085494b55a0f7c4c2b3d2d52cc712ce370 # v4.39.1
+        uses: github/codeql-action/init@2892aa5e19bbd11bc0cff5427e3b750a04d9e3c2 # v4.38.2
        with:
          languages: ${{ matrix.language }}
          build-mode: ${{ matrix.language == 'go' && 'autobuild' || 'none' }}
      - name: Perform CodeQL Analysis
        if: matrix.branch == ''
-        uses: github/codeql-action/analyze@9b3254085494b55a0f7c4c2b3d2d52cc712ce370 # v4.39.1
+        uses: github/codeql-action/analyze@2892aa5e19bbd11bc0cff5427e3b750a04d9e3c2 # v4.38.2
        with:
          category: "/language:${{ matrix.language }}"
      # Results from a checked-out line are filed under that line's ref, not the
      # default branch the schedule happened to run from.
      - name: Perform CodeQL Analysis (${{ matrix.branch }})
        if: matrix.branch != ''
-        uses: github/codeql-action/analyze@9b3254085494b55a0f7c4c2b3d2d52cc712ce370 # v4.39.1
+        uses: github/codeql-action/analyze@2892aa5e19bbd11bc0cff5427e3b750a04d9e3c2 # v4.38.2
        with:
          category: "/language:${{ matrix.language }}"
          ref: refs/heads/${{ matrix.branch }}
@@ -24,7 +24,7 @@ jobs:
        with:
          version: 10
          run_install: false
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: "24"
          cache: pnpm
@@ -20,7 +20,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: '22'
          cache: npm
@@ -24,7 +24,7 @@ jobs:
        with:
          version: 10
          run_install: false
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: "24"
          cache: pnpm
@@ -1,44 +0,0 @@
-name: Deploy ops events relay
-
-on:
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-concurrency:
-  group: deploy-ops-events
-  cancel-in-progress: false
-
-jobs:
-  deploy:
-    if: github.ref == 'refs/heads/platform'
-    runs-on: ubuntu-latest
-    environment: production
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
-        with:
-          node-version: '22'
-          cache: npm
-          cache-dependency-path: workers/ops-events/package-lock.json
-      - name: Validate
-        working-directory: workers/ops-events
-        run: |
-          npm ci
-          npm run typecheck
-          npm test
-      - name: Deploy
-        working-directory: workers/ops-events
-        env:
-          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}
-        run: npx wrangler deploy
-      - name: Sync secrets
-        working-directory: workers/ops-events
-        env:
-          CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}
-          GITHUB_WEBHOOK_SECRET: ${{ secrets.OPS_GITHUB_WEBHOOK_SECRET }}
-          OPS_LISTEN_TOKEN: ${{ secrets.OPS_LISTEN_TOKEN }}
-          OPS_EMIT_TOKEN: ${{ secrets.OPS_EMIT_TOKEN }}
-        run: |
-          node -e 'process.stdout.write(JSON.stringify({GITHUB_WEBHOOK_SECRET:process.env.GITHUB_WEBHOOK_SECRET||null,OPS_LISTEN_TOKEN:process.env.OPS_LISTEN_TOKEN||null,OPS_EMIT_TOKEN:process.env.OPS_EMIT_TOKEN||null}))' | npx wrangler secret bulk
@@ -23,12 +23,12 @@ jobs:
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7

-      - uses: pnpm/action-setup@ea17c68df8912ef543352723c149a84f56e3d413 # v6.1.0
+      - uses: pnpm/action-setup@0ebf47130e4866e96fce0953f49152a61190b271 # v6.0.9
        with:
          version: 10
          run_install: false

-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: "24"
          cache: pnpm
@@ -51,7 +51,7 @@ jobs:
          go-version-file: go.mod
          cache: true

-      - uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0 # v7
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
        with:
          python-version: '3.12'

@@ -1,77 +0,0 @@
-name: Sync in-app feedback
-
-# Turns feedback submitted from Reasonix Studio into issues and reports issue
-# state (in progress, fixed in a tag, won't fix, duplicate) back to the worker,
-# and mirrors the reporter's replies onto their issue as bot comments.
-# Text from the public reaches the script only through the worker's JSON and
-# the GitHub API, never through a shell command line or an expression.
-
-on:
-  schedule:
-    - cron: '*/10 * * * *'
-    - cron: '5 3 * * *'
-  workflow_dispatch:
-    inputs:
-      dry_run:
-        description: Print what would happen without creating issues or writing back
-        required: false
-        default: false
-        type: boolean
-      full:
-        description: Re-read every open entry instead of only recently changed issues
-        required: false
-        default: false
-        type: boolean
-
-permissions:
-  contents: read
-  issues: write
-
-concurrency:
-  group: feedback-sync
-  cancel-in-progress: false
-
-jobs:
-  sync:
-    if: github.repository == 'esengine/DeepSeek-Reasonix'
-    runs-on: ubuntu-latest
-    timeout-minutes: 8
-    steps:
-      - name: Check secret
-        id: gate
-        env:
-          FEEDBACK_ADMIN_TOKEN: ${{ secrets.FEEDBACK_ADMIN_TOKEN }}
-        run: |
-          if [ -z "$FEEDBACK_ADMIN_TOKEN" ]; then
-            echo "::notice::FEEDBACK_ADMIN_TOKEN is not set; feedback sync skipped"
-            echo "run=false" >> "$GITHUB_OUTPUT"
-          else
-            echo "run=true" >> "$GITHUB_OUTPUT"
-          fi
-      - if: steps.gate.outputs.run == 'true'
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          # git tag --contains needs full history; the tag API would cost a request per entry.
-          fetch-depth: 0
-          persist-credentials: false
-      - if: steps.gate.outputs.run == 'true'
-        uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
-        with:
-          node-version: '22'
-      - name: Sync
-        if: steps.gate.outputs.run == 'true'
-        env:
-          FEEDBACK_ADMIN_TOKEN: ${{ secrets.FEEDBACK_ADMIN_TOKEN }}
-          # crash.reasonix.io answers runner IPs with a managed challenge; workers.dev serves only the converter's admin paths.
-          FEEDBACK_BASE_URL: https://reasonix-crash-report.reasonix.workers.dev
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          DRY_RUN: ${{ github.event.inputs.dry_run || 'false' }}
-          FULL: ${{ github.event.inputs.full || 'false' }}
-          SCHEDULE: ${{ github.event.schedule }}
-        run: |
-          set -euo pipefail
-          args=()
-          if [ "$DRY_RUN" = "true" ]; then args+=(--dry-run); fi
-          # The daily schedule sweeps everything the change window could have missed.
-          if [ "$FULL" = "true" ] || [ "$SCHEDULE" = "5 3 * * *" ]; then args+=(--full); fi
-          node scripts/feedback-sync.mjs "${args[@]}"
@@ -1,32 +0,0 @@
-name: Deploy website
-
-# GitHub requires workflow_dispatch to be registered on the default branch.
-# The production implementation is the same workflow path on website.
-on:
-  workflow_dispatch:
-    inputs:
-      release_version:
-        description: "Optional product release version associated with this publication"
-        required: false
-        default: ""
-        type: string
-      release_request:
-        description: "Optional release workflow correlation ID"
-        required: false
-        default: ""
-        type: string
-      studio_version:
-        description: "Studio version that must be present in the generated changelog"
-        required: false
-        default: ""
-        type: string
-
-permissions: {}
-
-jobs:
-  require-website:
-    runs-on: ubuntu-latest
-    steps:
-      - run: |
-          echo "::error::Deploy the website with --ref website; production is restricted to that branch."
-          exit 1
@@ -43,7 +43,7 @@ jobs:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
        with:
          fetch-depth: 0
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: '22'
      - name: Test sweeper
@@ -46,7 +46,7 @@ jobs:
          version: 10
          run_install: false

-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: "24"
          cache: pnpm
@@ -99,7 +99,7 @@ jobs:
            -c.extraMetadata.version="${{ steps.version.outputs.version }}"
          sha256sum ../../dist/ReasonixStudio-windows-amd64-installer.exe | tee -a "$GITHUB_STEP_SUMMARY"

-      - uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7.0.2
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
        with:
          name: studio-selftest-windows-${{ steps.version.outputs.version }}
          path: dist/ReasonixStudio-windows-amd64-installer.exe
@@ -16,7 +16,7 @@ permissions:

 concurrency:
  group: studio-${{ github.ref }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+  cancel-in-progress: true

 jobs:
  frontend:
@@ -33,7 +33,7 @@ jobs:
          version: 10
          run_install: false

-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: "24"
          cache: pnpm
@@ -81,218 +81,9 @@ jobs:
            exit 1
          fi

-  # Real-Chromium guards that run without a kernel. Only the ones that passed
-  # three consecutive local runs are wired; the rest stay manual (see
-  # desktop/frontend-next/perf/README.md). Skipped unless the SPA changed.
-  browser-guards:
-    name: browser guards
-    runs-on: ubuntu-latest
-    timeout-minutes: 25
-    defaults:
-      run:
-        working-directory: desktop/frontend-next
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          fetch-depth: 0
-
-      - name: SPA changed?
-        id: spa
-        working-directory: .
-        run: |
-          if [ "${{ github.event_name }}" = "pull_request" ] &&
-             git diff --quiet "origin/${{ github.base_ref }}...HEAD" -- desktop/frontend-next; then
-            echo "run=false" >> "$GITHUB_OUTPUT"
-          else
-            echo "run=true" >> "$GITHUB_OUTPUT"
-          fi
-
-      - if: steps.spa.outputs.run == 'true'
-        uses: pnpm/action-setup@ea17c68df8912ef543352723c149a84f56e3d413 # v6.1.0
-        with:
-          version: 10
-          run_install: false
-
-      - if: steps.spa.outputs.run == 'true'
-        uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
-        with:
-          node-version: "24"
-          cache: pnpm
-          cache-dependency-path: desktop/frontend-next/pnpm-lock.yaml
-
-      - if: steps.spa.outputs.run == 'true'
-        name: Install
-        run: pnpm install --frozen-lockfile
-
-      - if: steps.spa.outputs.run == 'true'
-        name: Build the perf harness and install Chromium
-        run: |
-          pnpm exec vite build --config vite.perf.config.ts
-          pnpm exec playwright install --with-deps chromium
-
-      - if: steps.spa.outputs.run == 'true'
-        name: Guards
-        run: |
-          python3 -m http.server 4399 --directory dist-perf >/dev/null 2>&1 &
-          for _ in $(seq 30); do curl -sf http://localhost:4399/perf.html >/dev/null && break; sleep 1; done
-          node perf/intrinsic.mjs
-          node perf/firstscreen-fonts.mjs
-          node perf/contrast.mjs
-          node perf/stickyfold.mjs
-          node perf/crumb.mjs
-          node perf/hoverradius.mjs
-          node perf/jump-ends.mjs
-          node perf/modelrows.mjs
-          node perf/explorersize.mjs
-          node perf/git-panel.mjs
-          node perf/session-move-to.mjs
-          node perf/shell-refused.mjs
-          node perf/external-origin.mjs
-          node perf/rail-search-hold.mjs
-          node perf/rail-toggle-panes.mjs
-          node perf/session-move.mjs
-          node perf/session-ref.mjs
-          node perf/browser-full.mjs
-          node perf/mod-ui.mjs
-          node perf/codeeditor-size.mjs
-          node perf/composer-reflow.mjs
-          node perf/composer-toolbar-row.mjs
-          node perf/action-icons.mjs
-          node perf/rowhit.mjs
-          node perf/prefs-wide-centered.mjs
-          node perf/tab-pip.mjs
-          node perf/syntax-tokens.mjs
-          node perf/tab-close-inside.mjs
-          node perf/context-card-hover.mjs
-          node perf/sessquick.mjs
-          node perf/sessionrows.mjs
-          node perf/session-row-marker.mjs
-          node perf/session-selected-row.mjs
-          node perf/wheel.mjs
-          node perf/branch-tooltip.mjs
-          node perf/branch-alignment.mjs
-          node perf/branch-menu.mjs
-          node perf/image-preview.mjs
-          node perf/user-images.mjs
-          node perf/toast-anchor.mjs
-          node perf/addfolder-rest.mjs
-          node perf/scroll-jump.mjs
-          node perf/add-project-selects.mjs
-          node perf/project-header.mjs
-          node perf/setup-defer.mjs
-          node perf/narrow.mjs
-          node perf/settings-rows.mjs
-          node perf/narrow-steps.mjs
-          node perf/job-names.mjs
-          node perf/mcp-actions.mjs
-          node perf/web-access.mjs
-          node perf/peers-settings.mjs
-          node perf/output-cap.mjs
-          node perf/agent-run-tab.mjs
-          node perf/changes-cap.mjs
-          node perf/session-menu-fit.mjs
-          node perf/rail-rows-typography.mjs
-          node perf/settings-nav-fit.mjs
-          node perf/rail-rows-align.mjs
-          node perf/progress-tail.mjs
-          node perf/trust-wait.mjs
-          node perf/nest-nodes.mjs
-          node perf/provider-ports.mjs
-          node perf/peer-threads.mjs
-          node perf/text-size.mjs
-          node perf/fs-scale.mjs
-          node perf/composer-feel.mjs
-          node perf/side-panel.mjs
-          node perf/card-shell.mjs
-          node perf/line-cards.mjs
-          node perf/detail-panel.mjs
-          node perf/brand-lockup.mjs
-          node perf/notice-line.mjs
-          node perf/brought-line.mjs
-          node perf/modal-focus.mjs
-          node perf/focus-ring.mjs
-          node perf/narrow-forms.mjs
-          node perf/i18n-rendered.mjs
-          node perf/narrow-rail.mjs
-          node perf/selfcheck.mjs branch-tooltip
-
-  # What a hosted page does over the relay, end to end: the production bundle,
-  # the real Studio host and kernel, a real Chromium; only the accounts API and
-  # the relay are stand-ins (tools/relayfake). Event-transport changes are
-  # judged here, because a unit test of one side cannot see the other.
-  relay-harness:
-    name: relay harness
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          fetch-depth: 0
-
-      - name: Transport changed?
-        id: scope
-        env:
-          BASE_REF: ${{ github.base_ref }}
-        run: |
-          paths="desktop/frontend-next internal/frontend/serve internal/platform/remotecloud internal/platform/account tools/relayfake cmd/reasonix-studio-host"
-          if [ "$GITHUB_EVENT_NAME" = "pull_request" ] &&
-             git diff --quiet "origin/$BASE_REF...HEAD" -- $paths; then
-            echo "run=false" >> "$GITHUB_OUTPUT"
-          else
-            echo "run=true" >> "$GITHUB_OUTPUT"
-          fi
-
-      - if: steps.scope.outputs.run == 'true'
-        uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
-        with:
-          go-version-file: go.mod
-          cache: true
-
-      - if: steps.scope.outputs.run == 'true'
-        uses: pnpm/action-setup@ea17c68df8912ef543352723c149a84f56e3d413 # v6.1.0
-        with:
-          version: 10
-          run_install: false
-
-      - if: steps.scope.outputs.run == 'true'
-        uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
-        with:
-          node-version: "24"
-          cache: pnpm
-          cache-dependency-path: desktop/frontend-next/pnpm-lock.yaml
-
-      - if: steps.scope.outputs.run == 'true'
-        name: Build the host, the stand-ins and the page
-        run: |
-          go build -o desktop/frontend-next/relay/bin/studio-host ./cmd/reasonix-studio-host
-          go build -o desktop/frontend-next/relay/bin/relayfake ./tools/relayfake
-          cd desktop/frontend-next
-          pnpm install --frozen-lockfile
-          pnpm exec vite build --outDir dist-relay --emptyOutDir
-          pnpm exec playwright install --with-deps chromium
-
-      - if: steps.scope.outputs.run == 'true'
-        name: Scenarios (hosted page over the relay; local page on a flapping network)
-        working-directory: desktop/frontend-next
-        env:
-          RELAY_HOST_BIN: ${{ github.workspace }}/desktop/frontend-next/relay/bin/studio-host
-          RELAY_FAKE_BIN: ${{ github.workspace }}/desktop/frontend-next/relay/bin/relayfake
-          RELAY_DIST: ${{ github.workspace }}/desktop/frontend-next/dist-relay
-        run: |
-          node relay/live.mjs
-          node relay/storm.mjs
-
-  # A pull request with no platform-relevant change (scripts/ci-platform-scope.sh)
-  # moves the Windows and macOS legs onto an ubuntu runner with every step
-  # skipped, so the check names still report without a scarce runner. A push to
-  # studio, or a failed `scope`, runs all three.
  desktop:
    name: desktop (${{ matrix.os }})
-    needs: scope
-    if: always()
-    runs-on: ${{ github.event_name == 'pull_request' && needs.scope.outputs.platform == 'false' && matrix.os != 'ubuntu-22.04' && 'ubuntu-latest' || matrix.os }}
-    env:
-      RUN_STEPS: ${{ github.event_name != 'pull_request' || needs.scope.outputs.platform != 'false' || matrix.os == 'ubuntu-22.04' }}
+    runs-on: ${{ matrix.os }}
    strategy:
      fail-fast: false
      matrix:
@@ -301,11 +92,9 @@ jobs:
      run:
        working-directory: desktop
    steps:
-      - if: env.RUN_STEPS == 'true'
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7

-      - if: env.RUN_STEPS == 'true'
-        uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
+      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
        with:
          go-version-file: desktop/go.mod
          cache: true
@@ -318,97 +107,20 @@ jobs:
      # carries build dependencies of its own.
      # Same command as `make studio-test`; this job is already in desktop/.
      - name: go test
-        if: env.RUN_STEPS == 'true'
        run: go test ./...

-      - if: env.RUN_STEPS == 'true'
-        uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
-        with:
-          node-version: "24"
-
-      - name: CUA process and packaging contracts
-        if: env.RUN_STEPS == 'true'
-        timeout-minutes: 2
-        run: node --test --test-name-pattern=CUA electron/test/unit.mjs
-
-      - name: Install Linux native CUA fixture dependencies
-        if: env.RUN_STEPS == 'true' && runner.os == 'Linux'
-        run: sudo apt-get update && sudo apt-get install -y xvfb openbox dbus-x11 libgtk-3-dev at-spi2-core
-
-      - name: Native CUA application acceptance
-        if: env.RUN_STEPS == 'true'
-        timeout-minutes: 6
-        shell: bash
-        run: |
-          if [ "$RUNNER_OS" = Linux ]; then
-            dbus-run-session -- xvfb-run -a sh -c 'openbox >/tmp/reasonix-openbox.log 2>&1 & export NO_AT_BRIDGE=0 GTK_MODULES=gail:atk-bridge; node electron/test/cua-native.mjs'
-          else
-            node electron/test/cua-native.mjs
-          fi
-
      # The host sits in the main module and nothing else builds it, so its
      # dependency closure is checked here: it must stay CGO-free and
      # platform-neutral or the Electron bundle cannot ship it. Cross-compiling
      # is host-independent, so once is enough.
      - name: host stays pure Go and platform-neutral
-        if: env.RUN_STEPS == 'true' && runner.os == 'Linux'
+        if: runner.os == 'Linux'
        working-directory: .
        run: |
          for os in windows linux darwin; do
-            for arch in amd64 arm64; do
-              CGO_ENABLED=0 GOOS=$os GOARCH=$arch go build -o /dev/null ./cmd/reasonix-studio-host
-            done
+            CGO_ENABLED=0 GOOS=$os GOARCH=amd64 go build -o /dev/null ./cmd/reasonix-studio-host
          done

-  # The ARM64 leg queues on a scarcer runner pool, so a pull request runs it
-  # only when it touches what the leg verifies; a push to studio always does.
-  # The macOS and Windows package legs follow `platform`, the same rule as the
-  # desktop matrix above.
-  scope:
-    name: package matrix
-    runs-on: ubuntu-latest
-    outputs:
-      matrix: ${{ steps.pick.outputs.matrix }}
-      arm64: ${{ steps.pick.outputs.arm64 }}
-      platform: ${{ steps.pick.outputs.platform }}
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          fetch-depth: 0
-
-      - id: pick
-        env:
-          EVENT: ${{ github.event_name }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha }}
-        run: |
-          arm64=true
-          platform=true
-          if [ "$EVENT" = pull_request ]; then
-            arm64=false
-            platform=$(git diff -z --name-only --no-renames "$BASE_SHA" HEAD | bash scripts/ci-platform-scope.sh)
-            if git diff --name-only --no-renames "$BASE_SHA" HEAD |
-              grep -qE '^(desktop/electron/|desktop/packaging/|desktop/cmd/update-helper/|desktop/cmd/studio-manifest/|desktop/go\.(mod|sum)$|cmd/reasonix-studio-host/|internal/platform/(update|appupdate)/|go\.(mod|sum)$|scripts/(linux-install-smoke|verify-linux-package|verify-linux-appimage|verify-studio-artifacts)\.sh$|scripts/(windows-signing-lib|verify-windows-package|verify-windows-authenticode|install-smoke)\.ps1$|\.github/workflows/(studio|release-studio)\.yml$)'; then
-              arm64=true
-            fi
-          fi
-          legs='[
-            {"os":"macos-latest","arch":"arm64","platform":"darwin"},
-            {"os":"ubuntu-22.04","arch":"amd64","platform":"linux"},
-            {"os":"windows-latest","arch":"amd64","platform":"windows","ebarch":"x64"}
-          ]'
-          if [ "$platform" = false ]; then
-            legs="$(jq -c 'map(select(.platform == "linux"))' <<<"$legs")"
-          elif [ "$arm64" = true ]; then
-            legs="$(jq -c '. + [{"os":"windows-11-arm","arch":"arm64","platform":"windows","ebarch":"arm64"},{"os":"ubuntu-22.04-arm","arch":"arm64","platform":"linux"}]' <<<"$legs")"
-          else
-            legs="$(jq -c . <<<"$legs")"
-          fi
-          {
-            echo "arm64=$arm64"
-            echo "platform=$platform"
-            echo "matrix={\"include\":$legs}"
-          } >> "$GITHUB_OUTPUT"
-
  # The packaging contract, on real artifacts rather than on the config that
  # describes them. Unsigned on purpose: signing and notarization are a
  # distribution concern and belong to the release workflow, and treating their
@@ -416,12 +128,20 @@ jobs:
  # cross-compile while only Windows had ever produced a package.
  package:
    name: package (${{ matrix.os }})
-    needs: scope
    runs-on: ${{ matrix.os }}
-    timeout-minutes: 45
    strategy:
      fail-fast: false
-      matrix: ${{ fromJSON(needs.scope.outputs.matrix) }}
+      matrix:
+        include:
+          - os: macos-latest
+            arch: arm64
+            platform: darwin
+          - os: ubuntu-22.04
+            arch: amd64
+            platform: linux
+          - os: windows-latest
+            arch: amd64
+            platform: windows
    defaults:
      run:
        shell: bash
@@ -439,40 +159,24 @@ jobs:
          version: 10
          run_install: false

-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: "24"
          cache: pnpm
          cache-dependency-path: desktop/electron/pnpm-lock.yaml

-      # electron-builder shells out to dpkg-deb through fakeroot for the .deb and
-      # to rpmbuild for the .rpm.
+      # electron-builder shells out to dpkg-deb through fakeroot for the .deb.
      - name: Install Linux packaging deps
        if: runner.os == 'Linux'
-        run: sudo apt-get update && sudo apt-get install -y fakeroot rpm cpio xvfb xauth policykit-1
-
-      # The fpm electron-builder carries is an x86_64 program, so on an ARM64
-      # runner the .deb and .rpm are built with one installed from RubyGems.
-      - name: Use a native fpm on ARM64 Linux
-        if: runner.os == 'Linux' && runner.arch == 'ARM64'
-        run: |
-          sudo apt-get install -y ruby ruby-dev build-essential
-          sudo gem install --no-document fpm -v 1.16.0
-          echo "USE_SYSTEM_FPM=true" >> "$GITHUB_ENV"
+        run: sudo apt-get update && sudo apt-get install -y fakeroot

      - name: Build the SPA
        working-directory: desktop/frontend-next
        run: pnpm install --frozen-lockfile && pnpm build

-      # The Windows legs name their architecture instead of taking the
-      # toolchain's, so a runner whose Go or Node is the other architecture
-      # fails here rather than shipping a mislabelled package.
      - name: Build the kernel the shell speaks to
        working-directory: desktop/electron
-        env:
-          REASONIX_GOARCH: ${{ matrix.platform == 'windows' && matrix.arch || '' }}
-          REASONIX_VERSION: v0.0.0-ci
-        run: pnpm install --frozen-lockfile && pnpm test && pnpm run build:host
+        run: pnpm install --frozen-lockfile && pnpm run build:host

      # The .deb's absolute-path payload. A no-op off Linux, so it is not
      # guarded here: the one place it must not be forgotten is the one place
@@ -487,8 +191,7 @@ jobs:
        working-directory: desktop/electron
        env:
          CSC_IDENTITY_AUTO_DISCOVERY: "false"
-          ELECTRON_BUILDER_7Z_FILTER: ${{ matrix.platform == 'windows' && matrix.arch == 'arm64' && 'BCJ' || '' }}
-        run: pnpm run build:dist ${{ matrix.ebarch && format('--{0}', matrix.ebarch) || '' }} -c.extraMetadata.version=0.0.0-ci
+        run: pnpm run build:dist

      # The release vocabulary is GOARCH. electron-builder's own is not, and the
      # rename that reconciles them runs here rather than in the manifest.
@@ -504,99 +207,6 @@ jobs:
            exit 1
          fi

-      # An emulated x64 program runs on ARM64 Windows, so a package that merely
-      # starts proves nothing about its architecture: every PE image in the
-      # bundle has to be built for the one the package is named for.
-      - name: Every PE image is built for the package's architecture
-        if: matrix.platform == 'windows'
-        run: |
-          dir=$(find dist -maxdepth 1 -type d -name 'win*-unpacked' | head -1)
-          test -n "$dir"
-          node desktop/electron/packaging/pe.js "$dir" "${{ matrix.arch }}"
-
-      # An update starts the installer with --updated, and that is what keeps the
-      # shortcuts (and the taskbar pin hanging off them) from being deleted and
-      # recreated. Installed twice over itself the way an update does it, the
-      # desktop shortcut has to still be the same file, pointing at a path with
-      # no version in it, or every update costs the user their pin.
-      - name: An update keeps the shortcuts and they name no version
-        if: matrix.platform == 'windows'
-        shell: pwsh
-        run: |
-          $installer = (Get-ChildItem dist -Filter 'ReasonixStudio-windows-*-installer.exe' | Select-Object -First 1).FullName
-          $dir = Join-Path $env:RUNNER_TEMP 'rx-update-guard'
-          Start-Process $installer -ArgumentList '/S', "/D=$dir" -Wait
-          $lnk = Join-Path ([Environment]::GetFolderPath('Desktop')) 'Reasonix Studio.lnk'
-          $target = (New-Object -ComObject WScript.Shell).CreateShortcut($lnk).TargetPath
-          if ($target -ne (Join-Path $dir 'Reasonix Studio.exe')) { throw "the desktop shortcut targets $target, not the stable executable" }
-          $mark = [datetime]'2020-01-01'
-          (Get-Item $lnk).LastWriteTime = $mark
-          Start-Process $installer -ArgumentList '/S', '--updated', "/D=$dir" -Wait
-          if ((Get-Item $lnk).LastWriteTime -ne $mark) { throw 'an update rewrote the desktop shortcut' }
-          Start-Process (Join-Path $dir 'Uninstall Reasonix Studio.exe') -ArgumentList '/S', '/currentuser' -Wait
-
-      # The release packages a signed copy of the bundle and then reads both
-      # packages back against it. The same recipe on an unsigned copy, so the
-      # packaging contract (the installer's one application archive, the pinned
-      # elevate.exe, every PE image's architecture) is exercised on pull
-      # requests and not first by a release.
-      - name: The packages read back against the bundle they were made from
-        if: matrix.platform == 'windows'
-        shell: bash
-        env:
-          ELECTRON_BUILDER_7Z_FILTER: ${{ matrix.arch == 'arm64' && 'BCJ' || '' }}
-        run: |
-          rm -rf payload nsis-src packed zip-src
-          cp -r "$(find dist -maxdepth 1 -type d -name 'win*-unpacked' | head -1)" payload
-          rm -f payload/resources/elevate.exe
-          # The NSIS run writes elevate.exe into the tree it packs, so each
-          # packager gets a copy and payload stays what was signed.
-          cp -r payload zip-src
-          cp -r payload nsis-src
-          cd desktop/electron
-          pnpm exec electron-builder --prepackaged ../../zip-src --win zip --${{ matrix.ebarch }} \
-            -c.extraMetadata.version=0.0.0-ci -c.directories.output=../../packed-zip
-          pnpm exec electron-builder --prepackaged ../../nsis-src --win nsis --${{ matrix.ebarch }} \
-            -c.extraMetadata.version=0.0.0-ci -c.directories.output=../../packed
-          mv ../../packed-zip/ReasonixStudio-windows-${{ matrix.arch }}.zip ../../packed/
-          ls -la ../../packed
-
-      - name: verify-windows-package.ps1 accepts them
-        if: matrix.platform == 'windows'
-        shell: pwsh
-        run: |
-          $ErrorActionPreference = 'Stop'
-          . ./scripts/windows-signing-lib.ps1
-          $digest = Get-ManifestDigest -Manifest @(Get-TreeManifest -Root payload)
-          ./scripts/verify-windows-package.ps1 `
-            -PayloadDirectory payload `
-            -ExpectedPayloadDigest $digest `
-            -InstallerPath "packed/ReasonixStudio-windows-${{ matrix.arch }}-installer.exe" `
-            -PortableArchivePath "packed/ReasonixStudio-windows-${{ matrix.arch }}.zip" `
-            -OutputPath "$env:RUNNER_TEMP/verify-output.txt" `
-            -Architecture ${{ matrix.arch }}
-          Get-Content "$env:RUNNER_TEMP/verify-output.txt"
-
-      # The release's own install smoke, against the unsigned installer: the
-      # package is installed where it is meant to run, its tree checked, and
-      # removed.
-      - name: The package installs, runs and uninstalls
-        if: matrix.platform == 'windows'
-        shell: pwsh
-        run: |
-          ./scripts/install-smoke.ps1 `
-            -Installer "dist/ReasonixStudio-windows-${{ matrix.arch }}-installer.exe" `
-            -Architecture ${{ matrix.arch }}
-
-      - name: Keep the arm64 installer for the x64 refusal check
-        if: matrix.platform == 'windows' && matrix.arch == 'arm64'
-        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7.0.2
-        with:
-          name: studio-arm64-installer-ci
-          path: dist/ReasonixStudio-windows-arm64-installer.exe
-          if-no-files-found: error
-          retention-days: 1
-
      # The update channel is the signed manifest. A second answer to which
      # version is current is not a convenience. The .blockmap is the same
      # channel's differential index and carries the release prefix, so it also
@@ -623,17 +233,47 @@ jobs:
          echo "page:   ${page:-<missing>}"
          test -n "$host" -a -n "$page"

-      - name: Linux packages carry their declared architecture and update payload
+      # Read off the package rather than off the config that describes it. Both
+      # of these went wrong the same way once: electron-builder names a package
+      # after package.json unless told, and fpm records the owner it finds on
+      # disk. Either one leaves a Studio that cannot update itself on Linux and
+      # says nothing about it until a user asks.
+      - name: The .deb is Studio's own package and carries what updates it
        if: runner.os == 'Linux'
-        run: bash scripts/verify-linux-package.sh "dist/ReasonixStudio-linux-${{ matrix.arch }}.deb" "dist/ReasonixStudio-linux-${{ matrix.arch }}.rpm" "${{ matrix.arch }}"
+        run: |
+          deb=$(ls dist/ReasonixStudio-linux-*.deb)
+          name=$(dpkg-deb -f "$deb" Package)
+          if [ "$name" != "reasonix-studio" ]; then
+            echo "::error::the package is $name; dpkg holds that beside the installed reasonix-studio instead of upgrading it, and update.StudioLine asks for it by name"
+            exit 1
+          fi
+          dpkg-deb -c "$deb" > contents.txt
+          # The path is the polkit action's exec.path annotation, so it is the
+          # only one authorization is granted for; root because polkit will not
+          # run a helper the invoking user could have written. Either spelling
+          # of root: dpkg-deb prints the numeric ids it cannot map, and which
+          # one it prints is a property of the machine reading the package
+          # rather than of the package.
+          if ! grep -qE '^-rwxr-xr-x (root/root|0/0) +[0-9]+ .* \./usr/lib/reasonix-studio/reasonix-studio-update-helper$' contents.txt; then
+            echo "::error::no root-owned executable update helper at the path the polkit action names"
+            grep reasonix-studio-update-helper contents.txt || echo "(not in the package at all)"
+            exit 1
+          fi
+          if ! grep -q './usr/share/polkit-1/actions/io.reasonix.studio.update.policy$' contents.txt; then
+            echo "::error::the polkit action is not in the package, so pkexec has nothing to authorize against"
+            exit 1
+          fi

-      - name: The Linux package installs, runs, reinstalls and uninstalls
+      # The Electron shell sets the package's glibc floor; a Go binary linked
+      # with cgo would raise it to the runner's own.
+      - name: The .deb's Go binaries are statically linked
        if: runner.os == 'Linux'
-        run: bash scripts/linux-install-smoke.sh "dist/ReasonixStudio-linux-${{ matrix.arch }}.deb" "${{ matrix.arch }}" 0.0.0-ci
-
-      - name: The AppImage carries its architecture, the kernel and the page
-        if: runner.os == 'Linux'
-        run: bash scripts/verify-linux-appimage.sh "dist/ReasonixStudio-linux-${{ matrix.arch }}.AppImage" "${{ matrix.arch }}"
+        run: |
+          deb=$(ls dist/ReasonixStudio-linux-*.deb)
+          dpkg-deb -x "$deb" "$RUNNER_TEMP/deb"
+          host=$(find "$RUNNER_TEMP/deb" -type f -path '*/resources/bin/reasonix-studio-host')
+          test -n "$host"
+          node desktop/electron/packaging/elf.js "$host" "$RUNNER_TEMP/deb/usr/lib/reasonix-studio/reasonix-studio-update-helper"

      # Recorded, not asserted: the real bundle layout, so later work argues
      # from what electron-builder produced rather than from its documentation.
@@ -661,27 +301,3 @@ jobs:
            echo "::error::the manifest has no entry for ${{ matrix.platform }}-${{ matrix.arch }}"
            exit 1
          fi
-
-  # The arm64 installer carries one application and no other. Run anywhere but
-  # ARM64 Windows it has to say so and stop, instead of exiting 0 with nothing
-  # installed.
-  arm64-installer-refuses-x64:
-    name: arm64 installer refuses an x64 machine
-    needs: [scope, package]
-    if: needs.scope.outputs.arm64 == 'true'
-    runs-on: windows-latest
-    steps:
-      - uses: actions/download-artifact@9000827ccba6bdab643e8b6fd33ac0654aef8333 # v8.0.2
-        with:
-          name: studio-arm64-installer-ci
-          path: installer
-
-      - name: Run it silently
-        shell: pwsh
-        run: |
-          $ErrorActionPreference = 'Stop'
-          $target = Join-Path $env:RUNNER_TEMP 'rx-refused'
-          $run = Start-Process -FilePath installer/ReasonixStudio-windows-arm64-installer.exe -ArgumentList '/S', "/D=$target" -Wait -PassThru
-          Write-Host "exit code $($run.ExitCode)"
-          if ($run.ExitCode -eq 0) { throw 'the arm64 installer exited 0 on an x64 machine' }
-          if (@(Get-ChildItem -Recurse -Force $target -ErrorAction SilentlyContinue).Count -ne 0) { throw 'the arm64 installer wrote into the target on an x64 machine' }
@@ -14,7 +14,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: '22'
      - name: Update README contributor tables
@@ -33,6 +33,6 @@ jobs:
          body: |
            Refreshes the README acknowledgments tables from the GitHub contributors API.

-            - Source: `repos/esengine/DeepSeek-Reasonix/contributors?per_page=50&anon=1`
-            - The repository owner, declared aliases, tool identities, bots and entries without a GitHub login are left out (`scripts/credit-identities.json`).
+            - Source: `repos/esengine/DeepSeek-Reasonix/contributors?per_page=20&anon=1`
+            - Anonymous contributors are shown without email addresses.
          labels: documentation
@@ -27,7 +27,7 @@ jobs:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
        with:
          fetch-depth: 0
-      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
        with:
          node-version: '22'
      - name: Test chart generator
@@ -1,84 +0,0 @@
-name: Withdraw a Studio version
-
-# Pulls a published Studio release out of the update catalog (or puts it back),
-# so installs stop being offered it without waiting for a new release. Only the
-# catalog changes: the release's files stay where they are, and a withdrawn
-# release can be restored with the same run.
-on:
-  workflow_dispatch:
-    inputs:
-      version:
-        description: "Version to withdraw or restore, for example 2.36.0"
-        required: true
-        type: string
-      reason:
-        description: "Why it is withdrawn (shown to users on that version); ignored when restoring"
-        required: false
-        type: string
-      restore:
-        description: "Put a withdrawn version back instead of withdrawing it"
-        required: false
-        type: boolean
-        default: false
-
-concurrency:
-  group: studio-catalog
-  cancel-in-progress: false
-
-jobs:
-  withdraw:
-    name: ${{ inputs.restore && 'restore' || 'withdraw' }} ${{ inputs.version }}
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    environment: studio-release
-    permissions:
-      contents: write
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-        with:
-          ref: studio
-          persist-credentials: false
-
-      - name: Update the Studio catalog
-        env:
-          VERSION: ${{ inputs.version }}
-          REASON: ${{ inputs.reason }}
-          RESTORE: ${{ inputs.restore }}
-          R2_BUCKET: ${{ secrets.R2_BUCKET }}
-          HAS_R2: ${{ secrets.R2_ACCESS_KEY_ID != '' && secrets.R2_SECRET_ACCESS_KEY != '' && secrets.R2_ACCOUNT_ID != '' && secrets.R2_BUCKET != '' }}
-        run: |
-          if [ "$HAS_R2" != "true" ]; then
-            echo "::error::R2 credentials are not configured, so the catalog cannot be changed"
-            exit 1
-          fi
-          if ! [[ "$VERSION" =~ ^[0-9]+\.[0-9]+\.[0-9]+(-[0-9A-Za-z.]+)?$ ]]; then
-            echo "::error::version must look like 2.36.0"
-            exit 1
-          fi
-          aws configure set aws_access_key_id "${{ secrets.R2_ACCESS_KEY_ID }}"
-          aws configure set aws_secret_access_key "${{ secrets.R2_SECRET_ACCESS_KEY }}"
-          aws configure set default.region auto
-          endpoint="https://${{ secrets.R2_ACCOUNT_ID }}.r2.cloudflarestorage.com"
-          at="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
-          current=/tmp/studio-versions-current.json
-          next=/tmp/studio-versions-next.json
-          aws s3 cp "s3://${R2_BUCKET}/studio/versions.json" "$current" --endpoint-url "$endpoint"
-          if [ "$RESTORE" = "true" ]; then
-            bash scripts/withdraw-version.sh "$current" --restore "$VERSION" "$at" > "$next"
-            jq -e --arg v "$VERSION" '.versions | map(select((.version | ltrimstr("v")) == $v)) | length == 1' "$next" >/dev/null
-          else
-            bash scripts/withdraw-version.sh "$current" "$VERSION" "$REASON" "$at" > "$next"
-            jq -e --arg v "$VERSION" '(.versions | map(select((.version | ltrimstr("v")) == $v)) | length == 0) and (.withdrawn | map(select((.version | ltrimstr("v")) == $v)) | length == 1)' "$next" >/dev/null
-          fi
-          aws s3 cp "$next" "s3://${R2_BUCKET}/studio/versions.json" \
-            --endpoint-url "$endpoint" \
-            --content-type "application/json; charset=utf-8" \
-            --cache-control "public, max-age=300, stale-if-error=86400"
-          echo "Studio catalog updated for $VERSION (restore=$RESTORE)"
-
-      - name: Keep the GitHub release off the latest slot
-        if: ${{ !inputs.restore }}
-        env:
-          GH_TOKEN: ${{ github.token }}
-          VERSION: ${{ inputs.version }}
-        run: gh release edit "studio-v${VERSION}" --prerelease --latest=false --repo "$GITHUB_REPOSITORY"
@@ -15,7 +15,6 @@
 /runtime-resume
 /extension-protocol-gen
 /reasonix-plugin-example
-/mcp-line-counter-kit
 /reasonix-studio-host
 /signpath-contract
 /studio-delta
@@ -95,4 +94,3 @@ scripts/official-theme-art/__pycache__/
 # running, which on Windows is every rebuild while Studio is open.
 *.exe
 *.exe~
-/relayfake
@@ -11,7 +11,6 @@ fingerprint_files:
  - .github/workflows/release-studio.yml
  - .github/workflows/studio-certum-signing-smoke.yml
  - .github/actions/setup-certum/action.yml
-  - scripts/install-smoke.ps1
  - scripts/sign-certum.ps1
  - scripts/verify-windows-authenticode.ps1
  - scripts/verify-windows-package.ps1
@@ -117,22 +117,6 @@ go test ./internal/runtime/agent/ -v            # verbose, one package
 go test ./internal/tools/builtin/ -run TestGrep  # one test
 ```

-`testenv.IsolateUserState` installs a disposable user home and config-edit lock
-scope. Default-path tests can clear `REASONIX_HOME` without losing the test
-registry, while explicitly isolated Reasonix homes retain their own locks.
-
-`REASONIX_TEST_HOME` carries the scope to helper subprocesses. A helper that
-skips the isolation guard must install the inherited home with
-`testscope.InstallHome` before running tests.
-
-The environment variable alone cannot activate a test scope. Production
-executables still use the OS account's home for shared config locks.
-
-The guard removes its scratch home on normal completion. Scratch homes are
-created under the original home, so an interrupted run can still leave
-`.reasonix-test-home-*` directories; this is separate from persistent
-config-edit locks in the real `.reasonix/locks` directory.
-
 ### Code style

 - `gofmt` is enforced by CI — format before committing
@@ -197,35 +181,6 @@ porting to `studio`; do not open the same PR against both.
 5. Ensure `gofmt -l .` shows no changes
 6. Submit a pull request to that same branch

-## Pull request policy
-
-**Related changes.** If you have several related changes of one pattern, a
-tracking issue plus one pull request is easiest to review.
-
-**Issue first for features.** A change that adds behaviour rather than fixing a
-defect should link an issue the maintainers have agreed to before the work
-starts. A fix for a reproducible defect may open directly, with the failing test
-in the PR. Documentation-only and test-only PRs are reviewed at lower priority.
-
-**Review tiers.**
-
- Changes under the `sensitive:` paths declared in `REASONIX.md` get a full
-  review plus a security review, regardless of author.
- Every PR carries the four sections of the PR template: Cause, Blast radius,
-  Neighbouring behaviours tested, Why this layer.
- The new test must fail without the change.
- A UI change (`desktop/frontend-next/src` or the website pages) fills the
-  template's UI section: screenshots before and after, light and dark, wide
-  and phone width.
- The same section carries the line "Entries moved, hidden or removed".
- Hiding or removing a visible control needs a linked issue and a release-notes
-  line naming it, and updates `docs/STUDIO_PARITY.md` for a 1.x feature.
- Review holds these; no check reads them.
-
-**Transparency.** The PR template has an optional checkbox for disclosing AI
-assistance. It is for transparency only and does not change how a change is
-reviewed.
-
 ## Reporting issues

 Open an issue on GitHub with:
@@ -9,7 +9,7 @@ GOEXE := $(shell go env GOEXE)
 # One pin for the Makefile and the CI lint job; see .github/workflows/ci.yml.
 GOLANGCI_VERSION := $(shell cat .golangci-version)

-.PHONY: build vet fmt lint lint-go lint-install lint-cross lint-update coverage-gate coverage-gate-update pricecheck acp-conformance check test studio-test sdk-test sdk-test-race hooks cross clean studio
+.PHONY: build vet fmt lint lint-go lint-install lint-cross lint-update coverage-gate coverage-gate-update pricecheck check test studio-test sdk-test sdk-test-race hooks cross clean studio

 build:
 	CGO_ENABLED=0 go build -ldflags "$(LDFLAGS)" -o bin/reasonix$(GOEXE) ./cmd/reasonix
@@ -87,13 +87,6 @@ lint-cross:
 test: studio-test sdk-test
 	go test ./...

-# Drives `reasonix acp` with the official ACP TypeScript SDK and holds every
-# message it writes to the schema that SDK ships. Needs node and network for
-# `npm ci`, which runs without install scripts against a lockfile.
-acp-conformance: build
-	cd tools/acp-conformance && npm ci --ignore-scripts --no-audit --no-fund
-	cd tools/acp-conformance && REASONIX_BIN=$(CURDIR)/bin/reasonix$(GOEXE) npm test
-
 # The shell is a nested module, so the root `test` target never reaches it.
 studio-test:
 	cd desktop && go test ./...
@@ -18,8 +18,6 @@
  <a href="https://esengine.github.io/DeepSeek-Reasonix/">Website</a>
  &nbsp;·&nbsp;
  <strong><a href="https://discord.gg/XF78rEME2D">Discord</a></strong>
-  &nbsp;·&nbsp;
-  <strong><a href="https://qm.qq.com/q/i59b0z2R8s">QQ</a></strong>
 </p>

 <p align="center">
@@ -33,7 +31,6 @@
  <a href="https://github.com/esengine/DeepSeek-Reasonix/graphs/contributors"><img src="https://img.shields.io/github/contributors/esengine/DeepSeek-Reasonix.svg?style=flat-square&color=bc8cff&labelColor=161b22&logo=github&logoColor=white" alt="contributors"/></a>
  <a href="https://github.com/esengine/DeepSeek-Reasonix/discussions"><img src="https://img.shields.io/github/discussions/esengine/DeepSeek-Reasonix.svg?style=flat-square&color=58a6ff&labelColor=161b22&logo=github&logoColor=white" alt="Discussions"/></a>
  <a href="https://discord.gg/XF78rEME2D"><img src="https://img.shields.io/badge/discord-join-5865F2.svg?style=flat-square&labelColor=161b22&logo=discord&logoColor=white" alt="Discord"/></a>
-  <a href="https://qm.qq.com/q/i59b0z2R8s"><img src="https://img.shields.io/badge/QQ%20group-1093562660-12B7F5.svg?style=flat-square&labelColor=161b22&logo=tencentqq&logoColor=white" alt="QQ group 1093562660"/></a>
 </p>

 <p align="center">
@@ -43,73 +40,23 @@

 <br/>

-<p align="center"><strong>Open source · MIT · runs on your computer, model of your choice</strong></p>
-<h3 align="center">An open-source coding agent for your terminal, desktop, browser and editor.</h3>
-<p align="center">It reads your project, edits files and runs commands and tests. Have it ask before each step, or work on its own within limits you set. Files its edit tools changed can be rewound. New here? Start with the Studio desktop app.</p>
+<p align="center"><strong>Open source · MIT · a single Go binary</strong></p>
+<h3 align="center">A coding agent you can leave running.</h3>
+<p align="center">One local engine, four ways in — terminal, desktop app, browser, or your editor over ACP. Plan mode, permissions, a workspace sandbox and per-turn checkpoints keep a long autonomous run something you can still read and undo.</p>
 <p align="center">Maintained by <strong>Huahui Yu</strong>.</p>

 > [!IMPORTANT]
 > **Community · 加入社区** — bilingual Discord for setup help (`#help` / `#求助`), workflow showcases, and feature ideas. → **<https://discord.gg/XF78rEME2D>**
->
-> QQ group **DeepSeek-Reasonix官方群** (`1093562660`) → **<https://qm.qq.com/q/i59b0z2R8s>** ([QR code](./docs/assets/qq-group.svg))
-
-> Douyin: **做游戏的小鱼** (Douyin ID `22703872788`) · [QR code](./docs/assets/douyin.png)

 <br/>

-## ❤️Sponsor
-
-<table><tr><td width="180" align="center" valign="middle"><a href="https://www.rightapi.ai/register?aff=e083bf73" target="_blank" rel="sponsored noopener noreferrer"><img src="./.github/sponsor/rightcode.png" alt="Right Code" width="64"></a></td><td valign="middle">Thanks to Right Code for sponsoring this project! Right Code reliably provides routing services for models such as Claude Code, Codex, and Gemini, with both pay-as-you-go and monthly subscription billing options available. Invoices are available upon top-up, and enterprise and team users can receive dedicated one-on-one support. Register via <a href="https://www.rightapi.ai/register?aff=e083bf73" target="_blank" rel="sponsored noopener noreferrer">this link</a>.</td></tr><tr><td width="180" align="center" valign="middle"><a href="https://www.packyapi.ai/register?aff=rIbQ" target="_blank" rel="sponsored noopener noreferrer"><picture><source media="(prefers-color-scheme: dark)" srcset="./.github/sponsor/packycode-dark.svg"><img src="./.github/sponsor/packycode-light.svg" alt="PackyCode" width="80"></picture></a></td><td valign="middle">Thanks to PackyCode for sponsoring this project! Access leading AI models through PackyCode with one API endpoint and one API key, with automatic failover and dedicated high-speed routes for Codex and Claude Code. Register via <a href="https://www.packyapi.ai/register?aff=rIbQ" target="_blank" rel="sponsored noopener noreferrer">this link</a>.</td></tr></table>
-
-## What it is
-
-Reasonix is an open-source AI coding assistant. The program runs on your computer; the AI model is a service you configure, which receives what it reads. You describe a task in plain language; it reads code, edits files and runs commands and tests inside the project folder you choose, and shows you each step.
-
-You pick the model: a DeepSeek preset is built in, any OpenAI-compatible service is one config entry, and you pay the model provider directly.
-
-## What you get
-
- **Read, edit, run.** It reads code, edits files and runs tests and commands, all inside your project folder.
- **Permission levels.** Have it ask every time, or let everyday actions run on their own; your forbid rules always hold. To see the approach first, use plan mode: it starts work only once you approve.
- **Rewind.** A file is saved before it changes, so you can rewind by conversation turn without touching your git history. Only files its edit tools changed are covered, not what a shell command did.
-
-## 30-second start
-
-Download, add a model key, open a project. New here? Start with the Studio desktop app.
-
-1. Download and install the desktop app, Reasonix Studio, from the [releases page](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true).
-2. Open it and enter your model provider and key when asked.
-3. Open your project folder and write what you want, for example "add retry on failure to the http client".
-4. At the level you chose, it stops and asks before anything that needs your say-so. Allow it to continue, or refuse.
-
-Prefer a terminal? Download the `reasonix` command-line tool, run `reasonix setup` to pick a model, then run `reasonix`. See [Install](#install) and [Quick start](#quick-start).
-
-## Four ways to use it
-
- **Desktop app (Studio).** A graphical app; the best place to start.
- **Terminal.** A full-screen text interface; `reasonix run "task"` also works in scripts.
- **Browser.** `reasonix web` opens a web interface on your own machine.
- **Editor.** Install the extension in VS Code; other editors that support ACP (Agent Client Protocol, a common way for an editor to connect to an AI assistant) can connect to `reasonix acp`.
-
-All four are the same program underneath. What follows is for advanced users and contributors.
-
-## Use it from DeepSeek Harness
-
-DeepSeek Harness (`dsh`) can hand a task to Reasonix as a subagent over ACP. This repository publishes a plugin index for it:
-
-1. Run `reasonix setup` on the machine that runs `dsh`. In the DSH Web UI turn on Custom plugin markets (Plugins, More) and add this index in the Discover tab: `https://raw.githubusercontent.com/esengine/DeepSeek-Reasonix/studio/integrations/dsh-plugin/market-index.json`
-2. Install "Reasonix subagent" from it.
-3. Open the plugin's page and turn its switch on. While it is off the model has no `reasonix` tool and may run `reasonix -p` in its own shell, which looks like a delegation and is not one.
-
-Setup details, the permission policy and limits are in [Using Reasonix from DeepSeek Harness](./docs/DSH.md).
-
 ## Versions

 Reasonix ships on two lines. See the [version roadmap announcement](https://github.com/esengine/DeepSeek-Reasonix/discussions/10748) for the reasoning behind the split.

 | Line | Branch | Status | Get it |
 | --- | --- | --- | --- |
-| **Reasonix 2.x** | `studio` (this branch) | Active development | [Studio releases](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true) |
+| **Reasonix 2.x** | `studio` (this branch) | Active development, pre-release | [Studio releases](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true) |
 | **Reasonix 1.x** | [`main-v2`](https://github.com/esengine/DeepSeek-Reasonix/tree/main-v2) | Maintenance / stable | `npm i -g reasonix` · [desktop download](https://reasonix.io/?download=desktop#start) |

 - **Want something stable?** Stay on 1.x. It keeps receiving bug fixes,
@@ -128,8 +75,6 @@ or reimplemented on 2.x where they fit the new architecture.

 ## Features

-Technical highlights for advanced users.
-
 - **Config-driven.** Providers, the agent, enabled tools, and plugins are all
  declared in `reasonix.toml`. No hardcoded models.
 - **Multi-model & composable.** DeepSeek ships as a preset; any
@@ -154,14 +99,14 @@ Studio, and editor integrations all run the same local Reasonix engine.

 Download the package for your platform from the latest
 [Studio release](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true)
-(tagged `studio-v2.*`; candidates such as `-rc.N` are marked as
-pre-releases):
+(tagged `studio-v2.*`, published as pre-releases while 2.x is in active
+development):

 | Platform | Package | Architecture |
 | --- | --- | --- |
 | macOS | `.dmg` or `.zip` | Apple Silicon (`arm64`) / Intel (`amd64`) |
 | Windows | Installer `-installer.exe` or portable `.zip` | x64 |
-| Linux | `.deb`, `.rpm`, or `.AppImage` (download only) | x64 (`amd64`) / ARM64 (`arm64`, aarch64) |
+| Linux | `.deb` | x64 |

 Every package ships with a `.minisig` signature, and the release carries
 `SHA256SUMS`. Once installed, Studio updates itself in place.
@@ -169,19 +114,6 @@ Every package ships with a `.minisig` signature, and the release carries
 The same release also carries the 2.x `reasonix` CLI as archives for
 `darwin|linux|windows × amd64|arm64`.

-The Android build ships on that release as `ReasonixMobile-<version>.apk`
-(`arm64`, Android 8.0 or newer). The kernel and the page run on the phone, and
-the agent reaches your servers over SSH; pairing with a computer is done by
-scanning the code it shows.
-
-The APK is signed with the Android release key rather than minisign, so it
-carries a `.sha256` and no `.minisig`.
-
-When that key is absent the build is named
-`ReasonixMobile-<version>-unsigned.apk`. Android refuses to install an unsigned
-package, so those builds stay a workflow artifact for whoever wants to sign one,
-and the release carries no APK at all.
-
 ### Reasonix 1.x (stable)

 The 1.x CLI installs through npm on any supported platform, or Homebrew on
@@ -263,14 +195,12 @@ For advanced CLI usage and configuration, see the **[CLI reference](./docs/CLI.m
 ## Documentation

 - **Getting started:** [Guide](./docs/GUIDE.md) · [CLI reference](./docs/CLI.md) ·
-  [Configuration paths](./docs/CONFIG_PATHS.md) · [ACP editor integration](./docs/ACP.md) ·
-  [Use from DeepSeek Harness](./docs/DSH.md)
+  [Configuration paths](./docs/CONFIG_PATHS.md) · [ACP editor integration](./docs/ACP.md)
 - **Features & troubleshooting:** [Subagent profiles](./docs/SUBAGENT_PROFILES.md) ·
  [Context Engine v2](./docs/SESSION_MEMORY_RETRIEVAL.md) ·
  [Capability diagnostics](./docs/CAPABILITY_DIAGNOSTICS.md) ·
  [Recovery and updates](./docs/RECOVERY.md) ·
  [Checkpoints & rewind](./docs/CHECKPOINTS.md)
- **Agent presets:** [Session Agent compositions](./docs/AGENT_COMPOSITIONS.md)
 - **Engineering & migration:** [Spec](./docs/SPEC.md) ·
  [Task contracts & pause policy](./docs/TASK_CONTRACT.md) ·
  [Tool contract](./docs/TOOL_CONTRACT.md) · [Moving from 1.x to 2.x](./docs/MIGRATING.md) ·
@@ -301,11 +231,11 @@ A small list of folks whose work has shaped Reasonix the most — the current to
 <!-- reasonix-top-contributors:start -->
 | Contributor | Contributor | Contributor | Contributor |
 | --- | --- | --- | --- |
-| [**SivanCola**](https://github.com/SivanCola) | [**KHG420**](https://github.com/KHG420) | [**ttmouse**](https://github.com/ttmouse) | [**chuiwenwei**](https://github.com/chuiwenwei) |
-| [**lifu963**](https://github.com/lifu963) | [**XTLine**](https://github.com/XTLine) | [**HUQIANTAO**](https://github.com/HUQIANTAO) | [**xxoingr**](https://github.com/xxoingr) |
-| [**Harbor404**](https://github.com/Harbor404) | [**GTC2080**](https://github.com/GTC2080) | [**mchenziyi**](https://github.com/mchenziyi) | [**Li-Charles-One**](https://github.com/Li-Charles-One) |
-| [**light-front-theory**](https://github.com/light-front-theory) | [**imcreazy**](https://github.com/imcreazy) | [**eghrhegpe**](https://github.com/eghrhegpe) | [**noreply**](https://github.com/noreply) |
-| [**par73e**](https://github.com/par73e) | [**Bernardxu123**](https://github.com/Bernardxu123) | [**lancy69**](https://github.com/lancy69) | [**Guan0923**](https://github.com/Guan0923) |
+| [**SivanCola**](https://github.com/SivanCola) | [**esengine**](https://github.com/esengine) | [**ttmouse**](https://github.com/ttmouse) | [**lifu963**](https://github.com/lifu963) |
+| **reasonix** | [**HUQIANTAO**](https://github.com/HUQIANTAO) | [**GTC2080**](https://github.com/GTC2080) | [**mchenziyi**](https://github.com/mchenziyi) |
+| [**Li-Charles-One**](https://github.com/Li-Charles-One) | **merge-order-check** | [**light-front-theory**](https://github.com/light-front-theory) | **Yan Li** |
+| [**eghrhegpe**](https://github.com/eghrhegpe) | **wufengfan** | [**Bernardxu123**](https://github.com/Bernardxu123) | [**HaoyueQin**](https://github.com/HaoyueQin) |
+| [**CVEngineer66**](https://github.com/CVEngineer66) | [**KHG420**](https://github.com/KHG420) | [**JesonChou**](https://github.com/JesonChou) | [**SuMuxi66**](https://github.com/SuMuxi66) |
 <!-- reasonix-top-contributors:end -->

 Special thanks to [**Bernardxu123**](https://github.com/Bernardxu123) for designing the project logo and intro video.
@@ -335,9 +265,8 @@ coffee, not a contract — donations don't buy feature priority or change how
 issues get triaged.

 - **International** — PayPal: [paypal.me/yuhuahui](https://paypal.me/yuhuahui)
- **国内** — 微信支付 / 支付宝（扫码）
+- **国内** — 微信支付（扫码）

 <p align="center">
  <img src=".github/sponsor/wechat-pay.jpg" alt="WeChat Pay QR code" width="180"/>
-  <img src=".github/sponsor/alipay.jpg" alt="Alipay QR code" width="180"/>
 </p>
@@ -18,8 +18,6 @@
  <a href="https://esengine.github.io/DeepSeek-Reasonix/">官方网站</a>
  &nbsp;·&nbsp;
  <strong><a href="https://discord.gg/XF78rEME2D">Discord</a></strong>
-  &nbsp;·&nbsp;
-  <strong><a href="https://qm.qq.com/q/i59b0z2R8s">QQ</a></strong>
 </p>

 <p align="center">
@@ -33,7 +31,6 @@
  <a href="https://github.com/esengine/DeepSeek-Reasonix/graphs/contributors"><img src="https://img.shields.io/github/contributors/esengine/DeepSeek-Reasonix.svg?style=flat-square&color=bc8cff&labelColor=161b22&logo=github&logoColor=white" alt="contributors"/></a>
  <a href="https://github.com/esengine/DeepSeek-Reasonix/discussions"><img src="https://img.shields.io/github/discussions/esengine/DeepSeek-Reasonix.svg?style=flat-square&color=58a6ff&labelColor=161b22&logo=github&logoColor=white" alt="Discussions"/></a>
  <a href="https://discord.gg/XF78rEME2D"><img src="https://img.shields.io/badge/discord-join-5865F2.svg?style=flat-square&labelColor=161b22&logo=discord&logoColor=white" alt="Discord"/></a>
-  <a href="https://qm.qq.com/q/i59b0z2R8s"><img src="https://img.shields.io/badge/QQ%20group-1093562660-12B7F5.svg?style=flat-square&labelColor=161b22&logo=tencentqq&logoColor=white" alt="QQ group 1093562660"/></a>
 </p>

 <p align="center">
@@ -43,63 +40,13 @@

 <br/>

-<p align="center"><strong>开源 · MIT · 程序跑在你电脑上，模型自选</strong></p>
-<h3 align="center">终端、桌面、浏览器、编辑器里都能用的 AI 编程助手。</h3>
-<p align="center">它会读你的项目、改文件、跑命令和测试。每一步先问你，或在你划定的范围内自己做，由你选；它用编辑工具改过的文件可以退回去。第一次用，从桌面应用 Studio 开始。</p>
+<p align="center"><strong>开源 · MIT · 单个 Go 二进制</strong></p>
+<h3 align="center">可以一直开着跑的编码 Agent。</h3>
+<p align="center">一套本地引擎,四个入口——终端、桌面端、浏览器,或通过 ACP 接入你的编辑器。计划模式、权限、工作区沙箱与逐轮 checkpoint,让长时间自治运行始终可读、可撤销。</p>
 <p align="center">维护者：<strong>Huahui Yu</strong>。</p>

 > [!IMPORTANT]
 > **加入社区 · Community** — 双语 Discord，提供安装答疑（`#help` / `#求助`）、工作流展示与功能想法。→ **<https://discord.gg/XF78rEME2D>**
->
-> QQ 群 **DeepSeek-Reasonix官方群**（群号 `1093562660`）→ **<https://qm.qq.com/q/i59b0z2R8s>**（[二维码](./docs/assets/qq-group.svg)）
-
-> 抖音：**做游戏的小鱼**（抖音号 `22703872788`）· [二维码](./docs/assets/douyin.png)
-
-## ❤️赞助商
-
-<table><tr><td width="180" align="center" valign="middle"><a href="https://www.rightapi.ai/register?aff=e083bf73" target="_blank" rel="sponsored noopener noreferrer"><img src="./.github/sponsor/rightcode.png" alt="Right Code" width="64"></a></td><td valign="middle">感谢 Right Code 赞助了本项目！Right Code 稳定提供 Claude Code、Codex、Gemini 等模型的中转服务，并可选按量、包月两种计费模式。充值即可开票，企业、团队用户一对一对接。通过<a href="https://www.rightapi.ai/register?aff=e083bf73" target="_blank" rel="sponsored noopener noreferrer">此链接</a>注册。</td></tr><tr><td width="180" align="center" valign="middle"><a href="https://www.packyapi.ai/register?aff=rIbQ" target="_blank" rel="sponsored noopener noreferrer"><picture><source media="(prefers-color-scheme: dark)" srcset="./.github/sponsor/packycode-dark.svg"><img src="./.github/sponsor/packycode-light.svg" alt="PackyCode" width="80"></picture></a></td><td valign="middle">感谢 PackyCode 赞助了本项目！PackyCode 是一家稳定、高效的 API 中转服务商，统一域名、统一密钥、智能容灾切换。通过<a href="https://www.packyapi.ai/register?aff=rIbQ" target="_blank" rel="sponsored noopener noreferrer">此链接</a>注册。</td></tr></table>
-
-## 它是什么
-
-Reasonix 是一个开源的 AI 编程助手。程序跑在你的电脑上；AI 模型是你自己配置的服务，它读到的内容会发给这个服务。你用大白话描述任务，它在你选的项目文件夹里读代码、改文件、跑命令和测试，每一步都摆给你看。
-
-模型你自己选：内置 DeepSeek 预设，其他兼容 OpenAI 接口的服务填一条配置就行，费用直接付给模型服务商。
-
-## 你能得到什么
-
- **读、改、跑**：读你的项目、改文件、跑测试和命令，都在你选的项目文件夹里。
- **权限分档**：每次都问你，或让日常操作自动放行；你设的禁止规则始终有效。想先看方案，就开计划模式，批准后它才动手。
- **可回退**：改文件前先存快照，按对话轮次回退，不碰 git 历史。只覆盖编辑工具改的文件，命令行命令造成的改动不在内。
-
-## 30 秒上手
-
-下载、填模型密钥、打开项目。第一次用，从桌面应用 Studio 开始。
-
-1. 从[发布页](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true)下载并安装桌面应用 Reasonix Studio。
-2. 打开应用，按提示填入模型服务商和密钥。
-3. 打开你的项目文件夹，写下想做的事，例如“给 http 客户端加上失败重试”。
-4. 按你选的档位，该问你的操作它会先停下来问；同意就继续，不同意就拒绝。
-
-更喜欢终端的话，下载 `reasonix` 命令行后运行 `reasonix setup` 配好模型，再运行 `reasonix` 即可。详见[安装](#安装)和[快速开始](#快速开始)。
-
-## 四种用法
-
- **桌面应用（Studio）**：图形界面，适合第一次使用。
- **终端**：全屏文字界面，也能用 `reasonix run "任务"` 在脚本里运行。
- **浏览器**：`reasonix web` 在本机打开网页界面。
- **编辑器**：VS Code 安装扩展即可；其他支持 ACP（Agent Client Protocol，编辑器接入某个 AI 助手的通用约定）的编辑器，可连接 `reasonix acp`。
-
-四种用法背后是同一套程序。下面的内容写给进阶用户和贡献者。
-
-## 在 DeepSeek Harness 里使用
-
-DeepSeek Harness（`dsh`）可以通过 ACP 把任务交给 Reasonix 作为子智能体去做。本仓库为此发布了一个插件索引：
-
-1. 在运行 `dsh` 的机器上先运行 `reasonix setup`。在 DSH 网页界面打开“自定义插件市场”（插件，更多），再到 Discover 页添加这个索引：`https://raw.githubusercontent.com/esengine/DeepSeek-Reasonix/studio/integrations/dsh-plugin/market-index.json`
-2. 从中安装“Reasonix subagent”。
-3. 打开这个插件的页面，把它的开关打开。关着的时候模型没有 `reasonix` 工具，可能改在自己的 shell 里直接运行 `reasonix -p`，看起来像委派，其实不是。
-
-配置细节、权限策略和限制见 [在 DeepSeek Harness 中使用 Reasonix](./docs/DSH.md)。

 ## 版本

@@ -107,7 +54,7 @@ Reasonix 分为两条版本线，调整的原因见[版本路线公告](https://

 | 版本线 | 分支 | 状态 | 获取方式 |
 | --- | --- | --- | --- |
-| **Reasonix 2.x** | `studio`（当前分支） | 活跃开发 | [Studio 发布页](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true) |
+| **Reasonix 2.x** | `studio`（当前分支） | 活跃开发，预发布 | [Studio 发布页](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true) |
 | **Reasonix 1.x** | [`main-v2`](https://github.com/esengine/DeepSeek-Reasonix/tree/main-v2) | 维护 / 稳定 | `npm i -g reasonix` · [桌面端下载](https://reasonix.io/?download=desktop#start) |

 - **想要稳定**：继续使用 1.x。它会持续收到 Bug 修复、Provider / API 兼容、
@@ -122,8 +69,6 @@ Reasonix 分为两条版本线，调整的原因见[版本路线公告](https://

 ## 特性

-以下是写给进阶用户的技术要点。
-
 - **配置驱动**：provider、agent、启用的工具、插件全部在 `reasonix.toml` 中声明，
  内核无硬编码模型。
 - **多模型 · 可组合**：DeepSeek 作为预设内置；任何 OpenAI 兼容
@@ -143,28 +88,19 @@ Reasonix 分为两条版本线，调整的原因见[版本路线公告](https://
 ### Reasonix Studio（2.x）

 从最新的 [Studio 发布](https://github.com/esengine/DeepSeek-Reasonix/releases?q=studio-v&expanded=true)
-下载对应平台的安装包（标签为 `studio-v2.*`；`-rc.N` 等候选版本标为预发布）：
+下载对应平台的安装包（标签为 `studio-v2.*`，2.x 活跃开发期间以预发布形式发布）：

 | 平台 | 安装包 | 架构 |
 | --- | --- | --- |
 | macOS | `.dmg` 或 `.zip` | Apple Silicon（`arm64`）/ Intel（`amd64`） |
 | Windows | 安装器 `-installer.exe` 或便携 `.zip` | x64 |
-| Linux | `.deb`、`.rpm` 或 `.AppImage`（只提供下载） | x64（`amd64`）/ ARM64（`arm64`，aarch64） |
+| Linux | `.deb` | x64 |

 每个安装包都附带 `.minisig` 签名，发布页另有 `SHA256SUMS`。安装后 Studio 会在
 应用内自行更新。

 同一个发布页也提供 2.x 的 `reasonix` CLI 归档（`darwin|linux|windows × amd64|arm64`）。

-安卓版在同一个发布页上，文件名是 `ReasonixMobile-<版本>.apk`（`arm64`，Android 8.0 及以上）。
-
-这个 APK 用 Android 发布密钥签名，不走 minisign，因此只有一个 `.sha256`，没有 `.minisig`。
-
-该密钥缺失时产物名为 `ReasonixMobile-<版本>-unsigned.apk`。Android 不接受未签名的包，
-这类构建只留作 workflow 产物供自行签名，发布页上不会挂 APK。
-
-内核与界面都跑在手机上，Agent 通过 SSH 连接你自己的服务器；与电脑配对时扫它显示的二维码。
-
 ### Reasonix 1.x（稳定版）

 1.x CLI 在任意支持的平台上都可以通过 npm 安装，macOS 也可以使用 Homebrew：
@@ -240,13 +176,12 @@ CLI 进阶用法和详细配置见 **[CLI 命令参考](./docs/CLI.zh-CN.md)**

 - **开始使用：** [指南](./docs/GUIDE.zh-CN.md) ·
  [CLI 命令参考](./docs/CLI.zh-CN.md) · [配置路径](./docs/CONFIG_PATHS.md) ·
-  [ACP 编辑器接入](./docs/ACP.md) · [在 DeepSeek Harness 中使用](./docs/DSH.md)
+  [ACP 编辑器接入](./docs/ACP.md)
 - **功能与排障：** [子智能体 Profile](./docs/SUBAGENT_PROFILES.md) ·
  [Context Engine v2](./docs/SESSION_MEMORY_RETRIEVAL.md) ·
  [能力诊断](./docs/CAPABILITY_DIAGNOSTICS.md) ·
  [恢复与安全模式](./docs/RECOVERY.md) ·
  [Checkpoints 与 rewind](./docs/CHECKPOINTS.md)
- **Agent 预设：** [会话级组合](./docs/AGENT_COMPOSITIONS.zh.md)
 - **工程与迁移：** [规格](./docs/SPEC.md) ·
  [任务合约与暂停策略](./docs/TASK_CONTRACT.md) ·
  [工具合约](./docs/TOOL_CONTRACT.md) ·
@@ -278,11 +213,11 @@ CLI 进阶用法和详细配置见 **[CLI 命令参考](./docs/CLI.zh-CN.md)**
 <!-- reasonix-top-contributors:start -->
 | Contributor | Contributor | Contributor | Contributor |
 | --- | --- | --- | --- |
-| [**SivanCola**](https://github.com/SivanCola) | [**KHG420**](https://github.com/KHG420) | [**ttmouse**](https://github.com/ttmouse) | [**chuiwenwei**](https://github.com/chuiwenwei) |
-| [**lifu963**](https://github.com/lifu963) | [**XTLine**](https://github.com/XTLine) | [**HUQIANTAO**](https://github.com/HUQIANTAO) | [**xxoingr**](https://github.com/xxoingr) |
-| [**Harbor404**](https://github.com/Harbor404) | [**GTC2080**](https://github.com/GTC2080) | [**mchenziyi**](https://github.com/mchenziyi) | [**Li-Charles-One**](https://github.com/Li-Charles-One) |
-| [**light-front-theory**](https://github.com/light-front-theory) | [**imcreazy**](https://github.com/imcreazy) | [**eghrhegpe**](https://github.com/eghrhegpe) | [**noreply**](https://github.com/noreply) |
-| [**par73e**](https://github.com/par73e) | [**Bernardxu123**](https://github.com/Bernardxu123) | [**lancy69**](https://github.com/lancy69) | [**Guan0923**](https://github.com/Guan0923) |
+| [**SivanCola**](https://github.com/SivanCola) | [**esengine**](https://github.com/esengine) | [**ttmouse**](https://github.com/ttmouse) | [**lifu963**](https://github.com/lifu963) |
+| **reasonix** | [**HUQIANTAO**](https://github.com/HUQIANTAO) | [**GTC2080**](https://github.com/GTC2080) | [**mchenziyi**](https://github.com/mchenziyi) |
+| [**Li-Charles-One**](https://github.com/Li-Charles-One) | **merge-order-check** | [**light-front-theory**](https://github.com/light-front-theory) | **Yan Li** |
+| [**eghrhegpe**](https://github.com/eghrhegpe) | **wufengfan** | [**Bernardxu123**](https://github.com/Bernardxu123) | [**HaoyueQin**](https://github.com/HaoyueQin) |
+| [**CVEngineer66**](https://github.com/CVEngineer66) | [**KHG420**](https://github.com/KHG420) | [**JesonChou**](https://github.com/JesonChou) | [**SuMuxi66**](https://github.com/SuMuxi66) |
 <!-- reasonix-top-contributors:end -->

 特别感谢 [**Bernardxu123**](https://github.com/Bernardxu123) 设计的项目 logo和开场视频。
@@ -310,10 +245,9 @@ CLI 进阶用法和详细配置见 **[CLI 命令参考](./docs/CLI.zh-CN.md)**
 如果 Reasonix 帮你省了时间或 token，欢迎请杯咖啡。捐助不会换来 feature
 优先级，也不会影响 issue 的处理顺序——就是「谢谢」。

- **国内** — 微信支付 / 支付宝（扫下方二维码）
+- **国内** — 微信支付（扫下方二维码）
 - **海外** — PayPal: [paypal.me/yuhuahui](https://paypal.me/yuhuahui)

 <p align="center">
  <img src=".github/sponsor/wechat-pay.jpg" alt="微信支付收款码" width="180"/>
-  <img src=".github/sponsor/alipay.jpg" alt="支付宝收款码" width="180"/>
 </p>
@@ -219,11 +219,9 @@ otherwise the block the model already has goes on reading as current.
 - `#<note>` in chat quick-adds an always-on instruction. The `remember` tool
  instead saves a fallible background fact (frontmatter file + `MEMORY.md`
  index). Fact `type` classifies content; independent `scope` controls whether it
-  is project-only (the default) or explicitly global.
- The saved-fact index stays out of the stable prefix and is reached through the
-  `memory` tool; pinned bodies and global user/feedback bodies load into it, the
-  latter as lower-priority compatibility guidance. The current turn receives a
-  tail note.
+  is project-only (the default) or explicitly global. The index loads into the
+  stable prefix on the next session; global user/feedback bodies also load as
+  lower-priority compatibility guidance. The current turn receives a tail note.

 ## Notes

@@ -322,31 +320,6 @@ Use `go test ./path/to/target/` to detect cycles **before** pushing. A `[setup f
 - **Keep the PR diff minimal.** Only the files relevant to the PR's purpose — no stray changes from other branches.
 - **Amend, don't add commits, for review feedback** — keeps the commit history clean.

-## Change discipline
-
-**Assess before editing.** For every change, humans and agents list callers,
-consumers of the same state, sibling features, reload/replay/restart paths,
-error and edge states, TUI/CLI and desktop equivalents, and wire mirrors.
-
-A fix that breaks a neighbouring behaviour is a failed fix.
-
-**Test the neighbourhood.** UI changes add or extend interaction tests for
-adjacent functions of the same component or flow (composer: submit, clear,
-queue, withdraw, restore, steer, failure, replay after reload) and run the whole
-affected suite.
-
-Write tests first; they must fail without the change.
-
-**Fix at the owning layer.** Use the existing unified pattern: no special-case
-flags, duplicated paths, or workarounds at the symptom site.
-
-If the design is duplicated or wrong, make the smallest unifying change:
-one path instead of two, deleting what becomes dead. Use current language and
-framework idioms; do not add a new pattern where one exists.
-
-**PR evidence.** Every PR description carries: Cause, Blast radius,
-Neighbouring behaviours tested, and Why this layer.
-
 ## Reasonix host checks

 The paths below decide what the agent is allowed to do — a wrong change here
@@ -369,14 +342,9 @@ out of how a path is spelled, which cannot tell `internal/auth` from
 - sensitive: internal/ext/pluginpkg/**
 - sensitive: internal/base/netclient/**
 - sensitive: internal/safety/redirectguard/**
- sensitive: internal/safety/endpointclient/**
 - sensitive: internal/safety/egress/**
 - sensitive: internal/platform/browser/**
 - sensitive: internal/platform/computer/**
- sensitive: internal/platform/gitpush/**
- sensitive: internal/session/control/git_push.go
- sensitive: internal/frontend/serve/git_push.go
- sensitive: internal/frontend/serve/host_routes.go

 One declaration, two effects: the same list is also the coverage gate's subject.
 `make coverage-gate` holds each path to the coverage it already has, because
@@ -407,19 +375,14 @@ observation, never by a check that ran afterwards.
 What follows bounds Observable today. It is not a restatement of the invariant:
 moving a number here must not move the promise.

- `scanWorkspace` walks the workspace and every granted extra root recursively,
-  skips VCS store dirs, and is `complete` only below 50k files in total. Over that, or past a walk error, it
+- `scanWorkspace` walks the workspace recursively, skips VCS store dirs, and is
+  `complete` only below 50k files. Over that, or past a walk error, it
  establishes nothing and the mutation stays unproven.
 - Symlinks are recorded by `Lstat`, so a write through one lands outside what
  the walk compares. Confining the resolved target is the sandbox's job.
 - `internal/safety/sandbox` confines writes to the workspace, configured extras, temp
-  and toolchain caches on macOS and Linux.
- On Windows a user-enabled write-restricted token bounds the *creation* of files
-  (create, append, overwrite, `mkdir`, HKCU writes, global git config, writes through links).
- There, deleting or renaming away an existing file is not bounded, and with the
-  setting off nothing the host enforces bounds `Writable`.
- The caches a command may write can be poisoned for the user's later unsandboxed
-  builds, on every platform.
+  and toolchain caches — on macOS and Linux. Windows has no OS-level bash
+  sandbox, so there nothing the host enforces bounds `Writable`.
 - The governed host authorities are a *named* set — ssh-agent, docker, podman.
  DBus, Wayland, X11, gpg-agent, keyrings and arbitrary sockets stay reachable,
  so the claim is "these endpoints are governed", never "host IPC is isolated".
@@ -465,13 +428,9 @@ stop a connect. Possession is not authority: the key can be unreadable while an
 agent still signs with it.

 A protection claim exists only where a backend has a demonstrated enforcement
-primitive for that dimension. Unsupported dimensions stay explicitly unsupported,
-because a claim that outruns its enforcement is worse than an absent one: it is
-believed.
-
-Windows enforces Integrity on file creation only, when the user turns it on.
-Deletion and rename, Confidentiality and Authority (the named pipes of ssh-agent
-and docker, COM and RPC services) are not enforced there.
+primitive for that dimension. Unsupported dimensions stay explicitly unsupported
+— Windows currently enforces none of the three — because a claim that outruns
+its enforcement is worse than an absent one: it is believed.

 **Known unmeasured surface — macOS host-mediated delegation:** Apple Events and
 Mach-service lookup are not denied by the current allow-default profile. Prior
@@ -1,22 +0,0 @@
-# Build output and local tooling state.
-build/
-.gradle/
-.kotlin/
-local.properties
-*.iml
-.idea/
-
-# The signing key and the file that names it.
-#
-# A key is not part of the project: it is the identity one person ships under,
-# and publishing it would let anyone else sign an APK that installs over
-# everyone's. Neither the keystore nor its password belongs in version control.
-# `app/build.gradle` reads both from here when present and falls back to the
-# debug key when they are not, so a fresh clone builds without them.
-keystore/
-
-# Build inputs, produced by scripts/ before Gradle runs.
-# jniLibs/ holds the cross-compiled kernel, assets/page/ the built
-# frontend. assets/LICENSE is hand-placed and stays tracked.
-app/src/main/jniLibs/
-app/src/main/assets/page/
@@ -1,216 +0,0 @@
---
-owner: @esengine
-backup: @SivanCola
-status: active
-reviewed: 2026-10-20
---
-
-# Reasonix Mobile — the Android shell
-
-Runs Reasonix as an Android app: the kernel ships inside the APK and starts on the
-device, the interface is the same `desktop/frontend-next` build the desktop Studio
-serves, and the agent reaches servers over SSH.
-
-This file is for people building it. What a person installing it needs — what it
-does, how to pair with a computer, what it cannot do — belongs in the release
-notes.
-
-## One command
-
-```bash
-./scripts/build.sh release     # or debug
-```
-
-Three steps, in this order:
-
-1. `build-kernel.sh` → `app/src/main/jniLibs/arm64-v8a/libreasonix.so`
-2. `sync-page.sh` → `app/src/main/assets/page/`
-3. `./gradlew assembleRelease`
-
-Output: `app/build/outputs/apk/release/app-release.apk`
-
-The kernel and the page are **inputs to packaging, not Gradle build steps**.
-Running `gradle assembleRelease` on its own fails on purpose and names the script
-that produces whichever input is missing (`requireBuildInputs` in
-`app/build.gradle`): an APK with no kernel in it should not be buildable by
-accident.
-
-## Requirements
-
-| Component | Version |
-| --- | --- |
-| JDK | 21 (17 also works) |
-| Android SDK | `platforms;android-35`, `build-tools;35.0.0`, `platform-tools` |
-| Gradle | the committed wrapper, 8.14.3 (for AGP 8.11.1) |
-| Go | 1.26+ |
-| Node / pnpm | Node 24+ / pnpm 10 |
-
-```bash
-export ANDROID_HOME="$HOME/Android/Sdk"
-export ANDROID_SDK_ROOT="$ANDROID_HOME"
-export JAVA_HOME="/path/to/jdk-21"
-export PATH="$PATH:$(go env GOPATH)/bin"
-```
-
-`ANDROID_HOME` is the one thing nothing sets for you.
-
-## Scripts
-
-| Script | What it does |
-| --- | --- |
-| `build.sh [debug\|release]` | The whole chain |
-| `build-kernel.sh` | Cross-compiles the kernel only (`GOOS=android GOARCH=arm64 CGO_ENABLED=0`) |
-| `sync-page.sh` | Builds the frontend, copies `dist/` into assets, copies `LICENSE` |
-| `verify-page.sh` | **No phone needed**: starts the real kernel against the packaged page |
-| `run.sh [debug\|release]` | `adb install`, launch, follow the log |
-
-`verify-page.sh` is the one to run after a change: it starts the kernel with the
-same arguments `Kernel.kt` uses (`--auth token`, `--token-file`, `--port-file`,
-`-page`) and checks that it binds a port and answers 200. Without a device, that
-is as deep as the chain can be verified.
-
-## Where the code is
-
-| File | Owns |
-| --- | --- |
-| `MainActivity.kt` | The WebView, the back gesture, downloads, external links, renderer restart |
-| `Kernel.kt` | Unpacking the page, starting and stopping the kernel, waiting for the port, health probes, logs, the workspace directory |
-| `KernelService.kt` | The foreground service that keeps a long turn alive |
-| `NativeBridge.kt` | `@JavascriptInterface`: what the page may ask the phone for |
-| `BrowserShellServer.kt` | Loopback HTTP server the kernel's browser tools drive the overlay WebView through |
-| `Downloader.kt` | Writing a kernel-produced file through the system document picker |
-
-The page reaches the shell through `window.reasonixHost` and `ReasonixNative`,
-defined in `desktop/frontend-next/src/port/host.ts` (`AndroidHost`).
-
-**Phone and desktop share one page build.** What differs only on a phone is
-decided at runtime, not by a second set of components. `phoneShell()` in
-`src/port/host.ts` and the `[data-shell="android"]` selector in `mobile.css` both
-read `<html data-shell="android">`, which `index.html` sets before the bundle
-loads.
-
-One attribute feeding both is what keeps style and components from disagreeing
-about what a phone is.
-
-## What the page may ask the phone for
-
-`NativeBridge` is reachable from whatever document the WebView holds, so every
-method that touches the device asks `trusted()` first.
-
-The activity answers by comparing the **main frame's** origin against the
-kernel's and against the one paired computer the person scanned a code for. A
-page that navigated elsewhere is not trusted, even though it may still hold the
-object.
-
-Five methods answer without the check — `version`, `systemTheme`, `safeArea`,
-`setLayerOpen`, `setRunning` — because they report back what the page already
-knows about its own window and grant nothing.
-
-`openInApp(url)` is the pairing decision, not a navigation one: it loads a URL in
-this window only for the origin the scanner just returned, and only once.
-Without that, any page holding the kernel origin could hand itself the bridge.
-
-External links leave through the browser. A non-web scheme is routed to the
-system only for `mailto:` and `tel:`; everything else is dropped, because an
-`intent://` URL can start an arbitrary exported activity with caller-supplied
-extras.
-
-## Permissions, and what they cost
-
-`MANAGE_EXTERNAL_STORAGE` — opening a folder on the phone as a workspace. The
-kernel is a Go binary: it opens real paths with `open(2)`, and SAF hands back a
-`content://` URI it cannot open.
-
-Requested from the folder-picker action, never at launch. Android shows no dialog
-for it and opens system settings instead.
-
-**There is no OS-level sandbox around this kernel.** The PR sets the sandbox to
-off on Android, so confining the agent to the chosen folder is Reasonix's own
-write-root check and nothing else. A shell tool with all-files access can reach
-all shared storage.
-
-`REQUEST_IGNORE_BATTERY_OPTIMIZATIONS` — Doze freezes the app group, which stops
-a turn that is still running. Asked once when a turn actually runs, never at
-launch, and the person can refuse. Refusing costs battery life during long turns
-and nothing else.
-
-**Google Play restricts this permission**: fine for a sideloaded build, a
-store-policy risk if this is ever listed.
-
-`FOREGROUND_SERVICE` / `FOREGROUND_SERVICE_DATA_SYNC` — keeping the kernel alive
-while the screen is off. `dataSync` is capped per 24 hours from Android 15. It is
-the closest honest description of "an agent is working".
-
-`CAMERA` — reading the pairing code a computer shows. Requested when the person
-asks to scan, never at launch. `ScanActivity` is not exported.
-
-`READ_EXTERNAL_STORAGE` / `WRITE_EXTERNAL_STORAGE` — the same folder walk below
-API 33. `maxSdkVersion`-gated; above 33 they are superseded by the permission
-above.
-
-`allowBackup=false` with explicit extraction rules, so a `keystore` or a token in
-app storage never leaves the device in a cloud backup.
-
-## Cleartext HTTP
-
-The kernel serves its page on `127.0.0.1:<port>`, and a paired computer serves
-the same page over its LAN address. Both are plain HTTP.
-
-Neither address can be written into a `network-security-config`: the port is
-chosen at start, the LAN address at scan time, and the format matches literals
-and domains, not subnets.
-
-So the config permits cleartext and the boundary is in the code instead.
-`shouldOverrideUrlLoading` sends every other http(s) origin to the browser, and
-the bridge answers only the kernel and the just-paired computer. A plaintext page
-on the LAN cannot become the document holding the bridge.
-
-## What goes into the APK
-
-| Content | Size | Why it is shaped this way |
-| --- | --- | --- |
-| `lib/arm64-v8a/libreasonix.so` | 45 MB | Go static binary, stripped. It is a `.so` rather than a file in `assets/` because only `nativeLibraryDir` is guaranteed executable — a binary copied to app storage fails on a device whose data partition is mounted `noexec`. |
-| `assets/page/` | 14 MB | The frontend build, including the Inter and Noto Sans SC fonts, `noCompress` |
-| `assets/LICENSE` | 1 KB | The repository's MIT licence, verbatim. MIT requires the notice and the licence text to accompany a binary distribution. |
-
-`assets/LICENSE` is generated by `sync-page.sh` from the repository's root
-`LICENSE`, not committed a second time: a checked-in copy drifts when the root
-file is updated.
-
-It sits in `assets/` rather than `assets/page/` because that directory is removed
-and rebuilt on every build.
-
-A release APK is about 31 MB.
-
-## Building
-
-**The Gradle wrapper is committed**, jar and all. `build.sh` refuses to run
-without `gradlew` rather than generating one from whatever `gradle` is on `PATH`:
-the dependency versions are part of the build.
-
-`gradle-wrapper.properties` pins `distributionSha256Sum`, so a mirror that serves
-a different distribution zip fails verification instead of being unpacked and
-run. Dependency repositories are the official ones only — a mirror belongs in a
-user-level init script.
-
-**Signing.** A release with no key configured comes out **unsigned**
-(`app-release-unsigned.apk`) with a warning. It used to fall back to the debug
-key, which produced an artifact that looked shippable and was signed by a
-per-machine key nobody chose: it could not install over the previous build.
-
-Point the build at a real key by writing `keystore/keystore.properties`
-(`storeFile`, `storePassword`, `keyAlias`, and optionally `keyPassword`) or by
-setting the matching `REASONIX_KEYSTORE*` environment variables. Key material
-stays out of the repository and out of build logs.
-
-Debug builds always use the debug key. Signing a debug build with the release key
-would let anyone who can read the keystore produce a build that installs over the
-shipped one under the same signature.
-
-**Non-ASCII checkout paths.** AGP refuses to build under a path with non-ASCII
-characters on Windows unless `android.overridePathCheck=true` is set. That is a
-property of one machine's disk layout, not of this project, so it is not in the
-repository's `gradle.properties`.
-
-Set it in a user-level `~/.gradle/gradle.properties` or pass
-`-Pandroid.overridePathCheck=true`.
@@ -1,263 +0,0 @@
-// The app module.
-//
-// Two things ship inside the APK and neither is fetched at runtime: the kernel
-// binary, and the built page it serves. Bundling the page is what makes the app
-// work offline and start instantly; bundling the kernel is what lets it run
-// `reasonix serve` on the phone itself, which is the whole point — the agent
-// starts locally and reaches the server over SSH, so the interface, the model
-// catalogue and the credentials all live on the phone.
-//
-// The kernel is placed as `jniLibs/arm64-v8a/libreasonix.so`, not under
-// `assets/`. That is not a trick for its own sake: Android permits executing a
-// file under `nativeLibraryDir` and nothing under `filesDir` has that
-// guarantee, so a binary shipped the ordinary way would have to be copied out
-// and made executable at runtime — which fails outright on a device mounted
-// noexec, and fails confusingly on the ones that are merely unusual. `scripts/
-// build-kernel.sh` writes it there; the name is fixed by the packaging rules.
-plugins {
-    id "com.android.application"
-    id "org.jetbrains.kotlin.android"
-}
-
-// One version, read by the About screen and the update check.
-//
-// A published build takes its version from the tag it was cut for, so the About
-// screen and the update check name that release rather than the number this file
-// was last edited to. Without the property — a local build — the numbers below
-// are the answer.
-def publishedVersion = findProperty("reasonixVersion") as String
-def versionMajor = 2
-def versionMinor = 0
-def versionPatch = 0
-def versionBuild = 1
-if (publishedVersion) {
-    def parsed = publishedVersion =~ /^(\d+)\.(\d+)\.(\d+)/
-    if (!parsed.find()) {
-        throw new GradleException("reasonixVersion must start with MAJOR.MINOR.PATCH, got: ${publishedVersion}")
-    }
-    versionMajor = parsed.group(1) as int
-    versionMinor = parsed.group(2) as int
-    versionPatch = parsed.group(3) as int
-}
-
-// Where the two build inputs are expected. Both are produced from the checkout
-// beside this module, by the scripts named in each message.
-def upstream = rootProject.file("..")
-def pageDir = new File(projectDir, "src/main/assets/page")
-def kernelSo = new File(projectDir, "src/main/jniLibs/arm64-v8a/libreasonix.so")
-
-android {
-    namespace "io.reasonix.mobile.app"
-    compileSdk 35
-
-    defaultConfig {
-        applicationId "io.reasonix.mobile.app"
-        minSdk 26
-        // 35: the API level this app is actually built and tested against.
-        //
-        // It was briefly lowered to 26 while chasing an install failure that
-        // turned out to be a package-name collision with an unrelated app on the
-        // user's phone — the signature error named the wrong cause. Nothing about
-        // the target level ever caused it, so it goes back to the level the code
-        // is really written for.
-        //
-        // One consequence to keep in mind: from 35, Android 15 forces the window
-        // edge-to-edge and ignores a request to lay out inside the safe area. The
-        // window therefore *is* drawn under the status bar and the home indicator,
-        // and the only correct answer is for the page to pad itself — which is
-        // what `MainActivity` feeds real inset values for, since a WebView does
-        // not implement `env(safe-area-inset-*)`.
-        targetSdk 35
-        // versionCode must strictly increase, because Android refuses an update
-        // whose versionCode is not higher. The build digit is a single decimal
-        // place, so the scheme collides once a patch reaches 10: 2.55.10 and
-        // 2.56.0 both come out as 25601. Patch releases above 9 are not expected
-        // on Studio, and widening the digit now would renumber every build, so
-        // the constraint is recorded rather than fixed.
-        versionCode versionMajor * 10000 + versionMinor * 100 + versionPatch * 10 + versionBuild
-        versionName "${versionMajor}.${versionMinor}.${versionPatch}"
-
-        // The kernel is built for arm64 only, so every other ABI would ship an
-        // APK that installs and then cannot start. Naming the one real ABI keeps
-        // the store listing and the installer honest.
-        ndk {
-            abiFilters "arm64-v8a"
-        }
-
-        resourceConfigurations += ["zh", "en"]
-    }
-
-    // Android identifies an app by package name *and* signing key: an APK signed
-    // with a different key is not an update, it is an impostor, and the
-    // installer refuses it with "the signature does not match the installed
-    // app". Gradle's default release key is the debug key in the user's home
-    // directory, which is per-machine, so a build made on a second computer
-    // could never be installed over one made on the first.
-    //
-    // Whoever ships a build therefore pins a key — and that key is theirs, not
-    // the project's. It lives beside the checkout, in `keystore/`, which is
-    // ignored; so does the file naming it. A checkout without either signs with
-    // the debug key instead, which is what a contributor building locally wants
-    // and is what the README tells them to expect. Point the build at a real key
-    // by writing `keystore/keystore.properties` (storeFile, storePassword,
-    // keyAlias) or by setting REASONIX_KEYSTORE / REASONIX_KEYSTORE_PASSWORD /
-    // REASONIX_KEY_ALIAS.
-    def keyProperties = new Properties()
-    def keyPropertiesFile = rootProject.file("keystore/keystore.properties")
-    if (keyPropertiesFile.exists()) {
-        keyPropertiesFile.withInputStream { keyProperties.load(it) }
-    }
-    def keyStorePath = keyProperties.getProperty("storeFile") ?: System.getenv("REASONIX_KEYSTORE")
-    def keyStorePassword = keyProperties.getProperty("storePassword") ?: System.getenv("REASONIX_KEYSTORE_PASSWORD")
-    // The key's own password is a separate fact from the store's. PKCS#12 keystores
-    // almost always use one password for both, so falling back to the store's
-    // password keeps those working — but assuming they are equal is how a keystore
-    // that separates them gets a build failure at signing time, after the work is
-    // done, with an error that does not say which of the two was wrong.
-    def keyPasswordValue = keyProperties.getProperty("keyPassword") ?: System.getenv("REASONIX_KEY_PASSWORD") ?: keyStorePassword
-    def keyAliasName = keyProperties.getProperty("keyAlias") ?: System.getenv("REASONIX_KEY_ALIAS")
-    def haveReleaseKey = keyStorePath && keyStorePassword && keyAliasName &&
-            rootProject.file(keyStorePath).exists()
-
-    signingConfigs {
-        if (haveReleaseKey) {
-            releaseKey {
-                storeFile rootProject.file(keyStorePath)
-                storePassword keyStorePassword
-                keyAlias keyAliasName
-                keyPassword keyPasswordValue
-                // v1 is not optional here. Android 7+ reads v2 and ignores v1, so
-                // a v2-only APK installs on a modern phone — and then fails on the
-                // compatibility layers older devices and HarmonyOS' 卓易通 run,
-                // which know only the JAR signature and report its absence as a
-                // signature *mismatch*. Signing both is what makes one APK install
-                // everywhere.
-                enableV1Signing = true
-                enableV2Signing = true
-                enableV3Signing = true
-            }
-        }
-    }
-
-    // Nothing native is compiled here — the one .so is a prebuilt the kernel
-    // script writes — so the external native build is left alone entirely.
-    buildTypes {
-        debug {
-            applicationIdSuffix ".debug"
-            versionNameSuffix "-debug"
-            // The debug key, always. Signing a debug build with the release key
-            // would mean anyone who can read the release keystore can produce a
-            // build that installs over the shipped one under the same signature,
-            // and it buys nothing: the debug package is suffixed `.debug`, so the
-            // two already install side by side.
-        }
-        release {
-            // The kernel is a stripped Go binary and the page is minified, but
-            // the Kotlin shell still benefits and shipping the shrunk build is
-            // how a size regression gets noticed.
-            minifyEnabled true
-            shrinkResources true
-            proguardFiles getDefaultProguardFile("proguard-android-optimize.txt"), "proguard-rules.pro"
-            // A release with no key of its own comes out *unsigned*
-            // (`app-release-unsigned.apk`). It used to fall back to the debug key
-            // so the artifact would install, which produced a release that looked
-            // shippable and was signed by a per-machine key nobody chose: it could
-            // not be installed over the previous build, and every machine made a
-            // different one. An unsigned artifact says what it is.
-            if (haveReleaseKey) {
-                signingConfig signingConfigs.releaseKey
-            } else {
-                logger.warn(
-                    "release build has no signing key: producing an unsigned APK. " +
-                        "Write keystore/keystore.properties (storeFile, storePassword, keyAlias) " +
-                        "or set REASONIX_KEYSTORE / REASONIX_KEYSTORE_PASSWORD / REASONIX_KEY_ALIAS."
-                )
-            }
-        }
-    }
-
-    compileOptions {
-        sourceCompatibility JavaVersion.VERSION_17
-        targetCompatibility JavaVersion.VERSION_17
-    }
-    kotlinOptions {
-        jvmTarget = "17"
-    }
-
-    buildFeatures {
-        // AGP 8 stops generating BuildConfig unless asked. `Kernel` reads the
-        // version name and code out of it to decide whether the bundled page
-        // needs unpacking, so it is asked for.
-        buildConfig = true
-    }
-
-    packaging {
-        jniLibs {
-            // `extractNativeLibs=true`: the kernel must exist as a real file on
-            // disk to be executed, and an uncompressed-and-mmap'd .so is not
-            // one. This is the only reason the flag is on.
-            useLegacyPackaging true
-        }
-        resources {
-            excludes += ["META-INF/*.version", "META-INF/*.kotlin_module", "kotlin/**", "DebugProbesKt.bin"]
-        }
-    }
-
-    // The page is read out of assets by `Kernel.publish`, not by the resource
-    // system, so it is not compressed — a page is already-compressed text and
-    // shrinking it again only costs startup time.
-    androidResources {
-        noCompress += ["json", "woff2", "ttf", "woff", "txt", "svg"]
-    }
-
-    lint {
-        abortOnError true
-        warningsAsErrors false
-        checkReleaseBuilds true
-    }
-}
-
-dependencies {
-    implementation "androidx.core:core-ktx:1.15.0"
-    implementation "androidx.appcompat:appcompat:1.7.0"
-    implementation "androidx.activity:activity-ktx:1.9.3"
-    implementation "androidx.webkit:webkit:1.12.1"
-    implementation "androidx.lifecycle:lifecycle-runtime-ktx:2.8.7"
-    implementation "org.jetbrains.kotlinx:kotlinx-coroutines-android:1.9.0"
-
-    // Scanning the pairing code a computer shows.
-    //
-    // CameraX for the preview and the frame stream, ZXing for the decode, and
-    // deliberately not ML Kit: this app installs on phones whose HarmonyOS
-    // compatibility layer has no Google Play Services, so a decoder that needs
-    // GMS would work on the developer's phone and fail on the user's. ZXing
-    // core is pure Java with no runtime dependency at all, which is what makes
-    // the scanner work on both.
-    implementation "androidx.camera:camera-core:1.4.1"
-    implementation "androidx.camera:camera-camera2:1.4.1"
-    implementation "androidx.camera:camera-lifecycle:1.4.1"
-    implementation "androidx.camera:camera-view:1.4.1"
-    implementation "com.google.zxing:core:3.5.3"
-}
-
-// Fail before the packaging step with something a person can act on, rather
-// than shipping an APK whose kernel is missing or whose page is a directory the
-// build never filled. The two messages name the script that produces each.
-tasks.register("requireBuildInputs") {
-    doFirst {
-        if (!kernelSo.exists()) {
-            throw new GradleException(
-                "内核二进制缺失：${kernelSo}\n先运行 ./scripts/build-kernel.sh（在 android/ 目录下）交叉编译内核。")
-        }
-        if (!pageDir.exists() || !new File(pageDir, "index.html").exists()) {
-            throw new GradleException(
-                "前端页面缺失：${pageDir}/index.html\n先运行 ./scripts/sync-page.sh（在 android/ 目录下）构建并同步页面。")
-        }
-    }
-}
-
-tasks.configureEach { task ->
-    if (task.name.startsWith("merge") && (task.name.contains("Assets") || task.name.contains("JniLibFolders"))) {
-        task.dependsOn("requireBuildInputs")
-    }
-}
@@ -1,32 +0,0 @@
-# R8 rules for the shell.
-#
-# There is very little to keep here, and that is the point: the app's own code
-# is reached from the manifest and from Android's own callbacks, both of which
-# R8 already traces, and nothing in it is loaded reflectively. What follows is
-# the two places where a name is looked up rather than called.
-
-# The WebView's JavaScript bridge, if one is ever added, is resolved by name on
-# the Java side and by the string the page uses. Keeping annotated members makes
-# that survive shrinking rather than failing at runtime with a method-not-found
-# that only shows up on a release build.
-keepclassmembers class * {
-    @android.webkit.JavascriptInterface <methods>;
-}
-
-# The Kernel's child process is started by absolute path, and its own log is
-# read as text; neither is reflective, so nothing else needs keeping.
-
-# Line numbers make a stack trace from a release build readable, which is the
-# only way a crash report from a phone can be acted on.
-keepattributes SourceFile,LineNumberTable
-renamesourcefileattribute SourceFile
-
-# Kotlin's `@Metadata` annotation is present on every class the compiler touched
-# and nothing in this app reflects over it. It is tempting to `-assumenosideeffects`
-# it away, and that was tried: any pattern broad enough to name its members also
-# matches the methods every class inherits from `Object` — `equals`, `hashCode`,
-# `toString` — and R8 then treats those as removable, which is a silent
-# behaviour change for a few kilobytes. Left as a warning suppression instead:
-# the annotation is dropped from the shrunk output anyway, because nothing reads
-# it.
-dontwarn kotlin.Metadata
@@ -1,112 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!--
-  The shell's manifest.
-
-  There is no INTERNET-adjacent permission beyond INTERNET itself: the page and
-  the kernel talk over loopback, and reaching a server is the kernel's job over
-  a socket it owns. Reading and writing files goes through the system's document
-  picker (SAF), which hands back a grant for one file at a time — so no
-  storage permission is declared and none is needed.
-
-  `usesCleartextTraffic` is off here, and cleartext is re-enabled in the network
-  security config only because the kernel serves its page on `127.0.0.1` and a
-  paired computer serves the same page over its LAN address — both plain HTTP.
-  The config says what that costs and what bounds it.
-->
-<manifest xmlns:android="http://schemas.android.com/apk/res/android"
-    xmlns:tools="http://schemas.android.com/tools">
-
-    <uses-permission android:name="android.permission.INTERNET" />
-    <uses-permission android:name="android.permission.ACCESS_NETWORK_STATE" />
-    <!-- Only to tell the person the kernel is still running while the screen is
-         off; it is a foreground service with a dismissible notification, not a
-         background wake-up. -->
-    <uses-permission android:name="android.permission.FOREGROUND_SERVICE" />
-    <!-- From API 34 every foreground-service *type* needs its own permission as
-         well as the general one: naming dataSync in the manifest without asking
-         for it is a SecurityException at startForeground, on the one code path
-         that keeps a long task alive. -->
-    <uses-permission android:name="android.permission.FOREGROUND_SERVICE_DATA_SYNC" />
-    <uses-permission android:name="android.permission.POST_NOTIFICATIONS" />
-    <!-- Only to read the pairing code a computer shows. Requested at the
-         moment the person asks to scan, never at launch. -->
-    <uses-permission android:name="android.permission.CAMERA" />
-    <!-- The kernel is a child process, and a task left running is one the person
-         expects to finish with the screen off. Android's own doze can freeze the
-         whole app group instead, which is what the shell asks to be exempt from
-         — once, when a turn is actually running, never at launch. It is a
-         request the person can refuse, and refusing it only costs battery life
-         during a long turn. -->
-    <uses-permission android:name="android.permission.REQUEST_IGNORE_BATTERY_OPTIMIZATIONS" />
-    <!-- Every-file access, for opening a folder on this phone as a workspace.
-         The kernel is a plain Go binary: it opens paths with `open(2)`, and a
-         document URI is not a path it can open. SAF would hand back exactly
-         that unusable URI, so the one honest permission is the broad one —
-         granted by the person in system settings, which is the only place
-         Android offers it. Nothing here reads storage for any other purpose. -->
-    <uses-permission
-        android:name="android.permission.MANAGE_EXTERNAL_STORAGE"
-        tools:ignore="ScopedStorage" />
-    <!-- Below API 33 the same folder walk needs the older pair; from 33 they are
-         superseded by the permission above and asking for them would be a lie. -->
-    <uses-permission android:name="android.permission.READ_EXTERNAL_STORAGE" android:maxSdkVersion="32" />
-    <uses-permission android:name="android.permission.WRITE_EXTERNAL_STORAGE" android:maxSdkVersion="29" />
-
-    <!-- A phone without a touchscreen cannot drive this interface; saying so
-         keeps the listing honest rather than letting it install on a TV. -->
-    <uses-feature android:name="android.hardware.touchscreen" android:required="true" />
-    <!-- The camera is optional: a phone without one is still a fine client for
-         a kernel it already knows, and it can still pair by opening the link the
-         computer shows. Requiring it would take the app off those devices. -->
-    <uses-feature android:name="android.hardware.camera" android:required="false" />
-    <uses-feature android:name="android.hardware.camera.autofocus" android:required="false" />
-
-    <application
-        android:name=".App"
-        android:allowBackup="false"
-        android:dataExtractionRules="@xml/data_extraction_rules"
-        android:fullBackupContent="false"
-        android:icon="@mipmap/ic_launcher"
-        android:label="@string/app_name"
-        android:networkSecurityConfig="@xml/network_security_config"
-        android:supportsRtl="true"
-        android:theme="@style/Theme.Reasonix"
-        android:usesCleartextTraffic="false"
-        tools:targetApi="35">
-
-        <activity
-            android:name=".MainActivity"
-            android:exported="true"
-            android:configChanges="orientation|screenSize|screenLayout|keyboardHidden|uiMode|density|fontScale|smallestScreenSize"
-            android:launchMode="singleTask"
-            android:windowSoftInputMode="adjustResize"
-            android:theme="@style/Theme.Reasonix">
-            <intent-filter>
-                <action android:name="android.intent.action.MAIN" />
-                <category android:name="android.intent.category.LAUNCHER" />
-            </intent-filter>
-        </activity>
-
-        <!-- The kernel outlives the screen: a turn that takes minutes would
-             otherwise be killed when the phone is pocketed. The service is
-             `stopWithTask="false"` so swiping the app away ends it deliberately
-             rather than silently, and `foregroundServiceType="dataSync"` is the
-             closest honest description of "an agent is working". -->
-        <service
-            android:name=".KernelService"
-            android:exported="false"
-            android:foregroundServiceType="dataSync"
-            android:stopWithTask="false" />
-
-        <!-- Reading the pairing code a computer shows. Its own activity rather
-             than a dialog, because the camera preview needs the whole screen to
-             be worth pointing at anything. Not exported: only this app starts
-             it, and a scanner another app could open is one that could be made
-             to read a code on someone else's behalf. -->
-        <activity
-            android:name=".ScanActivity"
-            android:exported="false"
-            android:screenOrientation="portrait"
-            android:theme="@style/Theme.Reasonix.Scan" />
-    </application>
-</manifest>
@@ -1,59 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.app.Application
-import android.app.NotificationChannel
-import android.app.NotificationManager
-import android.content.Context
-
-/**
- * The process-wide facts this app needs.
- *
- * The kernel is one child process, so there is exactly one of it no matter how
- * many activities exist: it is held here, lazily, and the activity and the
- * service both reach the same instance. Nothing else is stored — the theme, the
- * language and every preference belong to the page and live in its localStorage.
- */
-class App : Application() {
-
-    /**
-     * The kernel, created on first use.
-     *
-     * Lazy rather than eager so that merely launching the process to answer a
-     * notification does not unpack a 15MB page.
-     */
-    val kernel: Kernel by lazy { Kernel(this) }
-
-    override fun onCreate() {
-        super.onCreate()
-        createChannel()
-    }
-
-    /**
-     * A channel the person can mute without muting the app, and one they can
-     * find: an agent that runs for minutes on a phone has to say so somewhere
-     * the system will not kill, and a notification is the only place Android
-     * offers. `IMPORTANCE_MIN` because the notification is a receipt, not an
-     * interruption — the work it reports is on another machine.
-     */
-    private fun createChannel() {
-        // No SDK_INT guard: minSdk is 26, which is the release NotificationChannel
-        // arrived in, so every device this app installs on has it.
-        val manager = getSystemService(Context.NOTIFICATION_SERVICE) as NotificationManager
-        if (manager.getNotificationChannel(CHANNEL_KERNEL) != null) return
-        val channel = NotificationChannel(
-            CHANNEL_KERNEL,
-            getString(R.string.channel_kernel),
-            NotificationManager.IMPORTANCE_MIN,
-        ).apply {
-            description = getString(R.string.channel_kernel_desc)
-            setShowBadge(false)
-            enableVibration(false)
-            setSound(null, null)
-        }
-        manager.createNotificationChannel(channel)
-    }
-
-    companion object {
-        const val CHANNEL_KERNEL = "reasonix-kernel"
-    }
-}
@@ -1,699 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.graphics.Bitmap
-import android.graphics.Canvas
-import android.os.Handler
-import android.os.Looper
-import android.util.Base64
-import android.util.Log
-import android.webkit.WebView
-import org.json.JSONArray
-import org.json.JSONObject
-import java.io.BufferedInputStream
-import java.io.ByteArrayOutputStream
-import java.io.OutputStream
-import java.net.ServerSocket
-import java.net.Socket
-import java.net.SocketException
-import java.util.concurrent.CountDownLatch
-import java.util.concurrent.TimeUnit
-import java.util.concurrent.atomic.AtomicBoolean
-import java.util.concurrent.atomic.AtomicReference
-
-/**
- * The kernel's browser, as a loopback HTTP server.
- *
- * The Go kernel (`internal/platform/browser/shell.go`) drives a WebView by
- * POSTing JSON to a handful of paths on 127.0.0.1. This class is the other end
- * of that conversation: it accepts connections on a random port, parses the
- * minimal HTTP the kernel sends, and translates each path into an operation on
- * the overlay WebView.
- *
- * The server is deliberately tiny — no connection pooling, no keep-alive, no
- * chunked encoding — because the kernel is its only client and sends one
- * request per connection. A `ServerSocket` accept loop on a daemon thread is
- * the whole transport; each request is handled synchronously on its own thread.
- *
- * WebView operations must run on the UI thread. Each handler posts to the main
- * looper and waits on a latch, with a timeout that bounds the worst case (a
- * page that hangs the JS thread) to something the kernel's own 30-second client
- * timeout will not outlast.
- *
- * Memory note on `/screenshot`: a full-page `WebView.draw` into an ARGB_8888
- * bitmap at device resolution can reach tens of MB for a long page. The bitmap
- * is downsampled by a factor of 2 in each dimension before PNG encoding, which
- * keeps the peak allocation under ~10 MB for a typical 1080p viewport. The
- * original is recycled immediately after the downsample.
- */
-/** One thing a page asked the network for, as the WebView's request
- *  interceptor reports it. The interceptor is handed the method and the URL
- *  and nothing about the response, so a caller may show exactly this much and
- *  must not read it as a completed exchange. */
-data class BrowserRequest(val method: String, val url: String, val kind: String)
-
-class BrowserShellServer(
-    /** The overlay WebView, when one already exists. The routes that act on an
-     *  open page — back, forward, reload, close — use this: with no page there
-     *  is nothing to act on, and creating one to do nothing to it would start a
-     *  renderer nobody asked for. */
-    private val existingView: () -> WebView?,
-    /** The overlay WebView, created on first use. The routes that drive a page
-     *  use this, because the kernel can open one before the panel has ever been
-     *  shown: the page's own panel asks for the view when it is displayed, and
-     *  an agent's `browser_open` can arrive first. Requiring the view to exist
-     *  already answered every such call with "browser view not ready", which
-     *  reads as a broken engine rather than as a view that was not created yet.
-     *  Called on the UI thread. */
-    private val ensureView: () -> WebView,
-    /** Why the page in the view is the browser's own error page rather than a
-     *  site, as `(url, chromium error)`; null when the last load succeeded.
-     *  `/read` asks this, because extracting `innerText` from the error page
-     *  returns the browser's own "this page could not be loaded" notice, which
-     *  the kernel would hand to the model as though the site had said it. */
-    private val loadFailure: () -> Pair<String, String>? = { null },
-    /** Begin a navigation: forget the load failure recorded for the page being
-     *  left, drop its request list, and start the wait a following `/read`
-     *  joins. Called on the UI thread, before the new address is loaded, so the
-     *  wait is in place before any callback for the new page can fire. */
-    private val startNavigation: () -> Unit = {},
-    /** What the page currently in the view has asked the network for, oldest
-     *  first. The WebView reports a request's method and URL to its interceptor
-     *  but nothing about the response, so that is all a caller may show. */
-    private val requests: () -> List<BrowserRequest> = { emptyList() },
-    /** Wait for the navigation last asked for to settle — the page arrived, or
-     *  its load failed — up to the given milliseconds. Called on the request's
-     *  own thread, never the UI thread: the settle is signalled from the UI
-     *  thread, so waiting on it there would deadlock. `/load` starts a
-     *  navigation and returns before the page has been visited; without this a
-     *  `/read` right behind it still sees the URL of the page before, and a
-     *  failure recorded for that page was handed back as the new one's answer. */
-    private val awaitNavigation: (Long) -> Unit = {},
-) {
-
-    companion object {
-        private const val TAG = "BrowserShellServer"
-
-        /** How long a UI-thread operation may take before the handler gives up.
-         *  The kernel's client timeout is 30 s; staying under it means the kernel
-         *  sees a clean error rather than a dropped connection. */
-        private const val UI_TIMEOUT_MS = 25_000L
-
-        /** The maximum text `/read` returns. Matches the kernel's `maxReadText`
-         *  (400 << 10). The shell truncates here and says so, rather than letting
-         *  the kernel truncate silently. */
-        private const val MAX_READ_TEXT = 400 * 1024
-
-        /** Screenshot downsample factor. 2 means half width and half height. */
-        private const val SCREENSHOT_SCALE = 2
-
-        /** The address the kernel dials. Both ends must agree on the family:
-         *  see [start] for what goes wrong when they do not. */
-        private const val LOOPBACK_V4 = "127.0.0.1"
-
-        /** How long `/read` waits for a navigation it follows to settle before
-         *  reading whatever the view is on. A page that never finishes loading
-         *  must not hold the tool call open to the kernel's own timeout. */
-        private const val NAV_SETTLE_MS = 10_000L
-
-        /** How long `/read` then waits for the page's own script to answer.
-         *  `NAV_SETTLE_MS + READ_SCRIPT_MS` has to stay under [UI_TIMEOUT_MS],
-         *  and that under the kernel's 30-second client timeout: a read that
-         *  waited out both budgets would outlast the caller, which would see a
-         *  dropped connection instead of this shell's answer. 10 + 15 = 25 s. */
-        private const val READ_SCRIPT_MS = 15_000L
-    }
-
-    @Volatile
-    private var serverSocket: ServerSocket? = null
-
-    /** The port the server is listening on. Valid after [start] returns. */
-    @Volatile
-    var port: Int = 0
-        private set
-
-    /**
-     * The bearer credential every request must carry. Minted by the activity
-     * and handed to the kernel through the child's environment, so the server
-     * answers only the process this shell started.
-     *
-     * A phone's loopback is not a boundary: any other app may connect to
-     * 127.0.0.1, and the routes here run script in and read text out of
-     * whatever page is loaded. Binding to `getLoopbackAddress()` keeps other
-     * machines out but not other apps, so the credential is what makes "only
-     * the kernel this shell started" true.
-     */
-    @Volatile
-    var token: String = ""
-        private set
-
-    private val mainHandler = Handler(Looper.getMainLooper())
-
-    /**
-     * Start listening on 127.0.0.1 with an ephemeral port. The port is available
-     * via [port] immediately after this returns, and the credential the kernel
-     * must present via [token].
-     */
-    fun start() {
-        token = newToken()
-        // The loopback address is spelled out as IPv4, not taken from
-        // `InetAddress.getLoopbackAddress()`. That call is allowed to answer
-        // either family and Android's libcore answers IPv6 (`::1`), while the
-        // kernel's client is hardcoded to `http://127.0.0.1:<port>`
-        // (`internal/assembly/boot/browser.go`). A socket bound to `::1` does
-        // not answer a dial to `127.0.0.1`, so every browser call failed with
-        // "connection refused" while the port was bound and reported — the
-        // failure looked like a server that never started. Both ends have to
-        // name the same family, and IPv4 is the one the kernel names.
-        val ss = ServerSocket(0, 16, java.net.InetAddress.getByName(LOOPBACK_V4))
-        serverSocket = ss
-        port = ss.localPort
-        Log.i(TAG, "browser shell server listening on $LOOPBACK_V4:$port")
-
-        Thread({
-            while (!ss.isClosed) {
-                try {
-                    val client = ss.accept()
-                    Thread({ handleClient(client, token) }, "browser-shell-conn").apply { isDaemon = true }.start()
-                } catch (_: SocketException) {
-                    // Server was closed; exit the loop.
-                }
-            }
-        }, "browser-shell-accept").apply { isDaemon = true }.start()
-    }
-
-    fun stop() {
-        runCatching { serverSocket?.close() }
-        serverSocket = null
-    }
-
-    /** A fresh 256-bit credential, hex-encoded, for one launch of the server. */
-    private fun newToken(): String {
-        val bytes = ByteArray(32)
-        java.security.SecureRandom().nextBytes(bytes)
-        return bytes.joinToString("") { "%02x".format(it) }
-    }
-
-    // ---- HTTP plumbing --------------------------------------------------------
-
-    private fun handleClient(client: Socket, expected: String) {
-        try {
-            client.soTimeout = 30_000
-            val input = BufferedInputStream(client.getInputStream())
-            val output = client.getOutputStream()
-
-            // Parse the request line.
-            val requestLine = readLine(input) ?: return
-            val parts = requestLine.split(" ", limit = 3)
-            if (parts.size < 2) return
-            val method = parts[0]
-            val path = parts[1].substringBefore('?')
-
-            // Read headers.
-            var contentLength = 0
-            var bearer: String? = null
-            while (true) {
-                val header = readLine(input) ?: break
-                if (header.isEmpty()) break
-                val lower = header.lowercase()
-                if (lower.startsWith("content-length:")) {
-                    contentLength = header.substringAfter(':').trim().toIntOrNull() ?: 0
-                } else if (lower.startsWith("authorization:")) {
-                    bearer = header.substringAfter(':').trim().removePrefix("Bearer ").trim()
-                }
-            }
-
-            // Nothing behind this server is safe to hand out on the strength of
-            // an address: the port is on loopback, which any other app on the
-            // phone can dial, and the routes run script in and read text out of
-            // the loaded page. The credential the kernel was started with is the
-            // only thing that says a caller is that kernel. Every route is
-            // refused without it, /health included — a probe needs no exemption.
-            if (expected.isEmpty() || bearer != expected) {
-                writeResponse(output, 401, """{"error":"unauthorized"}""")
-                return
-            }
-
-            // Read body.
-            val body = if (contentLength > 0) {
-                val buf = ByteArray(contentLength)
-                var read = 0
-                while (read < contentLength) {
-                    val n = input.read(buf, read, contentLength - read)
-                    if (n == -1) break
-                    read += n
-                }
-                String(buf, 0, read, Charsets.UTF_8)
-            } else ""
-
-            // Route.
-            val (status, responseBody) = route(method, path, body)
-            writeResponse(output, status, responseBody)
-        } catch (e: Exception) {
-            Log.w(TAG, "connection error: ${e.message}")
-        } finally {
-            runCatching { client.close() }
-        }
-    }
-
-    private fun route(method: String, path: String, body: String): Pair<Int, String> {
-        return when {
-            method == "GET" && path == "/health" -> 200 to """{"ok":true}"""
-            method == "POST" && path == "/load" -> handleLoad(body)
-            method == "POST" && path == "/close" -> handleClose()
-            method == "POST" && path == "/read" -> handleRead()
-            method == "POST" && path == "/screenshot" -> handleScreenshot()
-            method == "POST" && path == "/click" -> handleClick(body)
-            method == "POST" && path == "/eval" -> handleEval(body)
-            method == "POST" && path == "/back" -> handleBack()
-            method == "POST" && path == "/forward" -> handleForward()
-            method == "POST" && path == "/reload" -> handleReload()
-            method == "POST" && path == "/requests" -> handleRequests()
-            else -> 404 to """{"error":"not found"}"""
-        }
-    }
-
-    private fun writeResponse(out: OutputStream, status: Int, body: String) {
-        val reason = when (status) {
-            200 -> "OK"
-            404 -> "Not Found"
-            500 -> "Internal Server Error"
-            else -> "Error"
-        }
-        val bytes = body.toByteArray(Charsets.UTF_8)
-        val header = "HTTP/1.1 $status $reason\r\n" +
-            "Content-Type: application/json\r\n" +
-            "Content-Length: ${bytes.size}\r\n" +
-            "Connection: close\r\n" +
-            "\r\n"
-        out.write(header.toByteArray(Charsets.UTF_8))
-        out.write(bytes)
-        out.flush()
-    }
-
-    private fun readLine(input: BufferedInputStream): String? {
-        val sb = StringBuilder()
-        while (true) {
-            val b = input.read()
-            if (b == -1) return if (sb.isEmpty()) null else sb.toString()
-            if (b == '\r'.code) {
-                val next = input.read()
-                if (next == '\n'.code) return sb.toString()
-                sb.append(b.toChar())
-                if (next != -1) sb.append(next.toChar())
-                continue
-            }
-            if (b == '\n'.code) return sb.toString()
-            sb.append(b.toChar())
-        }
-    }
-
-    // ---- endpoint handlers ----------------------------------------------------
-
-    /** POST /load {"url":"..."} → {"url":"...","title":"..."} */
-    private fun handleLoad(body: String): Pair<Int, String> {
-        val json = runCatching { JSONObject(body) }.getOrNull()
-            ?: return 500 to """{"error":"invalid request body"}"""
-        val url = json.optString("url", "")
-        if (url.isEmpty()) return 500 to """{"error":"missing url"}"""
-
-        val result = AtomicReference("")
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            val wv = ensureView()
-            // The failure recorded for the page being left stops being this
-            // tab's, and the wait a following `/read` joins is armed before the
-            // new address is loaded, so no callback for the new page can fire
-            // ahead of it. `/load` answers before the page arrives, which is
-            // what left the old failure to be read back as the new page's.
-            startNavigation()
-            wv.loadUrl(url)
-            // The title and final URL are not available synchronously after
-            // loadUrl; answer with what was requested. The kernel stores this
-            // as the tab's URL and reads the real one on the next /read.
-            result.set(JSONObject().apply {
-                put("url", url)
-                put("title", "")
-            }.toString())
-            latch.countDown()
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout waiting for UI thread"}"""
-        }
-        return 200 to result.get()
-    }
-
-    /** POST /close {} → 200 */
-    private fun handleClose(): Pair<Int, String> {
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            val wv = existingView()
-            if (wv != null) {
-                wv.loadUrl("about:blank")
-                wv.clearHistory()
-            }
-            latch.countDown()
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout"}"""
-        }
-        return 200 to "{}"
-    }
-
-    /**
-     * POST /read {} → {"title":"...","url":"...","text":"..."}
-     *
-     * Injects a script that extracts readable text from the page, preferring
-     * `<article>` or `<main>` and falling back to `body.innerText`. The result
-     * is truncated to 400 KB (matching the kernel's `maxReadText`); when
-     * truncated, a note is appended so the model knows the text was cut.
-     */
-    private fun handleRead(): Pair<Int, String> {
-        // The page the caller just asked for has not necessarily been visited
-        // yet: `/load` returns as soon as it has started the navigation. Waiting
-        // here, off the UI thread, is what makes this read answer about that
-        // page instead of the one before it.
-        awaitNavigation(NAV_SETTLE_MS)
-        val result = AtomicReference("")
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            val wv = ensureView()
-            // A page that did not load is not content. Whether the view is on
-            // the browser's own error page is read from the view itself rather
-            // than from a remembered failure: a failure recorded for the page
-            // before a new navigation would otherwise be handed back as this
-            // page's answer, long after the page the caller asked for had
-            // loaded. The remembered failure is still what names the address
-            // that failed and why, since the error page's own URL does not.
-            if ((wv.url ?: "").startsWith("chrome-error:")) {
-                val (url, reason) = loadFailure() ?: ((wv.url ?: "") to "ERR_FAILED")
-                result.set(JSONObject().apply {
-                    put("url", url)
-                    put("title", "")
-                    put("reason", reason)
-                    put("text", "")
-                }.toString())
-                latch.countDown()
-                return@post
-            }
-            // The extraction script: prefer article/main for content pages,
-            // fall back to body. innerText preserves visual line breaks. It
-            // also stamps a ref on every element a person could act on, so the
-            // kernel can name one in an act step the way the desktop backend
-            // names an accessibility node. A ref already on an element is kept,
-            // so it stays the same across snapshots of one document; a fresh
-            // document starts the numbering over, and the kernel only accepts a
-            // ref from the snapshot it last read.
-            val script = """
-                (function(){
-                    var seq = window.__rxRefSeq || 0;
-                    var used = {};
-                    var marked = document.querySelectorAll('[data-rx-ref]');
-                    for (var i = 0; i < marked.length; i++) { used[marked[i].getAttribute('data-rx-ref')] = true; }
-                    function roleOf(el) {
-                        var tag = el.tagName.toLowerCase();
-                        var role = (el.getAttribute('role') || '').toLowerCase();
-                        if (role) return role;
-                        if (tag === 'a') return 'link';
-                        if (tag === 'button') return 'button';
-                        if (tag === 'select') return 'combobox';
-                        if (tag === 'textarea') return 'textbox';
-                        if (tag === 'input') {
-                            var t = (el.getAttribute('type') || 'text').toLowerCase();
-                            if (t === 'checkbox' || t === 'radio' || t === 'submit' || t === 'button' || t === 'reset') return t;
-                            return 'textbox';
-                        }
-                        return 'generic';
-                    }
-                    function nameOf(el) {
-                        var n = el.getAttribute('aria-label') || el.getAttribute('placeholder') || el.getAttribute('name') || '';
-                        if (!n) { n = (el.innerText || el.textContent || '').trim(); }
-                        return n.replace(/\s+/g, ' ').slice(0, 160);
-                    }
-                    var sel = 'a[href],button,input,select,textarea,[role=button],[role=link],[role=checkbox],[role=radio],[role=tab],[role=menuitem],[contenteditable=true]';
-                    var els = document.querySelectorAll(sel);
-                    var elements = [];
-                    for (var j = 0; j < els.length; j++) {
-                        var el = els[j];
-                        var ref = el.getAttribute('data-rx-ref');
-                        if (!ref) {
-                            do { seq++; ref = 'e' + seq; } while (used[ref]);
-                            el.setAttribute('data-rx-ref', ref);
-                            used[ref] = true;
-                        }
-                        var raw = (el.value != null && String(el.value) !== '') ? String(el.value).slice(0, 160) : '';
-                        elements.push({ref: ref, role: roleOf(el), name: nameOf(el), value: raw});
-                    }
-                    window.__rxRefSeq = seq;
-                    var el2 = document.querySelector('article') || document.querySelector('main') || document.body;
-                    var text = el2 ? (el2.innerText || '') : '';
-                    var title = document.title || '';
-                    var url = location.href;
-                    return JSON.stringify({title: title, url: url, text: text, elements: elements});
-                })()
-            """.trimIndent()
-            wv.evaluateJavascript(script) { value ->
-                // evaluateJavascript returns a JSON-encoded string (the result
-                // is double-encoded). Strip the outer quotes.
-                val parsed = parseJsResult(value)
-                if (parsed != null) {
-                    try {
-                        val obj = JSONObject(parsed)
-                        var text = obj.optString("text", "")
-                        if (text.length > MAX_READ_TEXT) {
-                            text = text.substring(0, MAX_READ_TEXT) +
-                                "\n\n[text truncated at 400 KB]"
-                        }
-                        result.set(JSONObject().apply {
-                            put("title", obj.optString("title", ""))
-                            put("url", obj.optString("url", ""))
-                            put("text", text)
-                            put("elements", obj.optJSONArray("elements") ?: JSONArray())
-                        }.toString())
-                    } catch (e: Exception) {
-                        result.set(JSONObject().apply {
-                            put("title", "")
-                            put("url", wv.url ?: "")
-                            put("text", parsed.take(MAX_READ_TEXT))
-                        }.toString())
-                    }
-                } else {
-                    result.set(JSONObject().apply {
-                        put("title", "")
-                        put("url", wv.url ?: "")
-                        put("text", "")
-                    }.toString())
-                }
-                latch.countDown()
-            }
-        }
-        if (!latch.await(READ_SCRIPT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout waiting for page script"}"""
-        }
-        return 200 to result.get()
-    }
-
-    /**
-     * POST /screenshot {} → {"data":"<base64>","mime":"image/png"}
-     *
-     * Draws the WebView into a bitmap, downsamples by SCREENSHOT_SCALE to keep
-     * memory bounded, and encodes as PNG → base64. For a 1080×2400 viewport the
-     * full bitmap is ~10 MB; the downsampled one is ~2.5 MB. The full bitmap is
-     * recycled immediately after scaling.
-     */
-    private fun handleScreenshot(): Pair<Int, String> {
-        val result = AtomicReference("")
-        val failed = AtomicBoolean(false)
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            val wv = ensureView()
-            try {
-                val width = wv.width.takeIf { it > 0 } ?: 1080
-                val height = wv.height.takeIf { it > 0 } ?: 1920
-                val bmp = Bitmap.createBitmap(width, height, Bitmap.Config.ARGB_8888)
-                val canvas = Canvas(bmp)
-                wv.draw(canvas)
-
-                // Downsample to bound the PNG size and the base64 string.
-                val scaled = Bitmap.createScaledBitmap(
-                    bmp,
-                    width / SCREENSHOT_SCALE,
-                    height / SCREENSHOT_SCALE,
-                    true,
-                )
-                if (scaled !== bmp) bmp.recycle()
-
-                val baos = ByteArrayOutputStream()
-                scaled.compress(Bitmap.CompressFormat.PNG, 85, baos)
-                scaled.recycle()
-                val b64 = Base64.encodeToString(baos.toByteArray(), Base64.NO_WRAP)
-                result.set(JSONObject().apply {
-                    put("data", b64)
-                    put("mime", "image/png")
-                }.toString())
-            } catch (e: Exception) {
-                // A 200 with an error body would be read by the kernel as a
-                // screenshot whose data is empty, and it would build a `data:`
-                // URI out of nothing. The status is what tells it this failed.
-                result.set("""{"error":"screenshot failed: ${e.message}"}""")
-                failed.set(true)
-            }
-            latch.countDown()
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout"}"""
-        }
-        if (failed.get()) return 500 to result.get()
-        return 200 to result.get()
-    }
-
-    /** POST /click {"selector":"..."} → {"ok":true|false} */
-    private fun handleClick(body: String): Pair<Int, String> {
-        val json = runCatching { JSONObject(body) }.getOrNull()
-            ?: return 500 to """{"error":"invalid request body"}"""
-        val selector = json.optString("selector", "")
-        if (selector.isEmpty()) return 500 to """{"error":"missing selector"}"""
-
-        val result = AtomicReference("")
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            val wv = ensureView()
-            // Escape the selector for embedding in a JS string.
-            val escaped = selector.replace("\\", "\\\\").replace("'", "\\'")
-            val script = """
-                (function(){
-                    try {
-                        var el = document.querySelector('$escaped');
-                        if (!el) return JSON.stringify({ok: false});
-                        el.click();
-                        return JSON.stringify({ok: true});
-                    } catch(e) {
-                        return JSON.stringify({ok: false});
-                    }
-                })()
-            """.trimIndent()
-            wv.evaluateJavascript(script) { value ->
-                val parsed = parseJsResult(value)
-                result.set(parsed ?: """{"ok":false}""")
-                latch.countDown()
-            }
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout"}"""
-        }
-        return 200 to result.get()
-    }
-
-    /** POST /eval {"script":"..."} → {"result":"..."} */
-    private fun handleEval(body: String): Pair<Int, String> {
-        val json = runCatching { JSONObject(body) }.getOrNull()
-            ?: return 500 to """{"error":"invalid request body"}"""
-        val script = json.optString("script", "")
-        if (script.isEmpty()) return 500 to """{"error":"missing script"}"""
-
-        val result = AtomicReference("")
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            val wv = ensureView()
-            wv.evaluateJavascript(script) { value ->
-                // The kernel expects a plain string in "result". The value from
-                // evaluateJavascript is JSON-encoded; unwrap it.
-                val unwrapped = parseJsResult(value) ?: ""
-                result.set(JSONObject().apply {
-                    put("result", unwrapped)
-                }.toString())
-                latch.countDown()
-            }
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout waiting for eval"}"""
-        }
-        return 200 to result.get()
-    }
-
-    /** POST /back → 200 */
-    private fun handleBack(): Pair<Int, String> {
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            existingView()?.let { if (it.canGoBack()) it.goBack() }
-            latch.countDown()
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout"}"""
-        }
-        return 200 to "{}"
-    }
-
-    /** POST /forward → 200 */
-    private fun handleForward(): Pair<Int, String> {
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            existingView()?.let { if (it.canGoForward()) it.goForward() }
-            latch.countDown()
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout"}"""
-        }
-        return 200 to "{}"
-    }
-
-    /** POST /reload → 200 */
-    private fun handleReload(): Pair<Int, String> {
-        val latch = CountDownLatch(1)
-        mainHandler.post {
-            val wv = existingView()
-            // A view sitting on the browser's own error page is not on the site,
-            // and reloading it would retry the error page itself. The address
-            // that failed is loaded again instead — which is the page a caller
-            // means by "reload" — read before `startNavigation` clears it.
-            val failed = if (wv != null && (wv.url ?: "").startsWith("chrome-error:"))
-                loadFailure()?.first else null
-            startNavigation()
-            if (wv != null) {
-                if (failed != null) wv.loadUrl(failed) else wv.reload()
-            }
-            latch.countDown()
-        }
-        if (!latch.await(UI_TIMEOUT_MS, TimeUnit.MILLISECONDS)) {
-            return 500 to """{"error":"timeout"}"""
-        }
-        return 200 to "{}"
-    }
-
-    /** POST /requests {} → {"requests":[{"method","url","kind"},…]} */
-    private fun handleRequests(): Pair<Int, String> {
-        val arr = JSONArray()
-        for (r in requests()) {
-            arr.put(JSONObject().apply {
-                put("method", r.method)
-                put("url", r.url)
-                put("kind", r.kind)
-            })
-        }
-        return 200 to JSONObject().apply { put("requests", arr) }.toString()
-    }
-
-    // ---- helpers ----------------------------------------------------------------
-
-    /**
-     * Unwrap the result of `evaluateJavascript`. The callback receives a
-     * JSON-encoded string: a JS string "hello" arrives as `"\"hello\""`, a
-     * number 42 as `"42"`, null as `"null"`. Strip the outer JSON encoding to
-     * get the raw value.
-     */
-    private fun parseJsResult(value: String?): String? {
-        if (value == null || value == "null") return null
-        // The value is a JSON literal. If it is a JSON string, decode it.
-        return try {
-            if (value.startsWith("\"") && value.endsWith("\"")) {
-                JSONObject().apply { }.let {
-                    // Use JSONObject's string decoding: wrap in an object.
-                    JSONObject("""{"v":$value}""").optString("v", "")
-                }
-            } else {
-                value
-            }
-        } catch (_: Exception) {
-            value
-        }
-    }
-}
@@ -1,119 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.content.Context
-import android.net.ConnectivityManager
-import android.net.LinkProperties
-import android.os.Build
-
-/**
- * The phone's DNS servers, as the kernel needs them.
- *
- * The kernel is a CGO-free Go binary, and Go's own source explains why that
- * matters:
- *
- * ```
- * // DNS requests don't work on Android, so prefer the cgo resolver.
- * // Issue #10714.
- * case "android":
- * ```
- *
- * A normal Android program resolves names through `getaddrinfo`, which reaches
- * the `netd` daemon and gets the servers of the network that is actually up.
- * A static binary has no cgo resolver, so Go reads `/etc/resolv.conf` instead —
- * and Android has no such file. Go then falls back to `127.0.0.1:53`, where
- * nothing listens, and every lookup fails: the setup screen reports
- * "连接失败" for every provider, including the official ones, on a phone whose
- * browser reaches them fine.
- *
- * So they are read here and handed to the kernel through `REASONIX_DNS`. This is
- * the same source the system itself uses, so a VPN, a private DNS hostname, or a
- * network that changed since boot is all reflected without this code knowing
- * anything about them.
- */
-object Dns {
-
-    /**
-     * The active network's servers, comma separated, or "" when there is none.
-     *
-     * Read from the *active* network rather than from every link, because a phone
-     * commonly has several at once — Wi-Fi, mobile data, a VPN — and the servers
-     * of a network that is not carrying traffic are the wrong ones to ask.
-     */
-    fun servers(context: Context): String {
-        val manager = context.getSystemService(Context.CONNECTIVITY_SERVICE) as? ConnectivityManager
-            ?: return ""
-        val properties: LinkProperties = runCatching { manager.activeNetwork }.getOrNull()
-            ?.let { network -> runCatching { manager.getLinkProperties(network) }.getOrNull() }
-            ?: return ""
-        return properties.dnsServers
-            .mapNotNull { address -> address.hostAddress }
-            .filter { it.isNotBlank() }
-            .filter { routable(it) }
-            .sortedBy { preference(it) }
-            .joinToString(",")
-    }
-
-    /**
-     * Whether an address can actually carry a query.
-     *
-     * A link-local address (fe80::/10) without its zone cannot: the packet has
-     * no interface to leave by, so it is dropped, nothing answers, and the
-     * lookup hangs until the kernel's own timeout expires — which is reported as
-     * a timeout rather than as a failure to resolve, and so reads like the
-     * network is slow instead of like the address is unusable.
-     *
-     * Android reports the zone as a `%wlan0` suffix on `hostAddress`, and the
-     * address is only usable with it kept. Nothing here trims it: an earlier
-     * version called `substringBefore('%')`, which turned a usable link-local
-     * server into an unroutable one.
-     */
-    private fun routable(host: String): Boolean =
-        !(host.startsWith("fe80:", ignoreCase = true) && !host.contains('%'))
-
-    /**
-     * IPv4 first, then IPv6.
-     *
-     * A phone commonly reports both, and its IPv6 entry is often a router's
-     * link-local address — reachable on one network and not the next. The
-     * kernel tries the servers in the order it is given them, and stops at the
-     * first that answers; putting the address family that shares the network's
-     * own numbering first means its timeout is spent on a server that works.
-     */
-    private fun preference(host: String): Int = if (host.contains(':')) 1 else 0
-
-    /**
-     * The private-DNS setting, for the one case the address list cannot answer.
-     *
-     * Since Android 9 a phone may use DNS-over-TLS, and then `dnsServers` holds a
-     * name rather than an address — a name this kernel cannot resolve *because*
-     * resolving it is the thing that is broken. There is no way for the shell to
-     * do that lookup for it, so this reports the situation instead of handing
-     * over an unusable value: the caller says so in the log, and the person can
-     * turn private DNS off or set it to automatic.
-     */
-    fun looksLikeDot(servers: String): Boolean =
-        servers.isNotBlank() && servers.split(',').any { part ->
-            val host = part.trim()
-            host.isNotEmpty() && host.none { it == ':' } && !host.all { it.isDigit() || it == '.' }
-        }
-
-    /** Whether the system requires a VPN or private DNS, which changes nothing
-     *  here but is worth saying in a bug report. */
-    fun describe(context: Context): String {
-        val manager = context.getSystemService(Context.CONNECTIVITY_SERVICE) as? ConnectivityManager
-            ?: return "no connectivity service"
-        val network = runCatching { manager.activeNetwork }.getOrNull() ?: return "no active network"
-        val properties = runCatching { manager.getLinkProperties(network) }.getOrNull()
-            ?: return "no link properties"
-        // The DHCP server arrived in API 30, not 23: asked for below that it is a
-        // NoSuchMethodError on a device this app supports (minSdk 26), and the
-        // answer is only ever read out of this string by a person debugging.
-        val dhcp = if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.R) {
-            properties.dhcpServerAddress?.hostAddress
-        } else {
-            null
-        }
-        return "iface=${properties.interfaceName} dns=${properties.dnsServers.joinToString(",") { it.hostAddress ?: "?" }}" +
-            (dhcp?.let { " dhcp=$it" } ?: "")
-    }
-}
@@ -1,87 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.content.Context
-import android.net.Uri
-import android.webkit.CookieManager
-import android.widget.Toast
-import kotlinx.coroutines.Dispatchers
-import kotlinx.coroutines.withContext
-import java.io.FileNotFoundException
-import java.net.HttpURLConnection
-import java.net.URL
-
-/**
- * Putting a file the page produced onto the phone's disk.
- *
- * A WebView has no downloads shelf and cannot write to the system's document
- * picker itself, so the two halves are joined here: the page asks, the picker
- * answers with a `content://` the app has been granted, and this copies the
- * bytes across.
- *
- * The request carries the WebView's own cookie, read out of the jar the page
- * authenticated in. That is the only reason this is not a plain `openStream`:
- * the kernel's download route is behind the session the page established, and a
- * shell that fetched without it would get a 403 and no explanation.
- */
-object Downloader {
-
-    suspend fun run(
-        context: Context,
-        url: String,
-        userAgent: String,
-        mime: String,
-        name: String,
-        target: Uri,
-    ): Result<Unit> = withContext(Dispatchers.IO) {
-        runCatching {
-            val connection = (URL(url).openConnection() as HttpURLConnection).apply {
-                connectTimeout = 15_000
-                readTimeout = 60_000
-                requestMethod = "GET"
-                if (userAgent.isNotBlank()) setRequestProperty("User-Agent", userAgent)
-                // The download route answers with the same session cookie the
-                // page is using; without it the kernel refuses the request.
-                cookieFor(url)?.let { setRequestProperty("Cookie", it) }
-            }
-            try {
-                val code = connection.responseCode
-                if (code !in 200..299) {
-                    throw IllegalStateException(
-                        context.getString(R.string.download_failed_status, code),
-                    )
-                }
-                val resolver = context.contentResolver
-                val output = resolver.openOutputStream(target, "w")
-                    ?: throw FileNotFoundException(target.toString())
-                output.use { out ->
-                    connection.inputStream.use { input -> input.copyTo(out) }
-                }
-                // `copyTo` answers with a byte count; what the caller wants to
-                // know is that there were no bytes missing, not how many.
-                Unit
-            } finally {
-                connection.disconnect()
-            }
-        }
-    }
-
-    /**
-     * The cookie the page authenticated with, if any.
-     *
-     * `CookieManager` is the WebView's own store, so this asks the exact jar the
-     * request will be checked against rather than a copy this shell kept.
-     */
-    private fun cookieFor(url: String): String? =
-        runCatching { CookieManager.getInstance().getCookie(url) }.getOrNull()?.takeIf { it.isNotBlank() }
-
-    /** A short receipt for a tap that otherwise does nothing visible. */
-    fun report(context: Context, name: String, result: Result<Unit>) {
-        val message = result.fold(
-            onSuccess = { context.getString(R.string.download_saved, name) },
-            onFailure = { failure ->
-                context.getString(R.string.download_failed, failure.message ?: "")
-            },
-        )
-        Toast.makeText(context, message, Toast.LENGTH_LONG).show()
-    }
-}
@@ -1,741 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.content.Context
-import android.content.res.AssetManager
-import org.json.JSONArray
-import org.json.JSONObject
-import android.util.Log
-import kotlinx.coroutines.Dispatchers
-import kotlinx.coroutines.delay
-import kotlinx.coroutines.withContext
-import java.io.File
-import java.io.FileOutputStream
-import java.net.HttpURLConnection
-import java.net.URL
-import java.security.SecureRandom
-import java.util.Locale
-
-/**
- * The Reasonix kernel, as this app runs it: one child process serving loopback,
- * plus the two things it needs on disk to start — the page it serves and a
- * credential for this launch.
- *
- * The shape is the desktop host's (`cmd/reasonix-studio-host`), narrowed to what
- * a phone does not have. That host binds a socket, mints a credential, speaks a
- * handshake down a pipe and holds a lease; here the credential goes through
- * `--token-file` and the bound address through `--port-file`, because a phone's
- * app process has no pipe to hand a child and no reason to care which port was
- * free. Everything after that — the kernel's own gate, its auth middleware, its
- * CSRF check — is untouched, so the security of the surface is the one upstream
- * tested, not one this shell invented.
- *
- * One instance lives in the app process. The activity asks it for a URL; the
- * foreground service asks it to stay alive. Both see the same process, so there
- * is nothing to synchronise beyond a lock on start/stop.
- */
-class Kernel(private val context: Context) {
-
-    /** Where everything the kernel owns lives. Nothing here is backed up. */
-    private val home = File(context.filesDir, "home")
-
-    /**
-     * The folder the agent works in, which is not the folder the kernel lives in.
-     *
-     * These used to be the same directory, and that is what made the workbench
-     * unreadable: the kernel's home holds `config.toml`, `projects/`, `cache/`,
-     * `stats/` and the rest of its own bookkeeping, and a workspace pointed at
-     * it shows all of that as if the person had put it there. Reported as "why
-     * does the default workspace have so many files, the desktop version does
-     * not generate these" — the desktop does generate them, in a different
-     * directory, because its workspace is whatever folder it was opened on.
-     *
-     * Split apart, the workbench shows what the agent produced and nothing else.
-     * `write_roots` follows the workspace, so the file tools stay confined to
-     * the folder the person actually works in rather than to the kernel's own
-     * home — which is the stricter answer, and the one the desktop already has.
-     */
-    private val workspaceDir = File(context.filesDir, "workspace")
-
-    /**
-     * The directory the kernel calls `REASONIX_HOME`, for the one caller that
-     * has to put a file where the kernel can read it — a private key the person
-     * picked from their Documents folder.
-     */
-    fun homeDir(): File = home
-
-    /** The folder the agent works in. */
-    fun workspaceDir(): File = workspaceDir
-
-    /** The built page, unpacked once per version. */
-    private val pageDir = File(context.filesDir, "page")
-
-    private val tokenFile = File(home, "remote/launch.token")
-    private val portFile = File(home, "remote/serve.port")
-    private val pidFile = File(home, "remote/serve.pid")
-
-    private val lock = Any()
-
-    @Volatile
-    private var process: Process? = null
-
-    /** The port the shell's browser HTTP server listens on, set by the
-     *  activity before `start()` is called. Zero means no server is up, and
-     *  the kernel leaves its browser tools suspended (the correct degradation
-     *  on a phone with no Chromium). */
-    @Volatile
-    var browserShellPort: Int = 0
-
-    /** The bearer credential for that server, set beside the port. Sent on
-     *  every browser call so the server can refuse anything that did not come
-     *  from the kernel this shell started — loopback on a phone is reachable
-     *  by every other app. Empty when there is no server. */
-    @Volatile
-    var browserShellToken: String = ""
-
-    /** What the WebView should load, once the kernel has bound a port. */
-    @Volatile
-    var origin: String? = null
-        private set
-
-    /**
-     * The launch credential, as the URL fragment the page trades for a cookie.
-     * Kept in memory only: the file on disk belongs to the kernel's own
-     * bookkeeping, and a phone's app data is not a place to leave a second copy.
-     */
-    @Volatile
-    private var token: String? = null
-
-    /**
-     * The URL the WebView loads first. It carries the credential in the
-     * fragment, which the page exchanges for an HttpOnly cookie before it makes
-     * any request — the same path a shared `reasonix web` link takes, so the
-     * token never enters a request line, a history entry or a log.
-     */
-    fun entryUrl(): String? {
-        val origin = origin ?: return null
-        val token = token ?: return null
-        return "$origin/#token=$token"
-    }
-
-    /** The plain origin, for reloads once the cookie is set. */
-    fun plainUrl(): String? = origin
-
-    fun isRunning(): Boolean = process?.isAlive == true
-
-    /**
-     * Bring the kernel up, or answer the URL of the one already up.
-     *
-     * Starting is idempotent by intent: the activity and the service both call
-     * this, and whichever arrives second joins the running kernel instead of
-     * racing it for the port.
-     */
-    suspend fun start(): Result<String> = withContext(Dispatchers.IO) {
-        synchronized(lock) {
-            origin?.let { return@withContext Result.success(it) }
-
-            runCatching {
-                val binary = kernelBinary()
-                publishPage()
-                prepareHome()
-                val launchToken = mintToken()
-                token = launchToken
-
-                val child = launch(binary, launchToken)
-
-                // The port is the kernel's answer, not this shell's guess. It
-                // writes the file after binding, so waiting on the file is
-                // waiting on the bind.
-                val bound = awaitPort(child)
-                val url = "http://$bound"
-
-                // A kernel that bound a port and then died would look healthy
-                // here and broken in the WebView; the two facts are checked
-                // together, and only a kernel that is answering is adopted.
-                awaitServing(child, url)
-                origin = url
-                process = child
-                url
-            }.onFailure { origin = null; process = null }
-        }
-    }
-
-    /**
-     * Stop the kernel.
-     *
-     * The graceful signal first: the kernel drains its listeners and closes the
-     * runtimes it opened, which is what lets a later launch adopt the sessions
-     * rather than find them locked. `destroyForcibly` is the second attempt, and
-     * only after a wait — a phone that is being shut down should not keep a
-     * half-killed agent.
-     */
-    fun stop() {
-        synchronized(lock) {
-            val p = process ?: return
-            process = null
-            origin = null
-            token = null
-            runCatching {
-                p.destroy()
-                if (!p.waitFor(3, java.util.concurrent.TimeUnit.SECONDS)) {
-                    p.destroyForcibly()
-                }
-            }
-            portFile.delete()
-            pidFile.delete()
-        }
-    }
-
-    // ---- the two inputs that ship inside the APK -------------------------
-
-    /**
-     * The kernel binary.
-     *
-     * It is packaged as a native library, so Android has already extracted it to
-     * a directory it allows execution from and given it the executable bit —
-     * which is the one thing a file under app storage does not get for free.
-     */
-    private fun kernelBinary(): File {
-        val dir = File(context.applicationInfo.nativeLibraryDir)
-        val binary = File(dir, "libreasonix.so")
-        if (!binary.exists()) {
-            throw IllegalStateException(
-                "内核未随 APK 安装：${binary.path}\n这是一次构建问题，请重新安装完整 APK。",
-            )
-        }
-        return binary
-    }
-
-    /**
-     * Write the bundled page where the kernel can serve it.
-     *
-     * It goes to app storage rather than being served out of the APK because the
-     * kernel's `--page` is a directory on a real filesystem; there is no
-     * archive-aware variant, and adding one would be a change to the kernel for
-     * this shell's convenience alone.
-     *
-     * Unpacking is versioned: the marker holds the app's version name, so an
-     * update replaces the page and a relaunch does not. A partially written
-     * directory — the app killed mid-copy — is caught by the same marker never
-     * having been written, so the next launch rewrites it.
-     */
-    private fun publishPage() {
-        val marker = File(pageDir, ".unpacked")
-        val stamp = "${BuildConfig.VERSION_NAME}:${BuildConfig.VERSION_CODE}"
-        if (marker.isFile && marker.readText().trim() == stamp) return
-
-        pageDir.deleteRecursively()
-        pageDir.mkdirs()
-        copyAssetTree(context.assets, "page", pageDir)
-
-        if (!File(pageDir, "index.html").isFile) {
-            throw IllegalStateException("随 APK 分发的页面不完整：缺少 index.html")
-        }
-        marker.writeText(stamp)
-    }
-
-    /**
-     * A directory tree out of assets.
-     *
-     * `AssetManager.list` on a file answers with an empty array, so a "file" is
-     * told from a directory by trying to open it — which is the only signal the
-     * API offers and the reason this is not a one-liner.
-     */
-    private fun copyAssetTree(assets: AssetManager, from: String, to: File) {
-        val children = assets.list(from).orEmpty()
-        if (children.isEmpty()) {
-            // A leaf: open it as a stream. A directory that happens to be empty
-            // is indistinguishable and harmless to skip.
-            runCatching {
-                assets.open(from).use { input ->
-                    FileOutputStream(to).use { output -> input.copyTo(output) }
-                }
-            }
-            return
-        }
-        to.mkdirs()
-        for (child in children) {
-            copyAssetTree(assets, "$from/$child", File(to, child))
-        }
-    }
-
-    /**
-     * Prepare the kernel's home.
-     *
-     * `REASONIX_HOME` is what every path in the kernel hangs off — config,
-     * sessions, the remote state directory its token file lands in — so pointing
-     * it at app storage is what keeps a phone's Reasonix separate from anything
-     * else and away from global storage entirely.
-     */
-    private fun prepareHome() {
-        File(home, "remote").mkdirs()
-        // The workspace directory not existing yet is what marks the first run
-        // after this build split it from the kernel's home. That is the one
-        // moment a migration is due, and doing it here means it happens before
-        // the kernel is asked for anything.
-        val firstSplit = !workspaceDir.exists()
-        workspaceDir.mkdirs()
-        if (firstSplit) {
-            adoptSessionsFromHome()
-        }
-        // Every launch, not only the first. The rewrite is cheap and
-        // idempotent, and an entry that survived the first pass would otherwise
-        // survive forever: `firstSplit` is spent the moment the directory
-        // exists, so a build that cleaned the list once and missed leaves the
-        // home on the sidebar for good. Running it unconditionally costs one
-        // small read and makes the whole thing self-correcting.
-        forgetHomeFromWorkspaceList()
-    }
-
-    /**
-     * Drop the kernel's own home from the sidebar's remembered projects.
-     *
-     * The list is written whenever a pane opens, and until this build the pane
-     * opened on the home itself — so an upgraded install carries a "project"
-     * that is really the kernel's own directory. Leaving it there would put the
-     * person back where they started: one tap on it and the workbench is full
-     * of `config.toml`, `cache/` and `projects/` again, which is the complaint
-     * this whole change answers. Only that one path is removed; every folder
-     * someone added on purpose stays.
-     *
-     * Parsed rather than pattern-matched, because the file's two shapes — a bare
-     * array of paths and an object with `paths` and `launch` — both have to keep
-     * whatever else they hold.
-     */
-    private fun forgetHomeFromWorkspaceList() {
-        val file = File(home, "serve-workspaces.json")
-        if (!file.isFile) return
-        val text = runCatching { file.readText() }.getOrNull() ?: return
-        val wrong = runCatching { homeNames() }.getOrNull() ?: return
-        val rewritten = runCatching { withoutHome(text, wrong) }.getOrNull() ?: return
-        if (!rewritten.dropped) return
-        runCatching { file.writeText(rewritten.text) }
-            .onSuccess { Log.i(TAG, "dropped the kernel home from the project list") }
-            .onFailure { Log.w(TAG, "could not rewrite the project list: ${it.message}") }
-    }
-
-    /**
-     * Every name of the kernel's home on this device — not just the one this
-     * process happens to write.
-     *
-     * Android shows one directory under two names: `/data/data/<pkg>` and
-     * `/data/user/0/<pkg>` are two mounts of the same tree, the same inode
-     * either way, and which name a path carries depends on how it was written
-     * rather than on where it points. The list is written by the kernel, whose
-     * working directory this shell set from a path it built itself; a lookup
-     * that compared against only one spelling would silently keep the entry
-     * this exists to remove. Both names are compared, so either spelling is
-     * recognised.
-     */
-    private fun homeNames(): Set<String> {
-        val path = home.absolutePath
-        val names = mutableSetOf(path)
-        val user = "/data/user/0/"
-        val data = "/data/data/"
-        when {
-            path.startsWith(user) -> names.add(data + path.removePrefix(user))
-            path.startsWith(data) -> names.add(user + path.removePrefix(data))
-        }
-        return names
-    }
-
-    /** What a rewrite produced, and whether it actually removed anything. */
-    private class Rewrite(val text: String, val dropped: Boolean)
-
-    /**
-     * The list without the home, in whichever shape it arrived, and whether
-     * anything was actually dropped from it.
-     *
-     * The flag matters because `JSONObject` re-serialises: it escapes every
-     * `/` and does not preserve key order, so a round trip never reproduces
-     * the original bytes. Deciding to write on "the text differs" alone would
-     * therefore rewrite the kernel's file on every single launch, log a drop
-     * that never happened, and race the kernel for a file it also writes.
-     */
-    private fun withoutHome(text: String, wrong: Set<String>): Rewrite {
-        if (text.trimStart().startsWith("[")) {
-            val paths = JSONArray(text)
-            val kept = JSONArray()
-            var dropped = false
-            for (i in 0 until paths.length()) {
-                val path = paths.optString(i)
-                if (path in wrong) {
-                    dropped = true
-                    continue
-                }
-                kept.put(path)
-            }
-            return if (dropped) Rewrite(kept.toString(), true) else Rewrite(text, false)
-        }
-        val list = JSONObject(text)
-        val kept = JSONArray()
-        var dropped = false
-        val paths = list.optJSONArray("paths")
-        if (paths != null) {
-            for (i in 0 until paths.length()) {
-                val path = paths.optString(i)
-                if (path in wrong) {
-                    dropped = true
-                    continue
-                }
-                kept.put(path)
-            }
-            if (dropped) list.put("paths", kept)
-        }
-        if (list.optString("launch") in wrong) {
-            list.remove("launch")
-            dropped = true
-        }
-        return if (dropped) Rewrite(list.toString(), true) else Rewrite(text, false)
-    }
-
-    /**
-     * Move the conversations that were filed under the kernel's own home.
-     *
-     * The transcript store is keyed by the workspace root — one directory per
-     * project, under `<state>/projects/<slug>/sessions` — and until this build
-     * the root was the home itself. Splitting the two therefore moves every
-     * existing conversation to a slug nothing looks under, which reads as "my
-     * history is gone" rather than as a reorganisation. So the old directory is
-     * renamed onto the new slug once.
-     *
-     * Renaming the project directory rather than just `sessions/` carries the
-     * rest of it too: `memory/` is filed the same way, and a project's memory
-     * lost while its transcripts were kept would be a stranger failure than
-     * either alone.
-     *
-     * The slug is the absolute path with its separators flattened, which is
-     * what `config.WorkspaceSlug` does on the kernel's side. Only the home's own
-     * slug is touched: a slug for some other folder is left exactly where it is,
-     * because those are projects someone added on purpose.
-     */
-    private fun adoptSessionsFromHome() {
-        val projects = File(home, "projects")
-        if (!projects.isDirectory) return
-        val from = projects.resolve(slugOf(home.absolutePath))
-        if (!from.isDirectory) return
-        val to = projects.resolve(slugOf(workspaceDir.absolutePath))
-        // A target already there means a previous run half-finished, or the
-        // person opened this folder before. Either way its contents win:
-        // merging two transcript directories is not a thing this can do safely,
-        // and the old one is still on disk to be looked at.
-        if (to.exists()) {
-            Log.i(TAG, "${to.name} already exists; leaving ${from.name} in place")
-            return
-        }
-        if (!from.renameTo(to)) {
-            Log.w(TAG, "could not move ${from.name} to ${to.name}; history stays under the old slug")
-        } else {
-            Log.i(TAG, "moved conversations from ${from.name} to ${to.name}")
-        }
-    }
-
-    /** The kernel's `config.WorkspaceSlug`, for the two paths that matter here. */
-    private fun slugOf(path: String): String = path.replace('/', '-').replace('\\', '-').replace(':', '-')
-
-    /**
-     * The launch credential.
-     *
-     * Minted here rather than read from configuration: a credential a person can
-     * set is one anything that can read their files can present. 32 bytes, hex,
-     * the same shape the desktop host mints.
-     */
-    private fun mintToken(): String {
-        val bytes = ByteArray(32)
-        SecureRandom().nextBytes(bytes)
-        val token = bytes.joinToString("") { "%02x".format(it) }
-        tokenFile.parentFile?.mkdirs()
-        tokenFile.writeText("$token\n")
-        // 0600 by intent, and the app's own private directory means it is
-        // already unreachable to every other app on the device.
-        runCatching {
-            tokenFile.setReadable(false, false)
-            tokenFile.setReadable(true, true)
-            tokenFile.setWritable(false, false)
-            tokenFile.setWritable(true, true)
-        }
-        return token
-    }
-
-    /**
-     * Start the child.
-     *
-     * `--addr 127.0.0.1:0` lets the kernel pick a free port; `--port-file`
-     * makes it announce which. `--auth token` is the mode a phone needs: the
-     * page authenticates, and a state-changing request without the credential
-     * is refused. `--no-open` is belt and braces — a serve never opens a
-     * browser, and a phone has no browser to open.
-     */
-    private fun launch(binary: File, token: String): Process {
-        portFile.delete()
-        pidFile.delete()
-        // The folder the agent works in, resolved before the command line is
-        // built so it can be named on it. See the note under the environment
-        // block for why naming it is not the same as setting the directory.
-        val workspace = launchWorkspaceDir()
-        val command = listOf(
-            binary.absolutePath,
-            "serve",
-            "--addr", "127.0.0.1:0",
-            "--auth", "token",
-            "--token-file", tokenFile.absolutePath,
-            "--port-file", portFile.absolutePath,
-            "--pid-file", pidFile.absolutePath,
-            "--no-open",
-            "-page", pageDir.absolutePath,
-            // The project root, said outright. Without this the kernel derives
-            // it from its working directory, and on this phone the two did not
-            // agree: the sidebar named the shared folder, `write_roots` confined
-            // the file tools to the app's own directory, and every write into the
-            // folder the person was looking at came back
-            // `workspace.write_outside_scope`. A flag is a fact the kernel was
-            // told rather than one it inferred from a mount it sees differently.
-            "--dir", workspace.absolutePath,
-        )
-        val builder = ProcessBuilder(command)
-        builder.environment().apply {
-            put("REASONIX_HOME", home.absolutePath)
-            put("HOME", home.absolutePath)
-            // The kernel's own language, for its tool failures, approval
-            // prompts and error sentences. Without this it detects from the
-            // POSIX locale variables, and Android sets none of them for an
-            // app: the kernel then answers in English while the page in front
-            // of the person is Chinese, and every sentence added for a Chinese
-            // reader was dead text. The tag, not the bare language, so a
-            // Traditional Chinese device resolves to zh-TW rather than zh.
-            // An explicit `language` in the kernel's config.toml still wins:
-            // DetectLanguage applies the configured value over this.
-            val languageTag = Locale.getDefault().toLanguageTag()
-            if (languageTag.isNotBlank() && languageTag != "und") {
-                put("REASONIX_LANG", languageTag)
-            }
-            // Android has no /tmp. Everything that wants a scratch directory —
-            // the shell tool, tool output spooling, sqlite journals — reads this.
-            put("TMPDIR", context.cacheDir.absolutePath)
-            // The phone's own DNS servers. A CGO-free Go binary cannot read them
-            // for itself — Android keeps them in `netd`, and Go looks for a
-            // `/etc/resolv.conf` that does not exist — so without this the kernel
-            // resolves nothing and every provider probe reports "连接失败" on a
-            // device whose browser reaches the same address fine.
-            val dns = Dns.servers(context)
-            if (dns.isNotBlank()) {
-                put("REASONIX_DNS", dns)
-                Log.i(TAG, "kernel DNS: $dns (${Dns.describe(context)})")
-                if (Dns.looksLikeDot(dns)) {
-                    // Private DNS (DNS-over-TLS) is on, and it hands back a
-                    // hostname this kernel cannot resolve by definition. Saying
-                    // so here is what turns "连不上" into something actionable.
-                    Log.w(TAG, "kernel DNS looks like a private-DNS hostname; turn private DNS off or set it to automatic")
-                }
-            } else {
-                Log.w(TAG, "kernel DNS: none reported by the system")
-            }
-            // The shell's browser server, when one is listening. The kernel
-            // reads this to bind its browser tools to the shell's WebView
-            // instead of looking for a Chromium binary that does not exist on
-            // a phone. When absent (zero), the variable is not set and the
-            // kernel keeps the browser tools suspended — the correct
-            // degradation for a phone with no engine to drive.
-            val shellPort = browserShellPort
-            if (shellPort > 0) {
-                put("REASONIX_SHELL_BROWSER_PORT", shellPort.toString())
-                // The credential the server checks. Set only alongside the
-                // port: a token with nothing listening would be a variable the
-                // kernel carries for no reason.
-                val shellToken = browserShellToken
-                if (shellToken.isNotBlank()) {
-                    put("REASONIX_SHELL_BROWSER_TOKEN", shellToken)
-                }
-            }
-        }
-        // The agent's folder, not the kernel's. `REASONIX_HOME` above is still
-        // where the config, the sessions and the transcripts live; this is what
-        // the workspace root resolves to, because the kernel reads the working
-        // directory for it.
-        //
-        // The folder the person last opened, when it is still usable — the same
-        // answer the desktop host gives (`launchWorkspace`), and the reason a
-        // phone session used to work in the app's own `files/workspace` while
-        // the sidebar said it was in the shared folder: the launch entry in
-        // `serve-workspaces.json` was written and then ignored, and `write_roots`
-        // is derived from the working directory, so the file tools were confined
-        // to the app's private folder and every write into the shared one was
-        // refused. `workspaceDir` is still the answer for a first run and for a
-        // folder since deleted, but the choice is never silent — see
-        // [launchWorkspaceDir]. `workspace` was resolved above so it could also
-        // be named on the command line.
-        builder.directory(workspace)
-        // `PWD` is what the kernel's own `os.Getwd` prefers: Go uses it whenever
-        // it names the same directory as `.`, which is exactly this case, and it
-        // is the only way to make the kernel resolve its root to the string this
-        // shell chose. Without it `os.Getwd` falls back to the kernel's view of
-        // the mount, and on Android `/storage/emulated/0/...` and the path the
-        // FUSE layer reports for it need not be spelled the same way. The kernel
-        // derives `write_roots` from that root, so the two spellings have to be
-        // the same one.
-        builder.environment().put("PWD", workspace.absolutePath)
-        builder.redirectErrorStream(true)
-        val process = builder.start()
-
-        // The kernel's own log is the only place a startup failure says why, and
-        // a phone has no terminal to read. It is drained on a thread and kept
-        // ring-buffered so the About screen can show the last of it.
-        Thread({ drain(process) }, "kernel-log").apply { isDaemon = true }.start()
-        return process
-    }
-
-    /**
-     * The folder the person last opened, or null when there is not one.
-     *
-     * `serve-workspaces.json` is the kernel's own record of the projects the
-     * sidebar lists, and its `launch` entry is which of them a session starts
-     * in. Both shapes the file has are read — a bare array of paths, and an
-     * object with `paths` and `launch` — because both are written in the wild
-     * and a reader that only knew one would silently fall back.
-     *
-     * The entry has to be one of the listed paths, which is the same rule the
-     * kernel applies when it reads this file (`workspacelist.Read` blanks a
-     * `launch` that is not in `paths`). It matters more here than there: the
-     * working directory decides `write_roots`, so whatever this answers is
-     * where the agent's file tools may write — and on a phone that permission
-     * reaches all of shared storage. A file with a `launch` outside its own
-     * list is one this shell did not write, and it gets the listed answer.
-     *
-     * Null for anything unusable: no file, unreadable JSON, a launch that is
-     * not among the listed folders, an entry that has since been deleted, or a
-     * path that is now a file. Every one of those means the same thing to the
-     * caller — start where this shell always started.
-     */
-    private fun rememberedLaunchDir(): File? {
-        val file = File(home, "serve-workspaces.json")
-        if (!file.isFile) return null
-        val text = runCatching { file.readText() }.getOrNull() ?: return null
-        val launch = runCatching {
-            // A bare array is the file's other shape — the one older kernels
-            // wrote, and therefore the one a person upgrading is most likely to
-            // have. Without this branch `JSONObject(text)` throws on it and the
-            // whole function falls back, which is the same mismatch (the
-            // sidebar naming one folder, the session working in another) this
-            // exists to fix. `withoutHome` below reads both shapes the same way.
-            if (text.trimStart().startsWith("[")) {
-                val arr = JSONArray(text)
-                if (arr.length() == 0) return@runCatching null
-                arr.optString(0)
-            } else {
-                val json = JSONObject(text)
-                val paths = json.optJSONArray("paths")
-                    ?.let { arr -> (0 until arr.length()).mapNotNull { arr.optString(it).takeIf(String::isNotBlank) } }
-                    .orEmpty()
-                if (paths.isEmpty()) return@runCatching null
-                json.optString("launch").takeIf { it.isNotBlank() && it in paths } ?: paths.first()
-            }
-        }.getOrNull()?.takeIf { it.isNotBlank() } ?: return null
-        val dir = runCatching { File(launch) }.getOrNull() ?: return null
-        if (!dir.isDirectory) {
-            // A launch entry this process cannot stat is exactly the failure the
-            // caller exists to avoid, and falling back silently is what made it
-            // take three releases to notice: the session opened in the app's own
-            // folder while the sidebar named the shared one. A phone has no
-            // terminal, so the log is the only place the reason can be said —
-            // and `exists`/`readable` tell apart "deleted" from "not permitted",
-            // which need different answers from the person.
-            Log.w(
-                TAG,
-                "launch entry $launch is not a usable directory " +
-                    "(exists=${dir.exists()}, readable=${dir.canRead()})",
-            )
-            return null
-        }
-        return dir
-    }
-
-    /**
-     * The folder the kernel should start in.
-     *
-     * Either the launch entry — the folder the person last opened, which is
-     * also the one the sidebar shows — or the app's own folder when there is
-     * none. Whichever it is, it is logged, because the two disagreeing is the
-     * bug this exists to fix and a silent choice is what let it survive.
-     */
-    private fun launchWorkspaceDir(): File {
-        val remembered = rememberedLaunchDir()
-        if (remembered != null) {
-            Log.i(TAG, "kernel workspace: ${remembered.absolutePath} (the remembered launch project)")
-            return remembered
-        }
-        Log.i(TAG, "kernel workspace: ${workspaceDir.absolutePath} (no remembered launch project)")
-        return workspaceDir
-    }
-
-    private val logLines = ArrayDeque<String>()
-
-    private fun drain(process: Process) {
-        runCatching {
-            process.inputStream.bufferedReader().useLines { lines ->
-                lines.forEach { line ->
-                    Log.i(TAG, line)
-                    synchronized(logLines) {
-                        logLines.addLast(line)
-                        while (logLines.size > LOG_KEEP) logLines.removeFirst()
-                    }
-                }
-            }
-        }
-    }
-
-    /** The kernel's tail, newest last, for a bug report or the About screen. */
-    fun recentLog(): String = synchronized(logLines) { logLines.joinToString("\n") }
-
-    private fun awaitPort(process: Process): String {
-        val deadline = System.currentTimeMillis() + START_TIMEOUT_MS
-        while (System.currentTimeMillis() < deadline) {
-            if (!process.isAlive) {
-                throw IllegalStateException("内核启动即退出。\n${recentLog().takeLast(2000)}")
-            }
-            val text = if (portFile.isFile) runCatching { portFile.readText().trim() }.getOrNull() else null
-            // The file is written by the kernel on the way to serving, and the
-            // kernel rewrites it with the real port when it was given :0.
-            if (!text.isNullOrBlank() && text.contains(":")) return text
-            Thread.sleep(POLL_MS)
-        }
-        throw IllegalStateException("内核在 ${START_TIMEOUT_MS / 1000}s 内没有报告监听地址。\n${recentLog().takeLast(2000)}")
-    }
-
-    /**
-     * Wait until the bound port answers.
-     *
-     * A port that is open but not yet serving is the normal first second of a
-     * kernel's life; a WebView pointed at it would show a connection error and
-     * cache the failure, so the wait belongs here rather than in the page.
-     */
-    private fun awaitServing(process: Process, origin: String) {
-        val deadline = System.currentTimeMillis() + START_TIMEOUT_MS
-        while (System.currentTimeMillis() < deadline) {
-            if (!process.isAlive) {
-                throw IllegalStateException("内核在开始服务前退出。\n${recentLog().takeLast(2000)}")
-            }
-            if (probe("$origin/")) return
-            Thread.sleep(POLL_MS)
-        }
-        throw IllegalStateException("内核已监听 $origin，但页面没有响应。\n${recentLog().takeLast(2000)}")
-    }
-
-    private fun probe(url: String): Boolean = runCatching {
-        (URL(url).openConnection() as HttpURLConnection).run {
-            connectTimeout = 1000
-            readTimeout = 1000
-            requestMethod = "GET"
-            try {
-                responseCode in 200..399
-            } finally {
-                disconnect()
-            }
-        }
-    }.getOrDefault(false)
-
-    private companion object {
-        const val TAG = "ReasonixKernel"
-        const val LOG_KEEP = 400
-        const val POLL_MS = 120L
-        const val START_TIMEOUT_MS = 30_000L
-    }
-}
@@ -1,139 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.Manifest
-import android.app.Notification
-import android.app.PendingIntent
-import android.app.Service
-import android.content.Context
-import android.content.Intent
-import android.content.pm.PackageManager
-import android.content.pm.ServiceInfo
-import android.os.Build
-import android.os.IBinder
-import androidx.core.app.NotificationCompat
-import androidx.core.app.NotificationManagerCompat
-import androidx.core.app.ServiceCompat
-import androidx.core.content.ContextCompat
-import kotlinx.coroutines.CoroutineScope
-import kotlinx.coroutines.Dispatchers
-import kotlinx.coroutines.SupervisorJob
-import kotlinx.coroutines.cancel
-import kotlinx.coroutines.launch
-
-/**
- * Keeps the kernel alive while the app is not on screen.
- *
- * The reason this exists at all: an agent turn that takes minutes is a normal
- * turn, and Android reclaims a backgrounded app's child process within seconds
- * of the screen going off. Without the service, pocketing the phone would kill
- * a running task — the one failure mode that makes a phone a bad client for
- * something that works on its own.
- *
- * A foreground service needs a notification the person can see, which is also
- * the honest thing to show: the kernel is running and what it is doing lives on
- * another machine.
- */
-class KernelService : Service() {
-
-    private val scope = CoroutineScope(SupervisorJob() + Dispatchers.Main)
-
-    override fun onBind(intent: Intent?): IBinder? = null
-
-    override fun onStartCommand(intent: Intent?, flags: Int, startId: Int): Int {
-        // Promoted before `start()`: a service that starts work first and
-        // notifies after has a window in which the system may kill it, and the
-        // work it would have been protecting is exactly what is being started.
-        startForegroundNotification()
-
-        val kernel = (application as App).kernel
-        scope.launch {
-            kernel.start().onFailure { failure ->
-                // A kernel that will not start is a fact the person needs, and a
-                // notification that only ever said "running" would be a lie.
-                notify(SERVICE_NOTICE, getString(R.string.notify_failed_title), failure.message.orEmpty())
-                stopSelf()
-            }
-        }
-        // Sticky, so a kernel killed under memory pressure comes back when there
-        // is room — which is what a person who left a task running expects.
-        return START_STICKY
-    }
-
-    override fun onDestroy() {
-        scope.cancel()
-        super.onDestroy()
-    }
-
-    private fun startForegroundNotification() {
-        val notification = buildNotification(
-            getString(R.string.notify_running_title),
-            getString(R.string.notify_running_body),
-        )
-        // `dataSync` matches the manifest's declaration. On 34+ the type is
-        // mandatory and a mismatch throws, so it is named in both places.
-        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.UPSIDE_DOWN_CAKE) {
-            ServiceCompat.startForeground(
-                this,
-                SERVICE_NOTICE,
-                notification,
-                ServiceInfo.FOREGROUND_SERVICE_TYPE_DATA_SYNC,
-            )
-        } else {
-            startForeground(SERVICE_NOTICE, notification)
-        }
-    }
-
-    private fun buildNotification(title: String, body: String): Notification {
-        val open = PendingIntent.getActivity(
-            this,
-            0,
-            Intent(this, MainActivity::class.java).apply {
-                flags = Intent.FLAG_ACTIVITY_NEW_TASK or Intent.FLAG_ACTIVITY_CLEAR_TOP
-            },
-            PendingIntent.FLAG_UPDATE_CURRENT or PendingIntent.FLAG_IMMUTABLE,
-        )
-        return NotificationCompat.Builder(this, App.CHANNEL_KERNEL)
-            .setSmallIcon(R.drawable.ic_notification)
-            .setContentTitle(title)
-            .setContentText(body)
-            .setContentIntent(open)
-            // The notification is a receipt, not an alarm; swiping it away
-            // leaves the kernel running, which is what the person meant by
-            // swiping it.
-            .setOngoing(false)
-            .setSilent(true)
-            .setPriority(NotificationCompat.PRIORITY_MIN)
-            .build()
-    }
-
-    private fun notify(id: Int, title: String, body: String) {
-        // The notification permission is granted at runtime from Android 13 and
-        // the person can refuse it. Without this check the call throws
-        // SecurityException and takes the service down with it — the opposite of
-        // what a notification that could not be shown should do. Below 13 the
-        // permission does not exist and posting always succeeds.
-        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.TIRAMISU &&
-            ContextCompat.checkSelfPermission(this, Manifest.permission.POST_NOTIFICATIONS) !=
-            PackageManager.PERMISSION_GRANTED
-        ) {
-            return
-        }
-        runCatching {
-            NotificationManagerCompat.from(this).notify(id, buildNotification(title, body))
-        }
-    }
-
-    companion object {
-        private const val SERVICE_NOTICE = 1
-
-        /** Idempotent: called on every start of the activity, and by the boot
-         *  path, and safe to repeat. */
-        fun start(context: Context) {
-            ContextCompat.startForegroundService(context, Intent(context, KernelService::class.java))
-        }
-
-        fun stop(context: Context) {
-            context.stopService(Intent(context, KernelService::class.java))
-        }
-    }
-}
@@ -1,744 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.content.ClipData
-import android.content.ClipboardManager
-import android.content.ContentValues
-import android.content.Context
-import android.content.res.Configuration
-import android.net.Uri
-import android.os.Build
-import android.os.Environment
-import android.os.Handler
-import android.os.Looper
-import android.provider.MediaStore
-import android.webkit.JavascriptInterface
-import android.webkit.MimeTypeMap
-import java.io.File
-import java.util.concurrent.CountDownLatch
-import java.util.concurrent.TimeUnit
-import java.util.concurrent.atomic.AtomicReference
-
-/**
- * What the page can ask the phone for.
- *
- * This is the whole surface, and it is deliberately tiny. Android's
- * `@JavascriptInterface` is a bridge into the app's process from a WebView whose
- * content is a web page: every method here is reachable by anything that can run
- * script in that page, so each one is written as a verb with no address
- * parameter it does not need, and none of them can name a path outside the
- * kernel's own home.
- *
- * It is deliberately *not* the Electron preload bridge. That one exposes window
- * management, a filesystem walker, an editor launcher and a browser view host —
- * all of which are meaningless on a phone, and the two that were not (open a URL,
- * save a file) already have a way to happen here: `WebViewClient` routes external
- * links to the browser, and the page's own download reaches the WebView's
- * `DownloadListener`.
- *
- * The one thing a phone must add is `importIntoHome`: the kernel runs on this
- * device and can only read paths under its own home, so a private key the person
- * has in their Documents folder has to be copied across. Nothing else the page
- * wants is out of its reach.
- */
-class NativeBridge(
-    private val context: Context,
-    private val kernel: Kernel,
-    private val onOpen: (String) -> Unit,
-) {
-
-    /**
-     * Whether the document calling may use this bridge.
-     *
-     * `addJavascriptInterface` makes the object reachable from whatever page the
-     * WebView holds, so a link the model produced — or a cleartext page on the
-     * LAN — would otherwise get `importIntoHome`, `pickFolder`, `hostCall` and
-     * the clipboard. The activity answers by comparing the main frame's origin
-     * against the kernel's and the one paired computer the person confirmed.
-     *
-     * Anything that reaches the device checks this first. The five methods that
-     * do not — version, systemTheme, safeArea, setLayerOpen, setRunning — report
-     * back what the page already knows about its own window and grant nothing.
-     */
-    var trusted: () -> Boolean = { false }
-
-    /** The version the page reads to decide whether this bridge is the one it
-     *  knows. Bumped when a method's contract changes, not when one is added. */
-    @JavascriptInterface
-    fun version(): String = BRIDGE_VERSION
-
-    /**
-     * Copy a file the person picks into the kernel's home, answering with its
-     * path there — or "" when they dismissed the picker.
-     *
-     * `kind` names the directory inside the home, because that is a fact about
-     * the kernel's layout and not about the file: an SSH private key belongs in
-     * `.ssh`, and a dotenv the kernel reads for passphrases belongs at the root
-     * of the home where `config.Load` looks for it.
-     *
-     * The path that comes back is *inside the home*, never the document
-     * provider's URI. That is the point: the kernel is a child process with no
-     * access to the provider, so a `content://` would be a path it cannot open.
-     */
-    @JavascriptInterface
-    fun importIntoHome(kind: String, name: String): String {
-        if (!trusted()) return ""
-        val dir = when (kind) {
-            "identity" -> File(kernel.homeDir(), ".ssh")
-            "env" -> kernel.homeDir()
-            else -> return ""
-        }
-        val safeName = sanitise(name)
-        if (safeName.isEmpty()) return ""
-
-        val picked = pickFile() ?: return ""
-        return try {
-            dir.mkdirs()
-            val target = File(dir, safeName)
-            context.contentResolver.openInputStream(picked)?.use { input ->
-                target.outputStream().use { output -> input.copyTo(output) }
-            } ?: return ""
-            // A private key with group or world read is one OpenSSH refuses to
-            // use, and the refusal reads as "bad permissions" from a long way
-            // away. Setting it here is what makes the file usable at all.
-            if (kind == "identity") {
-                runCatching {
-                    target.setReadable(false, false)
-                    target.setReadable(true, true)
-                    target.setWritable(false, false)
-                    target.setWritable(true, true)
-                }
-            }
-            target.absolutePath
-        } catch (_: Exception) {
-            // A copy that failed leaves nothing behind; the caller sees "" and
-            // the person can try again.
-            ""
-        }
-    }
-
-    /**
-     * Open a URL in the device's browser.
-     *
-     * The page asks for this when a link points somewhere outside the kernel —
-     * a provider's console, a documentation page — and loading it in place would
-     * strand the person in a window with no address bar and no way back.
-     */
-    @JavascriptInterface
-    fun openExternal(url: String) {
-        if (!trusted()) return
-        // Only what a browser can open. A `file://` or `content://` here would
-        // be the page reaching into the device through a method named for
-        // leaving it.
-        if (!url.startsWith("http://") && !url.startsWith("https://")) return
-        onOpen(url)
-    }
-
-    /** Put text on the clipboard. A WebView's own `navigator.clipboard` write is
-     *  gated behind a permission Android grants only to a focused document, so
-     *  the one thing the page actually offers to copy — a token, a command, a
-     *  path — has no other way out.
-     *
-     *  Answers whether the write landed. A platform that refuses it raises
-     *  rather than answering, and an exception must not cross
-     *  `@JavascriptInterface` as a silent success — the page would say 已复制
-     *  over an empty clipboard. Caught here, the answer is the one the caller
-     *  can act on: false, and the button says 复制失败 instead. */
-    @JavascriptInterface
-    fun copyText(text: String): Boolean = if (!trusted()) false else runCatching {
-        val clipboard = context.getSystemService(Context.CLIPBOARD_SERVICE) as ClipboardManager
-        clipboard.setPrimaryClip(ClipData.newPlainText("reasonix", text))
-    }.isSuccess
-
-    /**
-     * Write text where the person picks, and answer with the path it went to.
-     *
-     * The page's own route to a file — a blob URL on an `<a download>` — reaches
-     * nothing here: a WebView has no downloads shelf, and its `DownloadListener`
-     * fires only for a real navigation, not for a blob the page just built. So
-     * the text crosses the bridge and the shell opens the system's document
-     * picker, which is also the only way to reach a folder the person can find
-     * again afterwards.
-     *
-     * "" is a dismissal, the same answer `importIntoHome` gives, and the page
-     * reads the two the same way.
-     */
-    /**
-     * Put a copy of a file where a phone's own file manager will show it.
-     *
-     * `saveText` asks the person where to put it, which is the right question
-     * on a desktop and the wrong one here: the kernel's workspace is
-     * app-private storage, so the file someone wants is in a place no file
-     * manager can open, and a picker that starts at "recent" is another place
-     * to lose it. This answers the actual question — "where is it" — by
-     * putting it in Downloads, in a folder named after the app, under a name
-     * that is the file's own.
-     *
-     * The answer is the path relative to shared storage, so the page can say
-     * where the file went rather than only that it worked. "" means the write
-     * failed or the platform refused it.
-     *
-     * MediaStore on Android 10 and later, a plain file before that. The two
-     * are not interchangeable: from 10 the app may not write another app's
-     * files directly, and MediaStore is the only way in; before it there is no
-     * MediaStore collection for downloads and the public directory is writable
-     * with the storage permission this app does not ask for, so that branch is
-     * best-effort and says so by answering "".
-     */
-    @JavascriptInterface
-    fun exportToDownloads(name: String, content: String): String {
-        if (!trusted()) return ""
-        val safeName = sanitise(name).ifEmpty { "reasonix.txt" }
-        return exportFile(safeName, content).orEmpty()
-    }
-
-    @JavascriptInterface
-    fun saveText(name: String, content: String): String {
-        if (!trusted()) return ""
-        val safeName = sanitise(name).ifEmpty { "reasonix.txt" }
-        return saveFile(safeName, content) ?: ""
-    }
-
-    /**
-     * Hand text to whatever the device shares with — a chat app, the mail
-     * client, a note app.
-     *
-     * The share sheet rather than a second save dialog: "where is that file"
-     * is the question this answers, and on a phone a document someone wants to
-     * keep or send usually goes through one of those apps rather than a folder.
-     *
-     * Shared as text with the filename as its subject, which is what every
-     * target understands. A file attachment would need a `FileProvider` and a
-     * cache copy for the same words to arrive.
-     */
-    @JavascriptInterface
-    fun shareText(name: String, content: String): Boolean {
-        if (!trusted()) return false
-        val safeName = sanitise(name).ifEmpty { "reasonix.txt" }
-        val posted = java.util.concurrent.CountDownLatch(1)
-        var ok = false
-        main.post {
-            ok = runCatching { askForShare(safeName, content) }.getOrDefault(false)
-            posted.countDown()
-        }
-        runCatching { posted.await(SHARE_TIMEOUT_SECONDS, TimeUnit.SECONDS) }
-        return ok
-    }
-
-    /**
-     * The device's own light/dark answer.
-     *
-     * The page resolves its theme from `localStorage` first and only asks the
-     * system when nothing is saved, so this is what makes a fresh install match
-     * the phone instead of defaulting to light.
-     */
-    @JavascriptInterface
-    fun systemTheme(): String {
-        val night = context.resources.configuration.uiMode and Configuration.UI_MODE_NIGHT_MASK
-        return if (night == Configuration.UI_MODE_NIGHT_YES) "dark" else "light"
-    }
-
-    /**
-     * The window's real safe-area insets, in CSS pixels, as
-     * `top,right,bottom,left`.
-     *
-     * A WebView does not implement `env(safe-area-inset-*)`: those are an iOS
-     * Safari feature, and on Android every one of them resolves to `0`. So a
-     * page that pads itself with them — which `mobile.css` does — is un-padded
-     * on exactly the platform this shell exists for, and the first line of the
-     * interface sits under the status bar.
-     *
-     * The values have to come from here for a second reason too: from Android 15
-     * the window is edge-to-edge whether the app asks for it or not, so the bars
-     * really are drawn over the page and there is no window-manager switch that
-     * could move the content back down.
-     *
-     * Answered as a string because that is what crosses the bridge; the page
-     * sets it as a CSS custom property and `mobile.css` reads it.
-     */
-    @JavascriptInterface
-    fun safeArea(): String {
-        val insets = onInsets?.invoke() ?: return "0,0,0,0"
-        return "${insets.left},${insets.top},${insets.right},${insets.bottom}"
-    }
-
-    /**
-     * Read the pairing code a computer is showing.
-     *
-     * A phone cannot type what a computer renders as a QR code — it is a
-     * one-time code inside a URL, and the whole reason it is a picture is that
-     * nobody reads one aloud. So the page asks for the camera and gets the text
-     * back, which is the same direction the computer's own dialog describes.
-     *
-     * The answer is the code's own text, not a parsed result: what a pairing
-     * fragment means is the page's to decide, and it already knows how — the
-     * `#pair=` handling ships with the interface.
-     */
-    @JavascriptInterface
-    fun scanPairingCode(): String = if (!trusted()) "" else scanWithCamera()
-
-    /**
-     * Load a URL in this window instead of handing it to the browser.
-     *
-     * The page asks for this once, right after a scan: the code it read points at
-     * the computer's own Studio, and the whole point of scanning is to end up
-     * looking at that computer. Going out to a browser would leave the app on one
-     * screen and the paired machine on another, with the back gesture an exit
-     * rather than a return.
-     *
-     * Only http and https are accepted. The page is this repository's own build
-     * and is trusted, but a scanner reads whatever is held in front of it, so the
-     * one value that arrives here from outside is checked: a `file://` or
-     * `javascript:` that happened to be on a printed code must not become a
-     * navigation.
-     */
-    @JavascriptInterface
-    fun openInApp(url: String) {
-        if (!trusted()) return
-        val trimmed = url.trim()
-        if (!trimmed.startsWith("http://") && !trimmed.startsWith("https://")) return
-        main.post { onOpenInApp?.invoke(trimmed) }
-    }
-
-    /** How the activity loads a URL in the window it owns. Set by the activity. */
-    var onOpenInApp: ((String) -> Unit)? = null
-
-    // ---- the native browser view host bridge ----
-    //
-    // The page's `host.ts` looks for `window.reasonixHost` with Promise-based
-    // methods (showBrowserView, hideBrowserView, etc.). `@JavascriptInterface`
-    // is synchronous, so the round trip is: JS generates an id, calls
-    // `hostCall(id, method, args)`, Kotlin does the work asynchronously, then
-    // pushes the result back via `evaluateJavascript` calling
-    // `window.__reasonixHostReply(id, json)`. The injected script (in
-    // MainActivity) wires each Promise to that reply channel.
-
-    /** How the activity handles an async host call. Set by the activity. */
-    var onHostCall: ((id: String, method: String, args: String) -> Unit)? = null
-
-    /**
-     * Entry point for the page's `window.reasonixHost` Promise methods.
-     *
-     * `id` is a unique call identifier the JS side generated; `method` is the
-     * method name (e.g. "showBrowserView"); `args` is the JSON-encoded argument
-     * array. The activity handles the call asynchronously and replies through
-     * `evaluateJavascript` with `window.__reasonixHostReply(id, resultJson)`.
-     */
-    @JavascriptInterface
-    fun hostCall(id: String, method: String, args: String) {
-        if (!trusted()) return
-        onHostCall?.invoke(id, method, args)
-    }
-
-    /**
-     * Whether the page has a panel up that the back gesture should close first.
-     *
-     * Written by the page, read by the activity's back handler. The page owns the
-     * answer — it is the only thing that knows what is open, and a second
-     * implementation of that question in the shell would be one that drifts — so
-     * the page reports the fact and the shell keeps the action.
-     *
-     * Volatile because the two ends are on different threads: the page sets it
-     * from the WebView's JavaScript thread, the back gesture reads it on the main
-     * one, and a stale read here is a panel that does not close.
-     */
-    @Volatile
-    var layerOpen: Boolean = false
-        private set
-
-    @JavascriptInterface
-    fun setLayerOpen(open: Boolean) {
-        layerOpen = open
-    }
-
-    // ---- the picker, which is an Activity result in a synchronous world ----
-
-    private val pending = AtomicReference<Uri?>(null)
-    private val latch = AtomicReference<CountDownLatch?>(null)
-
-    /** The code the camera read, held between the activity's answer and the
-     *  wait below — the same round trip the file picker takes. */
-    private val scanned = AtomicReference<String?>(null)
-    private val scanLatch = AtomicReference<CountDownLatch?>(null)
-
-    /** The main thread, because starting an Activity result must happen there. */
-    private val main = Handler(Looper.getMainLooper())
-
-    /**
-     * How to ask for a file. Set by the activity, which owns the Activity Result
-     * API; it answers whether the ask was dispatched, not whether the person
-     * picked anything — that arrives through `deliver`.
-     */
-    var askForFile: () -> Boolean = { false }
-
-    /**
-     * How the activity reports the window's insets. Set by the activity, which is
-     * the only thing that can see them, and asked on every call rather than
-     * cached — a rotation, a fold, or the keyboard opening all change them, and a
-     * stale answer here is a stripe of interface under the status bar.
-     */
-    var onInsets: (() -> Insets?)? = null
-
-    /**
-     * How to open the camera. Set by the activity, which owns the Activity
-     * Result API; like `askForFile` it answers whether the ask was dispatched,
-     * and the code itself arrives through `deliverScan`.
-     */
-    var askForScan: () -> Boolean = { false }
-
-    /**
-     * How to put a file where the person picks. Set by the activity, which owns
-     * the Activity Result API; it answers whether the ask was dispatched, and
-     * the chosen place arrives through `deliverSave`.
-     */
-    var askForSave: (name: String, content: String) -> Boolean = { _, _ -> false }
-
-    /**
-     * How to open the share sheet. Set by the activity, which is the only thing
-     * that can start an Intent; it answers whether anything on the device took
-     * the text.
-     */
-    var askForShare: (name: String, content: String) -> Boolean = { _, _ -> false }
-
-    /** The place the save dialog answered with, held between the two halves. */
-    private val saveTarget = AtomicReference<Uri?>(null)
-    private val saveLatch = AtomicReference<CountDownLatch?>(null)
-
-    /**
-     * Ask where to put a file and write it there.
-     *
-     * The same latch round trip as `pickFile`, and for the same reason:
-     * `@JavascriptInterface` cannot suspend, so the WebView's thread waits while
-     * the picker is open — a modal panel the person is looking at.
-     */
-    private fun saveFile(name: String, content: String): String? {
-        val gate = CountDownLatch(1)
-        saveLatch.set(gate)
-        saveTarget.set(null)
-
-        val posted = CountDownLatch(1)
-        var dispatched = false
-        main.post {
-            dispatched = runCatching { askForSave(name, content) }.getOrDefault(false)
-            posted.countDown()
-        }
-        val ready = runCatching { posted.await(PICK_TIMEOUT_SECONDS, TimeUnit.SECONDS) }.getOrDefault(false)
-        if (!ready || !dispatched) {
-            saveLatch.set(null)
-            return null
-        }
-        return try {
-            if (!gate.await(PICK_TIMEOUT_SECONDS, TimeUnit.SECONDS)) null
-            else saveTarget.get()?.toString()
-        } catch (_: InterruptedException) {
-            null
-        } finally {
-            saveLatch.set(null)
-        }
-    }
-
-    /** Called by the activity once the bytes are written. A null uri is a dismissal. */
-    fun deliverSave(uri: Uri?) {
-        saveTarget.set(uri)
-        saveLatch.get()?.countDown()
-    }
-
-    /**
-     * Read the pairing code a computer is showing, and return it.
-     *
-     * The same shape as `pickFile` and for the same reason: a
-     * `@JavascriptInterface` method is called on the WebView's thread and must
-     * return a value, so the camera is opened on the main thread and this one
-     * waits. The block lasts as long as the person is aiming, which is why the
-     * timeout is generous — a scan that gave up after a few seconds would be
-     * one nobody could complete.
-     *
-     * An empty string is "no code", which is what a refusal, a cancellation and
-     * a timeout all are. The caller acts on the text, so a distinct error would
-     * only be a fourth way to say the same thing.
-     */
-    private fun scanWithCamera(): String {
-        val gate = CountDownLatch(1)
-        scanLatch.set(gate)
-        scanned.set(null)
-
-        val posted = CountDownLatch(1)
-        var dispatched = false
-        main.post {
-            dispatched = runCatching { askForScan() }.getOrDefault(false)
-            posted.countDown()
-        }
-        val ready = runCatching { posted.await(SCAN_TIMEOUT_SECONDS, TimeUnit.SECONDS) }.getOrDefault(false)
-        if (!ready || !dispatched) {
-            scanLatch.set(null)
-            return ""
-        }
-        return try {
-            if (!gate.await(SCAN_TIMEOUT_SECONDS, TimeUnit.SECONDS)) "" else scanned.get().orEmpty()
-        } catch (_: InterruptedException) {
-            ""
-        } finally {
-            scanLatch.set(null)
-        }
-    }
-
-    /** Called by the activity when the scanner answers. Null is a dismissal. */
-    fun deliverScan(text: String?) {
-        scanned.set(text)
-        scanLatch.get()?.countDown()
-    }
-
-    /**
-     * Ask for one file and wait for the answer.
-     *
-     * `@JavascriptInterface` methods cannot suspend: the WebView calls them on
-     * its own thread and expects a value back. So the ask is posted to the main
-     * thread and this one waits on a latch. The cost is a blocked JS thread for
-     * as long as the picker is open — which is a modal panel the person is
-     * looking at, with nothing else in the page running.
-     *
-     * A timeout is set rather than waiting forever: a picker that never answers
-     * (the activity destroyed underneath it) would otherwise pin a WebView
-     * thread until the process died.
-     */
-    private fun pickFile(): Uri? {
-        val gate = CountDownLatch(1)
-        latch.set(gate)
-        pending.set(null)
-
-        val posted = CountDownLatch(1)
-        var dispatched = false
-        main.post {
-            dispatched = runCatching { askForFile() }.getOrDefault(false)
-            posted.countDown()
-        }
-        val ready = runCatching { posted.await(PICK_TIMEOUT_SECONDS, TimeUnit.SECONDS) }.getOrDefault(false)
-        if (!ready || !dispatched) {
-            latch.set(null)
-            return null
-        }
-
-        return try {
-            if (!gate.await(PICK_TIMEOUT_SECONDS, TimeUnit.SECONDS)) null else pending.get()
-        } catch (_: InterruptedException) {
-            null
-        } finally {
-            latch.set(null)
-        }
-    }
-
-    /** Called by the activity when the picker answers. A null uri is a dismissal. */
-    fun deliver(uri: Uri?) {
-        pending.set(uri)
-        latch.get()?.countDown()
-    }
-
-    /** The folder under Downloads that this app's exports land in. */
-    private fun exportFolder(): String = "Reasonix"
-
-    /** The MIME type this name declares, or null when the platform does not know
-     *  it.
-     *
-     * Where the type is written, the name and the type agree, and that agreement
-     * is the whole fix: `.md` is answered `text/markdown` here, so MediaStore
-     * has nothing to correct and the export stays `报告.md` instead of being
-     * renamed `报告.md.txt` by the `text/plain` it used to be given. Where the
-     * type is unknown the field is left off, which leaves MediaStore to derive
-     * the same `application/octet-stream` it would have been told — the name
-     * still stands on its own, which is all this is for. */
-    private fun mimeFor(name: String): String? {
-        val ext = name.substringAfterLast('.', "").lowercase()
-        if (ext.isEmpty() || ext == name.lowercase()) return null
-        return runCatching { MimeTypeMap.getSingleton().getMimeTypeFromExtension(ext) }.getOrNull()
-    }
-
-    private fun exportFile(name: String, content: String): String? = runCatching {
-        val bytes = content.toByteArray(Charsets.UTF_8)
-        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.Q) {
-            val resolver = context.contentResolver
-            val values = ContentValues().apply {
-                put(MediaStore.Downloads.DISPLAY_NAME, name)
-                // The type is derived here rather than declared, and it is
-                // written only when the name really has an extension that maps
-                // to one.
-                //
-                // It used to be `text/plain` for everything, and MediaStore
-                // answers a name whose extension disagrees with the declared
-                // type by appending its own: exporting `报告.md` landed in
-                // Downloads as `报告.md.txt`, and any source file came out with
-                // the wrong second extension. Deriving the type from the name
-                // instead keeps the two in agreement — which is what stops the
-                // rename — and a name the platform has no type for is left off
-                // entirely rather than declared as `application/octet-stream`,
-                // so nothing is corrected there either.
-                mimeFor(name)?.let { put(MediaStore.Downloads.MIME_TYPE, it) }
-                put(
-                    MediaStore.Downloads.RELATIVE_PATH,
-                    Environment.DIRECTORY_DOWNLOADS + "/" + exportFolder(),
-                )
-                // Hidden until the bytes are all there, so a file manager that
-                // reads the folder mid-write cannot show half a document.
-                put(MediaStore.Downloads.IS_PENDING, 1)
-            }
-            val uri = resolver.insert(MediaStore.Downloads.EXTERNAL_CONTENT_URI, values)
-                ?: return@runCatching null
-            resolver.openOutputStream(uri)?.use { it.write(bytes) } ?: return@runCatching null
-            values.clear()
-            values.put(MediaStore.Downloads.IS_PENDING, 0)
-            resolver.update(uri, values, null, null)
-            "${Environment.DIRECTORY_DOWNLOADS}/${exportFolder()}/$name"
-        } else {
-            val dir = File(
-                Environment.getExternalStoragePublicDirectory(Environment.DIRECTORY_DOWNLOADS),
-                exportFolder(),
-            )
-            if (!dir.exists() && !dir.mkdirs()) return@runCatching null
-            val target = File(dir, name)
-            target.writeBytes(bytes)
-            target.absolutePath
-        }
-    }.getOrNull()
-
-    private fun sanitise(name: String): String {
-        // A name from the page is used as a filename, so it is reduced to one:
-        // no separators, no traversal, and nothing a filesystem treats as
-        // special. The default is the name the field it came from suggests.
-        val trimmed = name.trim().substringAfterLast('/').substringAfterLast('\\')
-        if (trimmed.isEmpty() || trimmed == "." || trimmed == "..") return ""
-        return trimmed.filter { it.isLetterOrDigit() || it in "._-" }.take(64)
-    }
-
-    /**
-     * Whether a turn is running, so the screen can be kept awake for one.
-     *
-     * The page owns the answer — it is the only thing that knows if the kernel
-     * is working — and the shell owns the flag, because a window flag is not
-     * something a page can set. Nothing is asked at launch: a window that is
-     * merely open is one the person may be reading, and holding the screen on
-     * for that is holding it on for nothing.
-     */
-    @JavascriptInterface
-    fun setRunning(running: Boolean) {
-        onRunning?.invoke(running)
-    }
-
-    /** How the activity applies that flag. Set by the activity. */
-    var onRunning: ((Boolean) -> Unit)? = null
-
-    /**
-     * Whether the app may read every file, which is what lets the kernel open a
-     * folder on this phone as a workspace.
-     *
-     * Asked rather than assumed: the permission is granted by hand in system
-     * settings, so the page has to be able to say "not yet" rather than fail on
-     * the first `stat`. Answered `true` below API 30, where ordinary read
-     * permission is enough and the concept does not exist.
-     */
-    @JavascriptInterface
-    fun hasAllFilesAccess(): Boolean = filesAccess?.invoke() ?: true
-
-    var filesAccess: (() -> Boolean)? = null
-
-    /** Open the system page where that permission is granted. The only place
-     *  Android offers it — there is no dialog for this one. */
-    @JavascriptInterface
-    fun requestAllFilesAccess() {
-        if (!trusted()) return
-        askForFilesAccess?.invoke()
-    }
-
-    var askForFilesAccess: (() -> Unit)? = null
-
-    /**
-     * Whether this app is exempt from the system's battery optimisation.
-     *
-     * A turn runs in a child process under a foreground service. Under doze the
-     * whole app group is frozen anyway, so the turn stops making progress while
-     * the screen is off — which is the state the person left it in on purpose.
-     * Unrestricted is the answer, and it is a question Android answers.
-     */
-    @JavascriptInterface
-    fun batteryUnrestricted(): Boolean = if (!trusted()) false else (batteryFree?.invoke() ?: true)
-
-    var batteryFree: (() -> Boolean)? = null
-
-    /** Ask for that exemption. Unlike the storage one, this has a real dialog. */
-    @JavascriptInterface
-    fun requestBatteryUnrestricted() {
-        if (!trusted()) return
-        askForBattery?.invoke()
-    }
-
-    var askForBattery: (() -> Unit)? = null
-
-    /**
-     * Ask for a folder on this phone and answer with its real filesystem path.
-     *
-     * "" is a dismissal, the same answer `importIntoHome` gives — and, unlike
-     * that one, an empty string is a normal outcome here rather than a failure.
-     * The `startIn` hint is a path the kernel knows; the picker may ignore it,
-     * which is the picker's business.
-     *
-     * The round trip is the same latch as `pickFile`, for the same reason: a
-     * document-tree picker is another activity, and this call cannot suspend.
-     */
-    @JavascriptInterface
-    fun pickFolder(startIn: String): String {
-        if (!trusted()) return ""
-        val gate = CountDownLatch(1)
-        folderLatch.set(gate)
-        folderPending.set(null)
-
-        val posted = CountDownLatch(1)
-        var dispatched = false
-        main.post {
-            dispatched = runCatching { askForFolder(startIn) }.getOrDefault(false)
-            posted.countDown()
-        }
-        val ready = runCatching { posted.await(PICK_TIMEOUT_SECONDS, TimeUnit.SECONDS) }.getOrDefault(false)
-        if (!ready || !dispatched) {
-            folderLatch.set(null)
-            return ""
-        }
-        return try {
-            if (!gate.await(PICK_TIMEOUT_SECONDS, TimeUnit.SECONDS)) "" else folderPending.get().orEmpty()
-        } catch (_: InterruptedException) {
-            ""
-        } finally {
-            folderLatch.set(null)
-        }
-    }
-
-    var askForFolder: (String) -> Boolean = { false }
-
-    private val folderPending = AtomicReference<String?>(null)
-    private val folderLatch = AtomicReference<CountDownLatch?>(null)
-
-    /** Called by the activity with the folder's real path, or null on dismissal. */
-    fun deliverFolder(path: String?) {
-        folderPending.set(path ?: "")
-        folderLatch.get()?.countDown()
-    }
-
-    companion object {
-        const val BRIDGE_VERSION = "4"
-        private const val PICK_TIMEOUT_SECONDS = 5L * 60
-
-        /** Long, because this is however long someone takes to aim. A shorter
-         *  one would turn "still lining it up" into a failed scan. */
-        private const val SCAN_TIMEOUT_SECONDS = 5L * 60
-
-        /** Only as long as the sheet takes to appear: the answer is "something
-         *  took it", which is known the moment the Intent resolves. */
-        private const val SHARE_TIMEOUT_SECONDS = 15L
-    }
-}
-
-/**
- * The window's insets, in CSS pixels.
- *
- * A named type rather than four numbers because the page reads them as one
- * thing: half an inset is a layout that is wrong in one direction, and a caller
- * that passed the wrong pair would be indistinguishable from a device with no
- * cutout at all.
- */
-data class Insets(val left: Int, val top: Int, val right: Int, val bottom: Int)
@@ -1,66 +0,0 @@
-package io.reasonix.mobile.app
-
-import androidx.camera.core.ImageProxy
-import com.google.zxing.BarcodeFormat
-import com.google.zxing.BinaryBitmap
-import com.google.zxing.DecodeHintType
-import com.google.zxing.MultiFormatReader
-import com.google.zxing.PlanarYUVLuminanceSource
-import com.google.zxing.common.HybridBinarizer
-
-/**
- * Reading a QR code out of one camera frame.
- *
- * ZXing core and nothing else: it is pure Java, so it runs identically on a
- * phone with Google Play Services and on one whose HarmonyOS compatibility
- * layer has none. A decoder that leaned on ML Kit would work on the machine it
- * was written on and fail on the one that needs it.
- *
- * The frame arrives as YUV_420_888, whose Y plane is already the greyscale
- * bitmap ZXing wants — no conversion, no colour maths, just the bytes. That is
- * the whole reason a phone can decode 30 frames a second without warming up.
- */
-object Qr {
-
-    private val reader = MultiFormatReader().apply {
-        setHints(
-            mapOf(
-                // Only the form the computer actually renders. Leaving the
-                // reader open to every 1D barcode makes each frame slower for
-                // a case that cannot occur, and a slower frame is a code that
-                // has already left the viewfinder.
-                DecodeHintType.POSSIBLE_FORMATS to listOf(BarcodeFormat.QR_CODE),
-                DecodeHintType.TRY_HARDER to true,
-            ),
-        )
-    }
-
-    /** The code's text, or null when this frame holds none. */
-    fun decode(image: ImageProxy): String? {
-        val plane = image.planes.firstOrNull() ?: return null
-        val buffer = plane.buffer
-        val data = ByteArray(buffer.remaining())
-        buffer.get(data)
-        // `rowStride` is padded to a hardware alignment and is usually wider
-        // than the picture. Passing the width as the row length is what tells
-        // ZXing where each row actually ends; getting it wrong shears the image
-        // and every code reads as garbage.
-        val source = PlanarYUVLuminanceSource(
-            data,
-            plane.rowStride,
-            image.height,
-            0,
-            0,
-            image.width,
-            image.height,
-            false,
-        )
-        return runCatching {
-            reader.decodeWithState(BinaryBitmap(HybridBinarizer(source))).text
-        }.getOrNull().also {
-            // The reader keeps per-frame state between calls; resetting is what
-            // stops a partial match in one frame from corrupting the next.
-            reader.reset()
-        }
-    }
-}
@@ -1,146 +0,0 @@
-package io.reasonix.mobile.app
-
-import android.Manifest
-import android.content.Intent
-import android.content.pm.PackageManager
-import android.os.Bundle
-import android.view.Gravity
-import android.view.ViewGroup
-import android.widget.FrameLayout
-import android.widget.TextView
-import androidx.activity.result.contract.ActivityResultContracts
-import androidx.appcompat.app.AppCompatActivity
-import androidx.camera.core.CameraSelector
-import androidx.camera.core.ImageAnalysis
-import androidx.camera.core.ImageProxy
-import androidx.camera.core.Preview
-import androidx.camera.lifecycle.ProcessCameraProvider
-import androidx.camera.view.PreviewView
-import androidx.core.content.ContextCompat
-import java.util.concurrent.Executors
-
-/**
- * The camera, held open just long enough to read one pairing code.
- *
- * A phone that wants to drive a computer cannot type the code that computer
- * shows: it is a base64 fragment inside a URL, and the whole point of showing it
- * as a QR code is that nobody reads one aloud. So this is the way in — the same
- * gesture the computer's own dialog asks for, from the other side.
- *
- * The decode is ZXing rather than ML Kit on purpose. ML Kit needs Google Play
- * Services, and this app installs on phones whose HarmonyOS compatibility layer
- * has none: a GMS-backed scanner would work on a developer's phone and fail on
- * the device that needs it. ZXing core is pure Java, so the scanner behaves the
- * same on both.
- *
- * It returns the first code it reads and finishes. A pairing code is
- * single-use and short-lived, so "keep scanning after a hit" is not a feature —
- * it is a way to spend a code that a later frame would have discarded.
- */
-class ScanActivity : AppCompatActivity() {
-
-    private lateinit var preview: PreviewView
-    private lateinit var hint: TextView
-    private val executor = Executors.newSingleThreadExecutor()
-
-    /** Set once a code is found, so the frames in flight behind it do not each
-     *  try to finish the activity. */
-    @Volatile
-    private var settled = false
-
-    private val requestCamera = registerForActivityResult(ActivityResultContracts.RequestPermission()) { granted ->
-        if (granted) start() else refuse()
-    }
-
-    override fun onCreate(savedInstanceState: Bundle?) {
-        super.onCreate(savedInstanceState)
-        val root = FrameLayout(this)
-        preview = PreviewView(this).apply {
-            layoutParams = FrameLayout.LayoutParams(
-                ViewGroup.LayoutParams.MATCH_PARENT,
-                ViewGroup.LayoutParams.MATCH_PARENT,
-            )
-            // The frames are analysed, not composed: filling the view and
-            // cropping is right here and would be wrong for a camera app.
-            scaleType = PreviewView.ScaleType.FILL_CENTER
-        }
-        hint = TextView(this).apply {
-            text = getString(R.string.scan_hint)
-            setTextColor(0xFFFFFFFF.toInt())
-            textSize = 14f
-            gravity = Gravity.CENTER
-            val pad = (16 * resources.displayMetrics.density).toInt()
-            setPadding(pad, pad, pad, pad)
-            // Readable over whatever the camera is pointing at, which is the
-            // one background a design system cannot choose.
-            setBackgroundColor(0x99000000.toInt())
-            layoutParams = FrameLayout.LayoutParams(
-                ViewGroup.LayoutParams.MATCH_PARENT,
-                ViewGroup.LayoutParams.WRAP_CONTENT,
-                Gravity.BOTTOM,
-            )
-        }
-        root.addView(preview)
-        root.addView(hint)
-        setContentView(root)
-
-        if (ContextCompat.checkSelfPermission(this, Manifest.permission.CAMERA) ==
-            PackageManager.PERMISSION_GRANTED
-        ) {
-            start()
-        } else {
-            requestCamera.launch(Manifest.permission.CAMERA)
-        }
-    }
-
-    private fun refuse() {
-        // No dialog of our own: the permission prompt already said no, and a
-        // second window repeating it is one more tap to dismiss. The caller
-        // gets the same answer as a cancelled scan.
-        setResult(RESULT_CANCELED)
-        finish()
-    }
-
-    private fun start() {
-        val providerFuture = ProcessCameraProvider.getInstance(this)
-        providerFuture.addListener({
-            val provider = runCatching { providerFuture.get() }.getOrNull() ?: return@addListener refuse()
-            val previewUseCase = Preview.Builder().build().also { it.setSurfaceProvider(preview.surfaceProvider) }
-            val analysis = ImageAnalysis.Builder()
-                // Only the newest frame matters: a stale one describes where the
-                // camera was, and a code that has already left the viewfinder
-                // must not win over one that is in it now.
-                .setBackpressureStrategy(ImageAnalysis.STRATEGY_KEEP_ONLY_LATEST)
-                .build()
-                .also { it.setAnalyzer(executor, ::analyse) }
-            runCatching {
-                provider.unbindAll()
-                provider.bindToLifecycle(this, CameraSelector.DEFAULT_BACK_CAMERA, previewUseCase, analysis)
-            }.onFailure { refuse() }
-        }, ContextCompat.getMainExecutor(this))
-    }
-
-    private fun analyse(image: ImageProxy) {
-        try {
-            if (settled) return
-            val text = Qr.decode(image) ?: return
-            if (settled) return
-            settled = true
-            setResult(RESULT_OK, Intent().putExtra(EXTRA_TEXT, text))
-            // Finishing from the analysis thread is the whole reason `settled`
-            // exists; the activity teardown has to happen on the main one.
-            runOnUiThread { finish() }
-        } finally {
-            image.close()
-        }
-    }
-
-    override fun onDestroy() {
-        super.onDestroy()
-        executor.shutdown()
-    }
-
-    companion object {
-        const val EXTRA_TEXT = "text"
-    }
-}
@@ -1,18 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- The mark, drawn rather than shipped as a bitmap: one R in the app's own ink
-     on the palette's ground, which scales to every launcher's mask without a
-     density bucket per size. The hole in the R is the glyph's own counter, cut
-     with evenOdd rather than painted over. -->
-<vector xmlns:android="http://schemas.android.com/apk/res/android"
-    android:width="108dp"
-    android:height="108dp"
-    android:viewportWidth="108"
-    android:viewportHeight="108">
-    <path
-        android:fillColor="#F4F1EA"
-        android:fillType="evenOdd"
-        android:pathData="M39,30 L39,78 L47,78 L47,60 L54.5,60 L65,78 L74.5,78 L62.5,58 C68.6,55.8 72.6,50.6 72.6,43.6 C72.6,35.4 66.8,30 57.6,30 Z M47,37 L56.4,37 C61.2,37 64.4,39.4 64.4,43.8 C64.4,48.2 61.2,50.9 56.4,50.9 L47,50.9 Z" />
-    <path
-        android:fillColor="#A6680C"
-        android:pathData="M62,66 L74,66 L74,78 L62,78 Z" />
-</vector>
@@ -1,14 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- The status-bar glyph, not the launcher icon: Android renders a notification
-     icon as a mask, so anything that is not a solid silhouette on transparency
-     comes out as a grey blob. -->
-<vector xmlns:android="http://schemas.android.com/apk/res/android"
-    android:width="24dp"
-    android:height="24dp"
-    android:viewportWidth="24"
-    android:viewportHeight="24"
-    android:tint="#FFFFFF">
-    <path
-        android:fillColor="#FFFFFF"
-        android:pathData="M3,4h18v2H3zM3,9h18v2H3zM3,14h12v2H3zM3,19h12v2H3zM17.5,13.2l1.6,2.6 1.6,-2.6h2.3l-2.8,4.3V21h-2.2v-3.5L15.2,13.2z" />
-</vector>
@@ -1,62 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!--
-  The screen between the tap and the page.
-
-  It exists for a reason that is not cosmetic: the kernel has to be unpacked and
-  started before there is a page to show, and on a cold start that is about a
-  second. A blank window for a second reads as a broken app; a line of text that
-  says what is happening does not.
-
-  The progress bar is indeterminate and driven by nothing: the WebView's own
-  progress is mirrored onto it while the page loads, and before that it simply
-  turns, which is the honest description of "we do not know how long".
-->
-<LinearLayout xmlns:android="http://schemas.android.com/apk/res/android"
-    android:layout_width="match_parent"
-    android:layout_height="match_parent"
-    android:background="@color/page"
-    android:orientation="vertical"
-    android:gravity="center"
-    android:paddingStart="32dp"
-    android:paddingEnd="32dp">
-
-    <TextView
-        android:id="@+id/splash_wordmark"
-        android:layout_width="wrap_content"
-        android:layout_height="wrap_content"
-        android:text="@string/app_name"
-        android:textColor="#8A857B"
-        android:textSize="22sp"
-        android:letterSpacing="0.08"
-        android:fontFamily="sans-serif-medium" />
-
-    <ProgressBar
-        android:id="@+id/splash_bar"
-        style="?android:attr/progressBarStyleHorizontal"
-        android:layout_width="132dp"
-        android:layout_height="wrap_content"
-        android:layout_marginTop="26dp"
-        android:indeterminate="true" />
-
-    <TextView
-        android:id="@+id/splash_text"
-        android:layout_width="wrap_content"
-        android:layout_height="wrap_content"
-        android:layout_marginTop="18dp"
-        android:gravity="center"
-        android:text="@string/splash_connecting"
-        android:textColor="#8A857B"
-        android:textSize="13sp"
-        android:lineSpacingMultiplier="1.5" />
-
-    <Button
-        android:id="@+id/splash_retry"
-        style="?android:attr/borderlessButtonStyle"
-        android:layout_width="wrap_content"
-        android:layout_height="wrap_content"
-        android:layout_marginTop="14dp"
-        android:minHeight="44dp"
-        android:text="@string/splash_retry"
-        android:textColor="#A6680C"
-        android:visibility="gone" />
-</LinearLayout>
@@ -1,13 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- The adaptive launcher icon.
-
-     minSdk is 26, which is the release that introduced adaptive icons, so this
-     is the only launcher asset the app needs: there is no device running it
-     that would ask for a bitmap fallback. -->
-<adaptive-icon xmlns:android="http://schemas.android.com/apk/res/android">
-    <background android:drawable="@color/ic_launcher_background" />
-    <!-- The safe zone is the middle 66dp of 108dp; the glyph is drawn inside it
-         so no launcher's mask can clip it. -->
-    <foreground android:drawable="@drawable/ic_launcher_foreground" />
-    <monochrome android:drawable="@drawable/ic_launcher_foreground" />
-</adaptive-icon>
@@ -1,26 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<resources>
-    <string name="app_name">Reasonix</string>
-
-    <string name="channel_kernel">Kernel running</string>
-    <string name="channel_kernel_desc">Shown while the Reasonix kernel runs in the background. Can be muted on its own in system settings.</string>
-
-    <string name="notify_running_title">Reasonix is running</string>
-    <string name="notify_running_body">A task is still going. Switch back to see it.</string>
-    <string name="notify_failed_title">Reasonix could not start</string>
-
-    <string name="err_kernel_unreachable">Could not reach the kernel. Reopen the app from your home screen.</string>
-
-    <string name="download_default_name">reasonix-download</string>
-    <string name="download_saved">Saved: %1$s</string>
-    <string name="download_failed">Save failed: %1$s</string>
-    <string name="download_failed_status">Save failed: the server answered %1$d</string>
-
-    <string name="splash_connecting">Starting the kernel…</string>
-    <string name="splash_retry">Retry</string>
-
-    <string name="scan_title">Scan the code on the computer</string>
-    <string name="scan_hint">Point at the QR code in the computer\'s device panel to connect to it.</string>
-    <string name="scan_button">Scan to connect</string>
-    <string name="scan_denied">No camera permission, so scanning is unavailable. Allow it in system settings.</string>
-</resources>
@@ -1,7 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- The dark palette's page colour. `DayNight` on the theme picks this up when
-     the system is dark, which is what keeps the frame before the WebView from
-     flashing light at someone using a dark phone at night. -->
-<resources>
-    <color name="page">#0E1014</color>
-</resources>
@@ -1,9 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- The light palette's page colour, and only what the window needs before the
-     WebView paints. The interface's own colours are `tokens.css`'s and are the
-     page's business; these two exist so the first frame is the app's ground
-     rather than the platform default white. -->
-<resources>
-    <color name="page">#F4F1EA</color>
-    <color name="boot">#07080C</color>
-</resources>
@@ -1,6 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- The launcher ground: the interface's dark page colour. It is the same in
-     both themes because a launcher icon is not themed by the app. -->
-<resources>
-    <color name="ic_launcher_background">#14161C</color>
-</resources>
@@ -1,35 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- Every string the shell itself shows. The interface's own text belongs to
-     the page and lives in its i18n bundle; these are only the few lines that
-     exist before it, or when it cannot load. -->
-<resources>
-    <string name="app_name">Reasonix</string>
-
-    <string name="channel_kernel">内核运行中</string>
-    <string name="channel_kernel_desc">Reasonix 内核在后台运行时显示，可在系统设置中单独关闭。</string>
-
-    <string name="notify_running_title">Reasonix 正在运行</string>
-    <string name="notify_running_body">任务继续进行，切回应用即可查看。</string>
-    <string name="notify_failed_title">Reasonix 无法启动</string>
-
-    <string name="err_kernel_unreachable">连接内核失败。请返回桌面重新打开应用。</string>
-
-    <string name="download_default_name">reasonix-download</string>
-    <string name="download_saved">已保存：%1$s</string>
-    <string name="download_failed">保存失败：%1$s</string>
-    <string name="download_failed_status">保存失败：服务器返回 %1$d</string>
-
-    <string name="splash_connecting">正在启动内核…</string>
-    <string name="splash_retry">重试</string>
-
-    <string name="wsperm_title">需要「所有文件访问」权限</string>
-    <string name="wsperm_body">要打开手机上的文件夹作为工作区，需要先允许本应用访问所有文件。点「去设置」，在列表里打开 Reasonix 的开关，然后返回这里。</string>
-    <string name="wsperm_go">去设置</string>
-    <string name="wsperm_later">以后再说</string>
-
-    <string name="scan_title">扫描电脑上的二维码</string>
-    <string name="scan_hint">对准电脑上「设备访问」里的二维码，识别后即可连接这台电脑。</string>
-    <string name="scan_button">扫码连接电脑</string>
-    <string name="scan_denied">没有相机权限，无法扫码。请在系统设置里允许。</string>
-    <string name="share_title">分享文件</string>
-</resources>
@@ -1,28 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- The shell's own furniture: a window background matching the page's light
-     palette before the WebView paints, and a transparent system bar so the
-     page's own safe-area padding is what the notch meets. The page themes
-     itself; this exists only so the first frame is not a white flash. -->
-<resources xmlns:tools="http://schemas.android.com/tools">
-
-    <style name="Theme.Reasonix" parent="Theme.AppCompat.DayNight.NoActionBar">
-        <item name="android:windowBackground">@color/page</item>
-        <item name="android:statusBarColor">@android:color/transparent</item>
-        <item name="android:navigationBarColor">@android:color/transparent</item>
-        <!-- The page draws its own edge-to-edge layout and pads itself with the
-             safe-area insets from mobile.css. Laying the window out edge to edge
-             while leaving the bars transparent is what makes those insets real
-             numbers instead of zero. -->
-        <item name="android:windowLayoutInDisplayCutoutMode" tools:targetApi="p">shortEdges</item>
-        <item name="android:enforceStatusBarContrast" tools:targetApi="q">false</item>
-        <item name="android:enforceNavigationBarContrast" tools:targetApi="q">false</item>
-    </style>
-
-    <!-- The camera. A black ground because the preview has not drawn yet and a
-         light window behind a dark viewfinder is a flash in the dark; the bars
-         stay transparent so the preview reaches the edges, which is what makes
-         a code near the top of the screen scannable. -->
-    <style name="Theme.Reasonix.Scan" parent="Theme.Reasonix">
-        <item name="android:windowBackground">@android:color/black</item>
-    </style>
-</resources>
@@ -1,31 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!--
-  Backups are off, and this file says why rather than leaving it implicit.
-
-  A Reasonix home holds provider keys, an account token and every session
-  transcript. Android's auto-backup would copy all of it to the person's cloud
-  drive and, on a device-to-device transfer, to a new phone — which is a
-  credential leaving the device by a route nobody chose. `allowBackup="false"`
-  in the manifest turns that off; these two stanzas are the belt to its braces,
-  excluding everything explicitly so that a later change to `allowBackup` cannot
-  quietly start copying credentials.
-
-  What a person wants to move between phones, the kernel already moves: its own
-  export writes a file the person hands to the system's picker.
-->
-<data-extraction-rules>
-    <cloud-backup>
-        <exclude domain="root" />
-        <exclude domain="file" />
-        <exclude domain="database" />
-        <exclude domain="sharedpref" />
-        <exclude domain="external" />
-    </cloud-backup>
-    <device-transfer>
-        <exclude domain="root" />
-        <exclude domain="file" />
-        <exclude domain="database" />
-        <exclude domain="sharedpref" />
-        <exclude domain="external" />
-    </device-transfer>
-</data-extraction-rules>
@@ -1,27 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!--
-  Cleartext HTTP is permitted, and the platform has no way to permit less.
-
-  Two kinds of page load over plain HTTP: the kernel's own page on
-  `127.0.0.1:<port>`, and the same page served by a paired computer on whatever
-  address its router handed out. The port is chosen when the kernel starts and
-  the LAN address when a code is scanned, so neither can be written here — and a
-  `domain-config` matches literals and domains, never a subnet.
-
-  So the boundary is not this file. `MainActivity` decides which cleartext origin
-  may become the window: `shouldOverrideUrlLoading` sends every http(s) origin
-  that is not the kernel's or the one just paired out to the browser, and the
-  native bridge answers only those two. A plaintext page on the LAN therefore
-  cannot become the document holding the bridge, which is what the reviewer's
-  "deliberate path for LAN pairing" asks for.
-
-  HTTPS is untouched: nothing here downgrades a request, and `trust-anchors`
-  keeps the system roots for every TLS connection that is made.
-->
-<network-security-config>
-    <base-config cleartextTrafficPermitted="true">
-        <trust-anchors>
-            <certificates src="system" />
-        </trust-anchors>
-    </base-config>
-</network-security-config>
@@ -1,15 +0,0 @@
-// The build itself lives in app/build.gradle; this file declares the plugin
-// versions once so the app module can be read on its own.
-//
-// AGP 8.11.1 is the pairing for the Gradle 8.14 toolchain this machine has:
-// AGP 8.7 and older refuse a Gradle that new, and AGP 8.13+ wants a Gradle this
-// one does not have. Both numbers are pinned rather than ranged, because a
-// build that silently moved either one would be a different build.
-plugins {
-    id "com.android.application" version "8.11.1" apply false
-    id "org.jetbrains.kotlin.android" version "2.1.20" apply false
-}
-
-tasks.register("clean", Delete) {
-    delete rootProject.layout.buildDirectory
-}
@@ -1,15 +0,0 @@
-# AndroidX is the only copy of the support libraries this project uses; the
-# non-transitive R class keeps the app's own resources from aliasing library
-# ones, which is what makes `R.string.*` unambiguous here.
-android.useAndroidX=true
-android.nonTransitiveRClass=true
-
-# The daemon is what keeps a second `gradle assembleDebug` from re-reading the
-# whole SDK, and the heap is sized for the R8 pass over the WebView and the
-# packaged page rather than for this handful of Kotlin files.
-org.gradle.jvmargs=-Xmx3g -Dfile.encoding=UTF-8
-org.gradle.daemon=true
-org.gradle.parallel=true
-org.gradle.caching=true
-
-kotlin.code.style=official
@@ -1,12 +0,0 @@
-distributionBase=GRADLE_USER_HOME
-distributionPath=wrapper/dists
-distributionUrl=https\://services.gradle.org/distributions/gradle-8.14.3-bin.zip
-# Pinning the distribution's checksum is what makes the download verifiable: a
-# mirror that served a different zip would fail here instead of being unpacked
-# and run. The value is Gradle's own, from
-# https://services.gradle.org/distributions/gradle-8.14.3-bin.zip.sha256
-distributionSha256Sum=bd71102213493060956ec229d946beee57158dbd89d0e62b91bca0fa2c5f3531
-networkTimeout=10000
-validateDistributionUrl=true
-zipStoreBase=GRADLE_USER_HOME
-zipStorePath=wrapper/dists
@@ -1,251 +0,0 @@
-#!/bin/sh
-
-#
-# Copyright © 2015-2021 the original authors.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#      https://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-#
-
-##############################################################################
-#
-#   Gradle start up script for POSIX generated by Gradle.
-#
-#   Important for running:
-#
-#   (1) You need a POSIX-compliant shell to run this script. If your /bin/sh is
-#       noncompliant, but you have some other compliant shell such as ksh or
-#       bash, then to run this script, type that shell name before the whole
-#       command line, like:
-#
-#           ksh Gradle
-#
-#       Busybox and similar reduced shells will NOT work, because this script
-#       requires all of these POSIX shell features:
-#         * functions;
-#         * expansions «$var», «${var}», «${var:-default}», «${var+SET}»,
-#           «${var#prefix}», «${var%suffix}», and «$( cmd )»;
-#         * compound commands having a testable exit status, especially «case»;
-#         * various built-in commands including «command», «set», and «ulimit».
-#
-#   Important for patching:
-#
-#   (2) This script targets any POSIX shell, so it avoids extensions provided
-#       by Bash, Ksh, etc; in particular arrays are avoided.
-#
-#       The "traditional" practice of packing multiple parameters into a
-#       space-separated string is a well documented source of bugs and security
-#       problems, so this is (mostly) avoided, by progressively accumulating
-#       options in "$@", and eventually passing that to Java.
-#
-#       Where the inherited environment variables (DEFAULT_JVM_OPTS, JAVA_OPTS,
-#       and GRADLE_OPTS) rely on word-splitting, this is performed explicitly;
-#       see the in-line comments for details.
-#
-#       There are tweaks for specific operating systems such as AIX, CygWin,
-#       Darwin, MinGW, and NonStop.
-#
-#   (3) This script is generated from the Groovy template
-#       https://github.com/gradle/gradle/blob/HEAD/platforms/jvm/plugins-application/src/main/resources/org/gradle/api/internal/plugins/unixStartScript.txt
-#       within the Gradle project.
-#
-#       You can find Gradle at https://github.com/gradle/gradle/.
-#
-##############################################################################
-
-# Attempt to set APP_HOME
-
-# Resolve links: $0 may be a link
-app_path=$0
-
-# Need this for daisy-chained symlinks.
-while
-    APP_HOME=${app_path%"${app_path##*/}"}  # leaves a trailing /; empty if no leading path
-    [ -h "$app_path" ]
-do
-    ls=$( ls -ld "$app_path" )
-    link=${ls#*' -> '}
-    case $link in             #(
-      /*)   app_path=$link ;; #(
-      *)    app_path=$APP_HOME$link ;;
-    esac
-done
-
-# This is normally unused
-# shellcheck disable=SC2034
-APP_BASE_NAME=${0##*/}
-# Discard cd standard output in case $CDPATH is set (https://github.com/gradle/gradle/issues/25036)
-APP_HOME=$( cd -P "${APP_HOME:-./}" > /dev/null && printf '%s\n' "$PWD" ) || exit
-
-# Use the maximum available, or set MAX_FD != -1 to use that value.
-MAX_FD=maximum
-
-warn () {
-    echo "$*"
-} >&2
-
-die () {
-    echo
-    echo "$*"
-    echo
-    exit 1
-} >&2
-
-# OS specific support (must be 'true' or 'false').
-cygwin=false
-msys=false
-darwin=false
-nonstop=false
-case "$( uname )" in                #(
-  CYGWIN* )         cygwin=true  ;; #(
-  Darwin* )         darwin=true  ;; #(
-  MSYS* | MINGW* )  msys=true    ;; #(
-  NONSTOP* )        nonstop=true ;;
-esac
-
-CLASSPATH="\\\"\\\""
-
-
-# Determine the Java command to use to start the JVM.
-if [ -n "$JAVA_HOME" ] ; then
-    if [ -x "$JAVA_HOME/jre/sh/java" ] ; then
-        # IBM's JDK on AIX uses strange locations for the executables
-        JAVACMD=$JAVA_HOME/jre/sh/java
-    else
-        JAVACMD=$JAVA_HOME/bin/java
-    fi
-    if [ ! -x "$JAVACMD" ] ; then
-        die "ERROR: JAVA_HOME is set to an invalid directory: $JAVA_HOME
-
-Please set the JAVA_HOME variable in your environment to match the
-location of your Java installation."
-    fi
-else
-    JAVACMD=java
-    if ! command -v java >/dev/null 2>&1
-    then
-        die "ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
-
-Please set the JAVA_HOME variable in your environment to match the
-location of your Java installation."
-    fi
-fi
-
-# Increase the maximum file descriptors if we can.
-if ! "$cygwin" && ! "$darwin" && ! "$nonstop" ; then
-    case $MAX_FD in #(
-      max*)
-        # In POSIX sh, ulimit -H is undefined. That's why the result is checked to see if it worked.
-        # shellcheck disable=SC2039,SC3045
-        MAX_FD=$( ulimit -H -n ) ||
-            warn "Could not query maximum file descriptor limit"
-    esac
-    case $MAX_FD in  #(
-      '' | soft) :;; #(
-      *)
-        # In POSIX sh, ulimit -n is undefined. That's why the result is checked to see if it worked.
-        # shellcheck disable=SC2039,SC3045
-        ulimit -n "$MAX_FD" ||
-            warn "Could not set maximum file descriptor limit to $MAX_FD"
-    esac
-fi
-
-# Collect all arguments for the java command, stacking in reverse order:
-#   * args from the command line
-#   * the main class name
-#   * -classpath
-#   * -D...appname settings
-#   * --module-path (only if needed)
-#   * DEFAULT_JVM_OPTS, JAVA_OPTS, and GRADLE_OPTS environment variables.
-
-# For Cygwin or MSYS, switch paths to Windows format before running java
-if "$cygwin" || "$msys" ; then
-    APP_HOME=$( cygpath --path --mixed "$APP_HOME" )
-    CLASSPATH=$( cygpath --path --mixed "$CLASSPATH" )
-
-    JAVACMD=$( cygpath --unix "$JAVACMD" )
-
-    # Now convert the arguments - kludge to limit ourselves to /bin/sh
-    for arg do
-        if
-            case $arg in                                #(
-              -*)   false ;;                            # don't mess with options #(
-              /?*)  t=${arg#/} t=/${t%%/*}              # looks like a POSIX filepath
-                    [ -e "$t" ] ;;                      #(
-              *)    false ;;
-            esac
-        then
-            arg=$( cygpath --path --ignore --mixed "$arg" )
-        fi
-        # Roll the args list around exactly as many times as the number of
-        # args, so each arg winds up back in the position where it started, but
-        # possibly modified.
-        #
-        # NB: a `for` loop captures its iteration list before it begins, so
-        # changing the positional parameters here affects neither the number of
-        # iterations, nor the values presented in `arg`.
-        shift                   # remove old arg
-        set -- "$@" "$arg"      # push replacement arg
-    done
-fi
-
-
-# Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
-DEFAULT_JVM_OPTS='"-Xmx64m" "-Xms64m"'
-
-# Collect all arguments for the java command:
-#   * DEFAULT_JVM_OPTS, JAVA_OPTS, and optsEnvironmentVar are not allowed to contain shell fragments,
-#     and any embedded shellness will be escaped.
-#   * For example: A user cannot expect ${Hostname} to be expanded, as it is an environment variable and will be
-#     treated as '${Hostname}' itself on the command line.
-
-set -- \
-        "-Dorg.gradle.appname=$APP_BASE_NAME" \
-        -classpath "$CLASSPATH" \
-        -jar "$APP_HOME/gradle/wrapper/gradle-wrapper.jar" \
-        "$@"
-
-# Stop when "xargs" is not available.
-if ! command -v xargs >/dev/null 2>&1
-then
-    die "xargs is not available"
-fi
-
-# Use "xargs" to parse quoted args.
-#
-# With -n1 it outputs one arg per line, with the quotes and backslashes removed.
-#
-# In Bash we could simply go:
-#
-#   readarray ARGS < <( xargs -n1 <<<"$var" ) &&
-#   set -- "${ARGS[@]}" "$@"
-#
-# but POSIX shell has neither arrays nor command substitution, so instead we
-# post-process each arg (as a line of input to sed) to backslash-escape any
-# character that might be a shell metacharacter, then use eval to reverse
-# that process (while maintaining the separation between arguments), and wrap
-# the whole thing up as a single "set" statement.
-#
-# This will of course break if any of these variables contains a newline or
-# an unmatched quote.
-#
-
-eval "set -- $(
-        printf '%s\n' "$DEFAULT_JVM_OPTS $JAVA_OPTS $GRADLE_OPTS" |
-        xargs -n1 |
-        sed ' s~[^-[:alnum:]+,./:=@_]~\\&~g; ' |
-        tr '\n' ' '
-    )" '"$@"'
-
-exec "$JAVACMD" "$@"
@@ -1,94 +0,0 @@
-@rem
-@rem Copyright 2015 the original author or authors.
-@rem
-@rem Licensed under the Apache License, Version 2.0 (the "License");
-@rem you may not use this file except in compliance with the License.
-@rem You may obtain a copy of the License at
-@rem
-@rem      https://www.apache.org/licenses/LICENSE-2.0
-@rem
-@rem Unless required by applicable law or agreed to in writing, software
-@rem distributed under the License is distributed on an "AS IS" BASIS,
-@rem WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-@rem See the License for the specific language governing permissions and
-@rem limitations under the License.
-@rem
-@rem SPDX-License-Identifier: Apache-2.0
-@rem
-
-@if "%DEBUG%"=="" @echo off
-@rem ##########################################################################
-@rem
-@rem  Gradle startup script for Windows
-@rem
-@rem ##########################################################################
-
-@rem Set local scope for the variables with windows NT shell
-if "%OS%"=="Windows_NT" setlocal
-
-set DIRNAME=%~dp0
-if "%DIRNAME%"=="" set DIRNAME=.
-@rem This is normally unused
-set APP_BASE_NAME=%~n0
-set APP_HOME=%DIRNAME%
-
-@rem Resolve any "." and ".." in APP_HOME to make it shorter.
-for %%i in ("%APP_HOME%") do set APP_HOME=%%~fi
-
-@rem Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
-set DEFAULT_JVM_OPTS="-Xmx64m" "-Xms64m"
-
-@rem Find java.exe
-if defined JAVA_HOME goto findJavaFromJavaHome
-
-set JAVA_EXE=java.exe
-%JAVA_EXE% -version >NUL 2>&1
-if %ERRORLEVEL% equ 0 goto execute
-
-echo. 1>&2
-echo ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH. 1>&2
-echo. 1>&2
-echo Please set the JAVA_HOME variable in your environment to match the 1>&2
-echo location of your Java installation. 1>&2
-
-goto fail
-
-:findJavaFromJavaHome
-set JAVA_HOME=%JAVA_HOME:"=%
-set JAVA_EXE=%JAVA_HOME%/bin/java.exe
-
-if exist "%JAVA_EXE%" goto execute
-
-echo. 1>&2
-echo ERROR: JAVA_HOME is set to an invalid directory: %JAVA_HOME% 1>&2
-echo. 1>&2
-echo Please set the JAVA_HOME variable in your environment to match the 1>&2
-echo location of your Java installation. 1>&2
-
-goto fail
-
-:execute
-@rem Setup the command line
-
-set CLASSPATH=
-
-
-@rem Execute Gradle
-"%JAVA_EXE%" %DEFAULT_JVM_OPTS% %JAVA_OPTS% %GRADLE_OPTS% "-Dorg.gradle.appname=%APP_BASE_NAME%" -classpath "%CLASSPATH%" -jar "%APP_HOME%\gradle\wrapper\gradle-wrapper.jar" %*
-
-:end
-@rem End local scope for the variables with windows NT shell
-if %ERRORLEVEL% equ 0 goto mainEnd
-
-:fail
-rem Set variable GRADLE_EXIT_CONSOLE if you need the _script_ return code instead of
-rem the _cmd.exe /c_ return code!
-set EXIT_CODE=%ERRORLEVEL%
-if %EXIT_CODE% equ 0 set EXIT_CODE=1
-if not ""=="%GRADLE_EXIT_CONSOLE%" exit %EXIT_CODE%
-exit /b %EXIT_CODE%
-
-:mainEnd
-if "%OS%"=="Windows_NT" endlocal
-
-:omega
@@ -1,64 +0,0 @@
-#!/usr/bin/env bash
-#
-# Build the Reasonix kernel for Android and place it where the APK packager
-# expects a native library.
-#
-# Why a shell script and not a Gradle task: the kernel is a Go module that lives
-# outside this Gradle build, with its own toolchain version pinned in go.mod. A
-# Gradle task that invoked `go build` would make the Android build depend on a Go
-# toolchain being configured for it, and would rebuild a 21MB binary on every
-# `assembleDebug` even when nothing in the kernel changed. This runs when you
-# mean it to, and the APK task only checks that the result is there.
-#
-# Output: app/src/main/jniLibs/arm64-v8a/libreasonix.so
-#   The name is not cosmetic. Android extracts files under `jniLibs/` to
-#   `nativeLibraryDir` and permits execution *there*; a binary shipped in
-#   `assets/` would have to be copied to app storage and chmod'd, which fails on
-#   a device whose data partition is mounted noexec. `lib*.so` is the shape the
-#   packager recognises as a native library, so the kernel wears it.
-set -euo pipefail
-
-here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-android="$here/.."
-upstream="$android/.."
-out="$android/app/src/main/jniLibs/arm64-v8a/libreasonix.so"
-
-if [ ! -d "$upstream" ]; then
-  echo "kernel source not found: $upstream" >&2
-  echo "run this from a checkout: the parent of android/ must be the kernel source." >&2
-  exit 1
-fi
-
-# GOOS=android needs nothing else: the kernel is CGO-free, which is what makes a
-# single static binary possible in the first place. `-s -w` drops the symbol and
-# DWARF tables — about a third of the file — and nothing on a phone reads them.
-export GOOS=android
-export GOARCH=arm64
-export CGO_ENABLED=0
-
-# The app version, read from the module's own build.gradle so the two cannot
-# drift. The official build stamps these through `-X` (see upstream's Makefile);
-# without them the kernel reports `version=dev revision=unknown`, which a phone
-# then shows next to the documentation it was built with and makes a bug report
-# from that build impossible to place. There is no `.git` here, so the revision
-# names the package the corpus shipped in rather than a commit hash — the honest
-# answer when there is no commit to name.
-app_gradle="$android/app/build.gradle"
-version_field() {
-  sed -n "s/^def $1 *= *\([0-9]*\).*/\1/p" "$app_gradle" | head -n1
-}
-app_version="${REASONIX_APP_VERSION:-$(version_field versionMajor).$(version_field versionMinor).$(version_field versionPatch)}"
-build_time_utc="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
-ldflags="-s -w"
-ldflags="$ldflags -X main.version=v$app_version"
-ldflags="$ldflags -X main.buildTimeUTC=$build_time_utc"
-ldflags="$ldflags -X reasonix/internal/tools/productdocs.linkedVersion=v$app_version"
-ldflags="$ldflags -X reasonix/internal/tools/productdocs.linkedRevision=apk-v$app_version"
-
-echo "building the kernel (android/arm64, v$app_version) ..."
-mkdir -p "$(dirname "$out")"
-cd "$upstream"
-go build -trimpath -ldflags="$ldflags" -o "$out" ./cmd/reasonix
-
-size=$(wc -c < "$out")
-echo "kernel ready: $out ($((size / 1024 / 1024)) MiB)"
@@ -1,45 +0,0 @@
-#!/usr/bin/env bash
-#
-# One command from a clean checkout to an installable APK.
-#
-# The order matters and is the whole content of this script: the kernel and the
-# page are inputs to packaging, not build steps inside it, so they are produced
-# first and the Gradle task only assembles what is already there. Running
-# `gradle assembleRelease` directly on a fresh checkout fails on purpose (see
-# `requireBuildInputs` in app/build.gradle) rather than producing an APK with no
-# kernel in it.
-#
-# Usage: ./scripts/build.sh [debug|release]     (default: release)
-set -euo pipefail
-
-here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-android="$here/.."
-variant="${1:-release}"
-
-case "$variant" in
-  debug|release) ;;
-  *) echo "usage: $0 [debug|release]" >&2; exit 2 ;;
-esac
-
-"$here/build-kernel.sh"
-"$here/sync-page.sh"
-
-cd "$android"
-
-# The wrapper is committed, jar and all, and it is the only thing that runs the
-# build. Generating one here would tie the dependency versions to whichever
-# `gradle` happened to be on PATH, which is the opposite of a pinned build.
-if [ ! -f gradlew ]; then
-  echo "gradlew is missing: this repository commits the Gradle wrapper, so the checkout is incomplete" >&2
-  echo "(a release build does not use `gradle` from PATH: that would tie dependency versions to the machine)" >&2
-  exit 1
-fi
-
-task="assemble$([ "$variant" = release ] && echo Release || echo Debug)"
-echo "packaging the APK ($variant) ..."
-./gradlew "$task" --no-daemon
-
-apk_dir="$android/app/build/outputs/apk/$variant"
-echo
-echo "done. Installable artifacts:"
-find "$apk_dir" -name '*.apk' -maxdepth 1 -print -exec ls -lh {} \;
@@ -1,43 +0,0 @@
-#!/usr/bin/env bash
-#
-# Install the built APK on a connected phone and show its log.
-#
-# `adb` comes from the Android SDK rather than this project, because the SDK is
-# already how the app is built. Point ANDROID_HOME at it, or have `adb` on PATH.
-#
-# Usage: ./scripts/run.sh [debug|release]
-set -euo pipefail
-
-here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-android="$here/.."
-variant="${1:-release}"
-
-adb="${ADB:-${ANDROID_HOME:-$HOME/AndroidSDK}/platform-tools/adb}"
-if ! command -v "$adb" >/dev/null 2>&1; then
-  adb="adb"
-fi
-
-apk_dir="$android/app/build/outputs/apk/$variant"
-apk=$(find "$apk_dir" -name '*.apk' -maxdepth 1 2>/dev/null | head -n1)
-if [ -z "$apk" ]; then
-  echo "no $variant APK found; run ./scripts/build.sh $variant first" >&2
-  exit 1
-fi
-
-echo "installing $apk ..."
-"$adb" install -r "$apk"
-
-# The device count is checked before the app is started so an emulator-less
-# machine gets "no devices" rather than a launch failure that reads like a bug.
-if [ "$("$adb" devices | grep -c 'device$')" -eq 0 ]; then
-  echo "no device attached; skipping the launch." >&2
-  exit 0
-fi
-
-package="io.reasonix.mobile.app"
-[ "$variant" = debug ] && package="io.reasonix.mobile.app.debug"
-"$adb" shell monkey -p "$package" -c android.intent.category.LAUNCHER 1 >/dev/null
-
-echo
-echo "following the log (Ctrl-C to stop) ..."
-"$adb" logcat -v color "ReasonixKernel:V" "AndroidRuntime:E" "*:S"
@@ -1,64 +0,0 @@
-#!/usr/bin/env bash
-#
-# Build the interface and copy it where the APK will carry it.
-#
-# The page that ships is the same build the desktop Studio serves — that is the
-# whole reason the mobile layout lives in `desktop/frontend-next` rather than in
-# an Android project's own web root. Nothing is edited on the way through: the
-# only thing this script does that `pnpm build` does not is copy `dist/` into the
-# app's assets.
-#
-# Output: app/src/main/assets/page/
-set -euo pipefail
-
-here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-android="$here/.."
-frontend="$android/../desktop/frontend-next"
-dest="$android/app/src/main/assets/page"
-
-if [ ! -d "$frontend" ]; then
-  echo "frontend project not found: $frontend" >&2
-  exit 1
-fi
-
-cd "$frontend"
-
-# pnpm is the package manager the lockfile is written for; npm would resolve a
-# different tree than the one that was tested.
-if ! command -v pnpm >/dev/null 2>&1; then
-  echo "pnpm 10 is required: npm install -g pnpm@10" >&2
-  exit 1
-fi
-
-if [ ! -d node_modules ]; then
-  echo "installing frontend dependencies ..."
-  pnpm install --frozen-lockfile
-fi
-
-echo "building the frontend ..."
-pnpm build
-
-rm -rf "$dest"
-mkdir -p "$dest"
-cp -R dist/. "$dest/"
-
-if [ ! -f "$dest/index.html" ]; then
-  echo "the build produced no index.html; the frontend build probably failed." >&2
-  exit 1
-fi
-
-count=$(find "$dest" -type f | wc -l)
-size=$(du -sk "$dest" | cut -f1)
-echo "page ready: $dest ($count files, $((size / 1024)) MiB)"
-
-# MIT requires the copyright notice and the licence text to travel with a binary
-# distribution, so the APK carries it. Generated from the repository's own
-# `LICENSE` rather than committed a second time: a checked-in copy is one more
-# thing that drifts when the root file is updated.
-root_license="$android/../LICENSE"
-if [ ! -f "$root_license" ]; then
-  echo "LICENSE not found: $root_license" >&2
-  exit 1
-fi
-cp "$root_license" "$android/app/src/main/assets/LICENSE"
-echo "licence: app/src/main/assets/LICENSE (copied from the repository's LICENSE)"
@@ -1,95 +0,0 @@
-#!/usr/bin/env bash
-#
-# Reproduce what `Kernel.kt` does, on this machine, against the page that is
-# actually packaged into the APK.
-#
-# This is the closest thing to an end-to-end test that runs without a phone: the
-# launch flags, the token file, the port file and the `--page` directory are the
-# same ones the Kotlin shell passes, and the page is read out of
-# `app/src/main/assets/page` rather than the frontend's own `dist/`. If this
-# serves a working interface, the only untested part is Android itself.
-set -euo pipefail
-
-here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-android="$here/.."
-page="$android/app/src/main/assets/page"
-kernel="${KERNEL_BIN:-$android/../reasonix-verify.exe}"
-
-work="${VERIFY_DIR:-/tmp/reasonix-verify}"
-rm -rf "$work"
-mkdir -p "$work/home/remote"
-
-if [ ! -f "$kernel" ]; then
-  echo "no runnable kernel; build one first:" >&2
-  echo "  (cd .. && go build -o reasonix-verify.exe ./cmd/reasonix)" >&2
-  exit 1
-fi
-
-if [ ! -f "$page/index.html" ]; then
-  echo "no packaged page: $page/index.html" >&2
-  exit 1
-fi
-
-# 32 random bytes as hex — the same shape `Kernel.mintToken` writes.
-token=$(od -An -tx1 -N32 /dev/urandom | tr -d ' \n')
-# The kernel refuses a token file other users can read.
-(umask 077; printf '%s\n' "$token" > "$work/home/remote/launch.token")
-
-export REASONIX_HOME="$work/home"
-export HOME="$work/home"
-export TMPDIR="$work"
-
-echo "kernel: $kernel"
-echo "page: $page"
-echo "token: ${token:0:8}... (${#token} chars)"
-
-"$kernel" serve \
-  --addr 127.0.0.1:0 \
-  --auth token \
-  --token-file "$work/home/remote/launch.token" \
-  --port-file "$work/home/remote/serve.port" \
-  --pid-file "$work/home/remote/serve.pid" \
-  --no-open \
-  -page "$page" \
-  > "$work/kernel.log" 2>&1 &
-echo $! > "$work/kernel.pid"
-
-# Wait for the port file the kernel writes after binding, which is exactly the
-# wait `Kernel.awaitPort` performs.
-for _ in $(seq 1 150); do
-  if [ -s "$work/home/remote/serve.port" ]; then break; fi
-  if ! kill -0 "$(cat "$work/kernel.pid")" 2>/dev/null; then
-    echo "the kernel exited; log:" >&2
-    cat "$work/kernel.log" >&2
-    exit 1
-  fi
-  sleep 0.2
-done
-
-if [ ! -s "$work/home/remote/serve.port" ]; then
-  echo "the kernel reported no listening address; log:" >&2
-  cat "$work/kernel.log" >&2
-  exit 1
-fi
-
-port=$(cat "$work/home/remote/serve.port")
-echo "listening on: $port"
-
-# The two things the WebView will do: fetch the page, and exist on the origin.
-probe=$(curl -s -o /dev/null -w '%{http_code}' "http://$port/")
-echo "GET /  -> $probe"
-if [ "$probe" != "200" ]; then
-  echo "the page did not answer; log:" >&2
-  cat "$work/kernel.log" >&2
-  exit 1
-fi
-
-cat > "$work/entry.txt" <<EOF
-origin=http://$port
-entry=http://$port/#token=$token
-token=$token
-EOF
-
-echo "entry written: $work/entry.txt"
-echo "kernel log (first 10 lines):"
-head -10 "$work/kernel.log"
@@ -1,31 +0,0 @@
-// Reasonix Mobile — the Android shell around the Reasonix kernel.
-//
-// The shell is deliberately thin: it unpacks the kernel and the built page that
-// ship inside the APK, starts `reasonix serve` on a loopback port, and points a
-// WebView at it. Everything the person sees and does is the same page the
-// desktop Studio and `reasonix web` serve, because it is literally the same
-// build of `desktop/frontend-next` — which is what keeps one interface honest
-// across the three surfaces, and why the page is packaged rather than fetched.
-
-pluginManagement {
-    repositories {
-        // The official repositories only. A mirror listed here would be an
-        // unverified trust anchor for anything that ends up in a signed build:
-        // the artifact it serves is whichever copy landed there. A machine that
-        // needs a mirror belongs on a user-level init script, where the choice
-        // stays with the person building rather than travelling with the source.
-        google()
-        mavenCentral()
-        gradlePluginPortal()
-    }
-}
-
-dependencyResolutionManagement {
-    repositories {
-        google()
-        mavenCentral()
-    }
-}
-
-rootProject.name = "ReasonixMobile"
-include(":app")
@@ -1,6 +1,6 @@
 # Reasonix Benchmarks

-Six harnesses live under `benchmarks/`; `cmd/e2ebench` also exposes a
+Five harnesses live under `benchmarks/`; `cmd/e2ebench` also exposes a
 SWE-bench Verified mode:

 - `e2e/` — the committed end-to-end task suite, driven by
@@ -19,10 +19,6 @@ SWE-bench Verified mode:
 - `fanout-width/` — a shape-instructed fleet whose members share nothing, for
  pricing what running them at the same time is worth. See
  [Pricing the run graph](#pricing-the-run-graph) below.
- `realrepo/` — long tasks over a pinned snapshot of this repository, for
-  measuring changes that cut model round trips (13–31 requests per task on
-  `deepseek-flash`). See
-  [its README](realrepo/README.md).

 ## Directory layout

@@ -420,11 +416,11 @@ own outcome).

 | Flag | Default | Purpose |
 | --- | --- | --- |
-| `-mode` | `suite` | `suite` \| `diff` \| `swebench` \| `compare` \| `traj` (`diff` generates tests for the PR diff; `swebench` runs the official per-instance evaluation; `compare` renders KPI/Pareto readouts from 2+ `-json` reports; `traj` re-digests recorded trajectory files without spending tokens; `audit` checks a batch's published figures against its retained rows). |
+| `-mode` | `suite` | `suite` \| `diff` \| `swebench` \| `compare` \| `traj` (`diff` generates tests for the PR diff; `swebench` runs the official per-instance evaluation; `compare` renders KPI/Pareto readouts from 2+ `-json` reports; `traj` re-digests recorded trajectory files without spending tokens). |
 | `-suite` | `benchmarks/e2e` | Suite root (must contain `tasks/<id>/`). |
 | `-task` | *(all)* | Suite mode: run only these comma-separated task IDs (e.g. `-task fix-add-bug`); unknown IDs fail with the available list. |
 | `-attempts` | `1` | Suite and diff modes: retry a task until an attempt passes, up to N; enables the `Pass@≤N` KPI, and TTCS charges a retried solve with its failed attempts' wall. |
-| `-trials` | `1` | Suite mode: run every task N times and keep every outcome. Adds a **Reliability** line: `pass^N` (complete tasks that passed every trial) and `pass@N` (complete tasks with any pass), both over the tasks that ran all N trials only (`n/a` when none did), mean `pass@1` over every graded trial, and the flaky tasks by name; a task the token budget cut short is listed on an **Incomplete** line and is in neither. `Pass@≤N` asks whether the agent can solve a task; `pass^N` asks whether it does so every time. Cannot be combined with `-attempts`. |
+| `-trials` | `1` | Suite mode: run every task N times and keep every outcome. Adds a **Reliability** line: `pass^N` (tasks that passed every trial), `pass@N` (any trial), mean `pass@1` over all trials, and the flaky tasks by name. `Pass@≤N` asks whether the agent can solve a task; `pass^N` asks whether it does so every time. Cannot be combined with `-attempts`. |
 | `-tape-record DIR` | — | Suite mode, with `-meter`: record every provider exchange at the meter, one tape per task and trial (`DIR/<task>/trial-N/0001.request.json`, `.response.json`, `.response.body`). Taped runs use a fixed workdir path and pin the seeded files' times, because both reach the model through the workspace block and file listings. |
 | `-tape-replay DIR` | — | Suite mode, with `-meter`: answer each run from its tape instead of the provider, and add a **Replay** line naming the first request whose body differs from the recorded one, down to the message (`request 2, message 4 (tool)`), the tool list, or a field. The diverging request is kept as `NNNN.replayed.json` beside the recorded one. A kernel change that should not alter what the model is asked replays with every request matched. |
 | `-bin` | `reasonix` | Path to the reasonix binary. |
@@ -433,7 +429,6 @@ own outcome).
 | `-ablate` | *(none)* | Ablation arm: comma-separated subsystems to switch off — `evidence`, `planner`, `subagent`, `retrieval`, `compaction`, `upstream` and `recall-search`; `none` \| `all`. |
 | `-out` | *(stdout)* | Write the markdown report here. |
 | `-json` | *(none)* | Write the JSON report here (optional). |
-| `-batch` | *(none)* | Suite mode: write `results.jsonl` (one row per attempt) and `batch.json` (manifest v1) here, so the report's figures can be audited later. Audit mode: read the batch from here. See [Auditing a batch](#auditing-a-batch). |
 | `-trajectories` | *(none)* | Suite mode: write one `<task-id>.trajectory.jsonl` per task into this directory (the agent's full event stream with timestamps — see `reasonix run --trajectory`). The report gains a time-attribution line (tools vs. model) and each JSON result a `trajectory` digest. |
 | `-force-planner` | `false` | Suite mode: prefix each prompt with a plan-first directive so the two-model turn engages regardless of the planner gate. Use for the "with planner" arm of an A/B; results carry `plan_forced` so arms are only comparable with equal forcing. |
 | `-anchor` | `blind` | Suite mode: which hypothesis the agent holds before it looks at anything — `blind` (none, the control) \| `correct` \| `wrong`. The seeded arms prefix each prompt with the task's authored seed and **skip** tasks that have none, so an unseeded control run never lands in a seeded denominator. Results carry `anchor`. See [Anchor resistance](#anchor-resistance). |
@@ -456,65 +451,10 @@ Diff-mode flags:
 | `-timeout` | `1200` | Agent timeout in seconds (diff mode). |
 | `-attempts` | `1` | Diff mode: retry up to N times until a run passes (stochastic agent). |

-## Auditing a batch
-
-`-batch DIR` writes two files from a suite run: `results.jsonl` (one row per
-attempt — the raw record every figure is computed from) and `batch.json`
-(manifest v1). `-mode audit -batch DIR` re-derives the report's headline figures
-from `results.jsonl`, prints the manifest's claim beside the recomputed value,
-and exits non-zero naming any figure that moved. It also refuses a batch whose
-manifest is missing, records no commit, or whose rows are truncated.
-
-This is an **aggregation audit, not a replay**: it proves the published figures
-follow from the retained rows. It does not re-execute the model, tools or
-environment, and cannot show the experiment would reproduce. A real replay would
-need the commit, a corpus fingerprint, the model and provider configuration, the
-seed, the flags, the raw results and the runtime version; it is not implemented.
-
-`schema_version` is top level. The rest is three groups:
-
-| Group | Fields | Answers |
-| --- | --- | --- |
-| `over` | `command.mode` and `command.axes` (every flag that changes what a figure means; paths omitted); `tasks.count` and `tasks.digest` (sha256 over sorted, NUL-separated task ids — the corpus identity, never the task text); `filters`; `denominator` (`unit`, `definition`, `value`) | what the figures are over |
-| `producer` | `commit` (or an explicit `unknown`, never a guess), `commit_source`, `dirty`, `model`, `provider`, `seed`, `started_at`, `finished_at`, `harness_version`, `go_version` | which run produced them |
-| `trust` | `results_sha256`, `claimed` (the headline figures), `ledger`, `retention` | why they can still be checked |
-
-The ledger accounts for every planned run, and the audit checks that the counts
-form a funnel (`scored ≤ completed ≤ attempted`, `failed ≤ scored`,
-`scored = pass1 + failed`, `rows = attempted + skipped`). A row count is never a
-denominator: `denominator.value` is a count of scored tasks, and `rows` is
-published only to show what was retained.
-
-| Count | Meaning |
-| --- | --- |
-| `planned` | graded runs the batch set out to attempt: filtered tasks × trials |
-| `attempted` | retained rows for runs that started (skipped rows are not attempts) |
-| `completed` | attempts that ended and left a row; this harness writes a row only after an attempt ends, so it equals `attempted` and the equality is checked, not assumed |
-| `scored` | completed first attempts of solvable, non-skipped tasks — the pass@1 denominator |
-| `failed` | scored attempts the grader did not pass |
-| `unmeasured` | attempts whose cost and tokens are unknown because no metrics landed; kept out of spend totals, never counted as zero |
-| `excluded` | rows kept out of the accuracy denominator: skipped, `no_solution`, and retry entries |
-
-Spend and cache fields are omitted from `claimed` when nothing was measured, so
-an absent figure reads as unknown rather than as a measured zero. `seed` is
-omitted for the same reason: the harness does not set a sampling seed.
-`results_sha256` covers the values and the presence of every row; line endings,
-indentation and key order are normalised away because none of them is evidence,
-while number spelling is kept (a rewritten `1` is not `1.0`).
-A successful audit records the verifier version (`audit.verifier_version`) and
-the digest it checked in `batch.json`, so a re-verification can tell which code
-checked the batch.
-
-`results.jsonl` is retained unredacted beside the manifest — task prompts and
-agent notes included — and is not public. The manifest itself carries a task-id
-digest, no prompt text, no absolute paths, no endpoints or keys and no model
-output. A claim whose rows are gone cannot be audited: there is nothing to
-recompute from, and this tool will not certify it.
-
 ## Dataset retention

-Keep every `-json` report, `-trajectories` directory and `-batch` directory from
-real runs: they are the accumulating corpus — per-task contracts-to-be, full event
+Keep every `-json` report and `-trajectories` directory from real runs: they
+are the accumulating corpus — per-task contracts-to-be, full event
 trajectories, checkpoint oracle verdicts, stop curves and phase traces — that
 any future offline learning (routing, stop policies, budgets) would train
 and evaluate on. The control plane stays deterministic and interpretable
@@ -1,35 +0,0 @@
-# Compaction prefix probe
-
-## Method
-
-1. Measured on 2026-10-17 against the official `https://api.deepseek.com` OpenAI adapter, using the model returned by `/models`: `deepseek-flash`. Thinking was disabled. Inputs were synthetic; no workspace history or credentials are recorded here.
-2. Each of three text and three rich-history trials sent a main request to warm the prefix, then called the actual agent summary path, then an independent summary control. Rich history included a synthetic closed `read_file` call, result and reasoning. Warmup outputs were discarded, never executed.
-3. A summary had a 1,024-token output cap; warmups had 64. Success means a heading, preservation of the fixture's `ParseDuration` and zero-value constraint, no emitted tool calls, and `finish_reason=stop`. This checks the fixture's contract, not arbitrary-summary factual completeness.
-4. First-attempt reuse success counts successful reused summaries / six reused summaries. Fallback rate counts independent requests initiated by the reuse coordinator / six reuse trials; the six explicit independent controls are excluded. Both were 6/6 success and 0/6 fallbacks. Independent controls also succeeded 6/6.
-5. The raw result file records provider-reported usage and full stream latency, measured from adapter `Stream` until channel close. Token rows below sum only the three indicated summary requests; they exclude warmup, model listing and any earlier exploratory requests. Cache residency was uncontrolled; prior probes had warmed these synthetic prefixes.
-
-## Results
-
-| History | Shape | N | Prompt | Hit | Miss | Output | Median ms | Estimated USD |
-| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
-| Text | Reused | 3 | 3,714 | 3,072 | 642 | 636 | 1,713 | 0.000487116 |
-| Text | Independent | 3 | 1,218 | 768 | 450 | 644 | 1,610 | 0.000456204 |
-| Rich | Reused | 3 | 3,780 | 3,072 | 708 | 742 | 1,774 | 0.000560616 |
-| Rich | Independent | 3 | 1,341 | 768 | 573 | 921 | 1,815 | 0.000640854 |
-
-1. Estimates use the [official pricing](https://api-docs.deepseek.com/quick_start/pricing/) checked on the measurement date: off-peak Flash USD per million tokens, hit 0.003, miss 0.15, output 0.6. October 17 was a Saturday. Formula: `(hit*0.003 + miss*0.15 + output*0.6)/1_000_000`. These are token estimates, not account deductions.
-2. Reuse hit fractions were 82.7% (text) and 81.3% (rich), versus 63.1% and 57.3% for controls. This small fixture does **not** establish cost savings: retaining the system/catalog increased both total input and input cost. Text total cost increased 6.8%; rich total decreased 12.5% because output was shorter, not because input was cheaper.
-3. The six warmups consumed 7,032 prompt and 231 output tokens, estimated USD 0.000346680. All 18 recorded generation calls together cost an estimated USD 0.002491470. No failed-digest retry occurred in these live trials; the deterministic fallback tests establish its behavior and accounting.
-4. A rejected nonempty reused answer earns one independent-shape fallback. A separate truncation retry may enlarge the output cap, for at most three logical model calls total. Every actual attempt emits and records its own spend; returned usage is their sum. Empty output, stream errors and cancellation do not earn shape retries.
-5. These sequential, six-pair trials establish interoperability for one model, mode and short fixture. They do not estimate production failure probability, thinking-mode behavior, other models, cold-cache cost, or large-history savings.
-
-## Reproduction
-
-1. Provide a test key through an existing secret environment configuration. Do not put credentials in source, command arguments or captured logs. Select a model currently returned by `/models`.
-2. Run the opt-in probe; it skips without `DEEPSEEK_API_KEY`. The dedicated build tag avoids enabling unrelated live tests.
-
-```sh
-REASONIX_LIVE_MODEL=deepseek-flash go test -tags live_compaction ./internal/runtime/agent -run '^TestLiveCompactionPrefix$' -v -count=1
-```
-
-3. This probe exercises actual summary construction and the HTTP adapter. The hermetic `TestEffectCompactionSharesThePreviouslySentPrefix` separately exercises manual, automatic pressure and rich history through `boot.Build`, comparing actual JSON request-prefix and schema bytes. Its sparse-history case also proves an interleaved retained user turn survives the fold.
@@ -1,512 +0,0 @@
-[
-  {
-    "model": "deepseek-flash",
-    "rich": false,
-    "trial": 1,
-    "results": [
-      {
-        "kind": "warm",
-        "milliseconds": 828,
-        "digest": false,
-        "grounded": false,
-        "tool_calls": 1,
-        "usage": {
-          "PromptTokens": 1127,
-          "CompletionTokens": 38,
-          "TotalTokens": 1165,
-          "CacheHitTokens": 896,
-          "CacheMissTokens": 231,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "tool_calls",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "reuse",
-        "milliseconds": 2006,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 1238,
-          "CompletionTokens": 234,
-          "TotalTokens": 1472,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 214,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "independent",
-        "milliseconds": 1817,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 406,
-          "CompletionTokens": 234,
-          "TotalTokens": 640,
-          "CacheHitTokens": 256,
-          "CacheMissTokens": 150,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      }
-    ]
-  },
-  {
-    "model": "deepseek-flash",
-    "rich": false,
-    "trial": 2,
-    "results": [
-      {
-        "kind": "warm",
-        "milliseconds": 820,
-        "digest": false,
-        "grounded": false,
-        "tool_calls": 1,
-        "usage": {
-          "PromptTokens": 1127,
-          "CompletionTokens": 38,
-          "TotalTokens": 1165,
-          "CacheHitTokens": 896,
-          "CacheMissTokens": 231,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "tool_calls",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "reuse",
-        "milliseconds": 1633,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 1238,
-          "CompletionTokens": 187,
-          "TotalTokens": 1425,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 214,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "independent",
-        "milliseconds": 1555,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 406,
-          "CompletionTokens": 202,
-          "TotalTokens": 608,
-          "CacheHitTokens": 256,
-          "CacheMissTokens": 150,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      }
-    ]
-  },
-  {
-    "model": "deepseek-flash",
-    "rich": false,
-    "trial": 3,
-    "results": [
-      {
-        "kind": "warm",
-        "milliseconds": 905,
-        "digest": false,
-        "grounded": false,
-        "tool_calls": 1,
-        "usage": {
-          "PromptTokens": 1127,
-          "CompletionTokens": 38,
-          "TotalTokens": 1165,
-          "CacheHitTokens": 896,
-          "CacheMissTokens": 231,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "tool_calls",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "reuse",
-        "milliseconds": 1713,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 1238,
-          "CompletionTokens": 215,
-          "TotalTokens": 1453,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 214,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "independent",
-        "milliseconds": 1610,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 406,
-          "CompletionTokens": 208,
-          "TotalTokens": 614,
-          "CacheHitTokens": 256,
-          "CacheMissTokens": 150,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      }
-    ]
-  },
-  {
-    "model": "deepseek-flash",
-    "rich": true,
-    "trial": 1,
-    "results": [
-      {
-        "kind": "warm",
-        "milliseconds": 561,
-        "digest": false,
-        "grounded": false,
-        "tool_calls": 1,
-        "usage": {
-          "PromptTokens": 1217,
-          "CompletionTokens": 39,
-          "TotalTokens": 1256,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 193,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "tool_calls",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "reuse",
-        "milliseconds": 1774,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 1260,
-          "CompletionTokens": 256,
-          "TotalTokens": 1516,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 236,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "independent",
-        "milliseconds": 2271,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 447,
-          "CompletionTokens": 381,
-          "TotalTokens": 828,
-          "CacheHitTokens": 256,
-          "CacheMissTokens": 191,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      }
-    ]
-  },
-  {
-    "model": "deepseek-flash",
-    "rich": true,
-    "trial": 2,
-    "results": [
-      {
-        "kind": "warm",
-        "milliseconds": 610,
-        "digest": false,
-        "grounded": false,
-        "tool_calls": 1,
-        "usage": {
-          "PromptTokens": 1217,
-          "CompletionTokens": 39,
-          "TotalTokens": 1256,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 193,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "tool_calls",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "reuse",
-        "milliseconds": 2263,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 1260,
-          "CompletionTokens": 276,
-          "TotalTokens": 1536,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 236,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "independent",
-        "milliseconds": 1783,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 447,
-          "CompletionTokens": 242,
-          "TotalTokens": 689,
-          "CacheHitTokens": 256,
-          "CacheMissTokens": 191,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      }
-    ]
-  },
-  {
-    "model": "deepseek-flash",
-    "rich": true,
-    "trial": 3,
-    "results": [
-      {
-        "kind": "warm",
-        "milliseconds": 834,
-        "digest": false,
-        "grounded": false,
-        "tool_calls": 1,
-        "usage": {
-          "PromptTokens": 1217,
-          "CompletionTokens": 39,
-          "TotalTokens": 1256,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 193,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "tool_calls",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "reuse",
-        "milliseconds": 1575,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 1260,
-          "CompletionTokens": 210,
-          "TotalTokens": 1470,
-          "CacheHitTokens": 1024,
-          "CacheMissTokens": 236,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      },
-      {
-        "kind": "independent",
-        "milliseconds": 1815,
-        "digest": true,
-        "grounded": true,
-        "tool_calls": 0,
-        "usage": {
-          "PromptTokens": 447,
-          "CompletionTokens": 298,
-          "TotalTokens": 745,
-          "CacheHitTokens": 256,
-          "CacheMissTokens": 191,
-          "CacheWriteTokens": 0,
-          "CacheWriteBilledTokens": 0,
-          "ReasoningTokens": 0,
-          "FinishReason": "stop",
-          "Estimated": false,
-          "RequestCount": 1,
-          "ServerToolRequests": 0,
-          "ContextPromptTokens": 0,
-          "ContextCompletionTokens": 0,
-          "ContextReasoningTokens": 0,
-          "ContextCacheHitTokens": 0,
-          "ContextCacheMissTokens": 0
-        }
-      }
-    ]
-  }
-]
@@ -3,12 +3,12 @@
 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=preflight     build all 12 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=run-boundary  Experiment I, paired: 6 tasks across their own cue boundary
+    -mode=run-index     Experiment I, full: 6 tasks x 4 index 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
@@ -22,28 +22,17 @@ 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).
+**Fold index marginal utility: unresolved.** Both the positive and the negative
+estimates are invalidated. The first Stage 2 ran in a contaminated environment.
+The clean paired batch found a consistent effect (six of six tasks searched less
+and paged in less with the cue visible), and the same-batch dose-response did
+not reproduce it (`boundary-aligned 0/4`, recall-token delta reversing sign).
+The most likely reason is that two of the six index tasks carry heavy-tail
+stopping behaviour that swamps the effect being measured.

-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.
+Existing evidence is not sufficient to tune the shipped 1% budget in either
+direction. Reopening this wants a new corpus, not more samples: see
+"Index corpus v2" below.

 **Things that turned out not to be problems**, each after being measured rather
 than argued about:
@@ -85,83 +74,11 @@ 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
+## Index corpus v2, if this is reopened

-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
+The present index tasks are not a fit substrate: two of six carry stopping
+behaviour that dominates the signal. A task should qualify only if a run-index
+run reports 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`,
@@ -169,7 +86,8 @@ class comes from the three counts in that report — `searches_after_sufficient`
 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.
+nothing in the workspace. Keep i03 and i04 — as a stopping corpus, not an
+efficiency one.

 `-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
@@ -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)
-	}
-}
@@ -18,8 +18,7 @@ import (
 // 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.
+// failure stages, in funnel order. Exactly one is reported per run.
 const (
 	stageReadMissed   = "NoSearch"           // never asked, and never read the target either
 	stageSearchMiss   = "SearchMiss"         // searched, target never came back
@@ -27,7 +26,6 @@ const (
 	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
@@ -91,15 +89,6 @@ type contextMetrics struct {
 	// 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.
@@ -118,12 +107,6 @@ type contextMetrics struct {
 	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.
@@ -470,12 +453,6 @@ type funnel struct {
 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]++
@@ -645,7 +622,7 @@ func reportStopping(all []contextMetrics) string {
 	var extraRounds, extraTokens []float64
 	sufficient, answeredNext := 0, 0
 	for _, m := range all {
-		if m.Failed || m.Contaminated {
+		if m.Contaminated {
 			continue
 		}
 		classes[m.StoppingClass]++
@@ -26,12 +26,6 @@ const (
 	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
@@ -51,8 +45,6 @@ type contextTask struct {
 	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.
@@ -129,20 +121,6 @@ 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{
 		{
@@ -214,104 +192,10 @@ func searchTasks() []contextTask {
 	}
 }

-// 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 {
+func indexTasks() []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,
+			ID: "i01-transport-fallback", Experiment: experimentIndex, TargetKind: "tool_input",
 			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")},
@@ -322,7 +206,7 @@ func stoppingIndexTasks() []contextTask {
 			PlantBody: "fallback selector = {{selector}}",
 		},
 		{
-			ID: "i02-scheduler-handoff", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
+			ID: "i02-scheduler-handoff", Experiment: experimentIndex, TargetKind: "tool_input",
 			Prompt:     "scheduler handoff 当时记录的 grace 和 marker 是什么？",
 			ProbeQuery: "scheduler handoff grace marker",
 			Vars: []varSpec{
@@ -336,7 +220,7 @@ func stoppingIndexTasks() []contextTask {
 			PlantBody: "handoff grace = {{grace}}\nmarker = {{marker}}",
 		},
 		{
-			ID: "i03-coalescing", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
+			ID: "i03-coalescing", Experiment: experimentIndex, TargetKind: "tool_input",
 			Prompt:     "job completion 的 coalescing probe 最后记录的 window 和 marker 是多少？",
 			ProbeQuery: "coalescing policy completion window",
 			Vars: []varSpec{
@@ -350,7 +234,7 @@ func stoppingIndexTasks() []contextTask {
 			PlantBody: "completion coalesce window = {{window}}\nmarker = {{marker}}",
 		},
 		{
-			ID: "i04-recovery-fence", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
+			ID: "i04-recovery-fence", Experiment: experimentIndex, TargetKind: "tool_input",
 			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{
@@ -364,7 +248,7 @@ func stoppingIndexTasks() []contextTask {
 			PlantBody: "{{marker}}\nexpected epoch {{expected}}\nobserved epoch {{observed}}",
 		},
 		{
-			ID: "i05-cache-probe", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
+			ID: "i05-cache-probe", Experiment: experimentIndex, TargetKind: "tool_input",
 			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")},
@@ -375,7 +259,7 @@ func stoppingIndexTasks() []contextTask {
 			PlantBody: "prefix salt = {{salt}}\nscope = {{scope}}",
 		},
 		{
-			ID: "i06-dispatch-handoff", Experiment: experimentIndex, TargetKind: "tool_input", Role: roleStopping,
+			ID: "i06-dispatch-handoff", Experiment: experimentIndex, TargetKind: "tool_input",
 			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")},
@@ -388,12 +272,6 @@ func stoppingIndexTasks() []contextTask {
 	}
 }

-// 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()...)
 }
@@ -454,9 +332,6 @@ func validateCorpus() error {
 			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)
 			}
@@ -121,7 +121,7 @@ var doseMetrics = []struct {
 func reportDoseResponse(all []contextMetrics) string {
 	valid := map[string]map[string]contextMetrics{}
 	for _, m := range all {
-		if m.Failed || m.Contaminated || m.FailureStage == "RunError" {
+		if m.Contaminated || m.FailureStage == "RunError" {
 			continue
 		}
 		scale := strings.TrimPrefix(m.Arm, "index-")
@@ -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)
-}
@@ -36,7 +36,7 @@ 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.ChunkText, Text: "Earlier work continued across several files; nothing outstanding was recorded."}
 	ch <- provider.Chunk{Type: provider.ChunkDone}
 	close(ch)
 	return ch, nil
@@ -5,7 +5,6 @@ import (
 	"crypto/rand"
 	"fmt"
 	"hash/fnv"
-	"io"
 	mathrand "math/rand"
 	"strings"
 )
@@ -91,25 +90,15 @@ func seededRand(parts ...string) *mathrand.Rand {
 // 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
+	if _, err := rand.Read(seed[:]); err != nil {
+		panic("contextbench: no entropy for a live run: " + err.Error())
 	}
 	var n int64
 	for _, b := range seed {
 		n = n<<8 | int64(b)
 	}
-	return mathrand.New(mathrand.NewSource(n)), nil
+	return mathrand.New(mathrand.NewSource(n))
 }

 // instantiate substitutes {{name}} and refuses anything left over. The grammar
@@ -54,22 +54,15 @@ func main() {
 	case "preflight":
 		os.Exit(runPreflight(tasks, root))
 	case "run-search":
-		os.Exit(runExperiment(*mode, experimentSearch, root, *dry, tasks))
+		os.Exit(runExperiment(experimentSearch, root, *dry, tasks))
 	case "run-boundary":
-		os.Exit(runBoundaries(*mode, root))
+		os.Exit(runBoundaries(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))
+		os.Exit(runExperiment(experimentIndex, root, *dry, tasks))
 	default:
 		fmt.Fprintf(os.Stderr, "unknown mode %q\n", *mode)
 		os.Exit(2)
@@ -1,10 +1,7 @@
 package main

 import (
-	"bytes"
 	"fmt"
-	"maps"
-	mathrand "math/rand"
 	"os/exec"
 	"slices"
 	"strings"
@@ -22,24 +19,16 @@ func TestCorpusIsWellFormed(t *testing.T) {
 	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)
+	if got := len(indexTasks()); got != 6 {
+		t.Errorf("index experiment has %d tasks, want 6", 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.
+	// Two per tier 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])
@@ -47,21 +36,6 @@ func TestCorpusIsWellFormed(t *testing.T) {
 	}
 }

-// 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) {
@@ -85,46 +59,20 @@ func TestNoAnswerLiteralExistsInTheRepository(t *testing.T) {
 	}
 }

-// 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)))
+// Two runs of the same task must not share an answer, or one leaked run leaks
+// every future one.
+func TestInstancesDifferBetweenRuns(t *testing.T) {
+	for _, task := range allTasks() {
+		a, err := instantiateTask(task, liveRand())
 		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)
+		b, err := instantiateTask(task, liveRand())
+		if err != nil {
+			t.Fatal(err)
 		}
-		c, _ := instantiateTask(task, other)
-		if maps.Equal(a.Vars, c.Vars) {
-			t.Errorf("%s: different entropy gave the same values", task.ID)
+		if slices.Equal(a.AnswerMarkers, b.AnswerMarkers) {
+			t.Errorf("%s produced the same answers twice: %v", task.ID, a.AnswerMarkers)
 		}
 	}
 }
@@ -130,7 +130,7 @@ type queryAudit struct {
 func auditQueries(runs []contextMetrics) queryAudit {
 	a := queryAudit{Reformulations: map[string]int{}, MissedProbeTerms: map[string]int{}}
 	for _, r := range runs {
-		if r.Failed || r.Contaminated {
+		if r.Contaminated {
 			continue
 		}
 		a.Runs++
@@ -14,6 +14,7 @@ import (

 	"reasonix/internal/assembly/boot"
 	"reasonix/internal/contract/ablation"
+	"reasonix/internal/contract/event"
 	"reasonix/internal/contract/provider"
 	"reasonix/internal/contract/tool"
 	_ "reasonix/internal/model/openai"
@@ -85,7 +86,6 @@ func runAdversarial(root string) int {
 			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 {
@@ -132,11 +132,10 @@ func runOne(p provider.Provider, t contextTask, arm ablation.Set, armName, root
 	}
 	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)
+	}, event.Discard)
 	a.LoadProjectionSidecar(f.Path)
 	// From here the transcript lives only in memory. On disk it is the answer.
 	sealFixture(f.Path)
@@ -162,10 +161,6 @@ func runOne(p provider.Provider, t contextTask, arm ablation.Set, armName, root
 	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"
@@ -174,16 +169,6 @@ func runOne(p provider.Provider, t contextTask, arm ablation.Set, armName, root
 	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) {
@@ -227,10 +212,7 @@ func isolateWorkspace() (func(), error) {

 // 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()
+func runBoundaries(root string) int {
 	root, err := filepath.Abs(root)
 	if err != nil {
 		fmt.Fprintln(os.Stderr, err)
@@ -250,18 +232,16 @@ func runBoundaries(mode, root string) int {

 	var all []contextMetrics
 	var cueSide, noCueSide []contextMetrics
-	for _, t := range efficiencyTasks() {
+	for _, t := range indexTasks() {
 		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 {
@@ -269,14 +249,14 @@ func runBoundaries(mode, root string) int {
 			} 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",
+			fmt.Printf("%-24s %-14s cue=%-3v %-14s search=%d read=%d direct=%-5v recall-tok=%-5d escape=%d\n",
 				t.ID, name, side.has, m.FailureStage, m.SearchCalls, m.ReadCalls,
-				m.CueDirectRead, m.RecallReturnedTokens, promptTokenSuffix(m), m.EscapeCalls)
+				m.CueDirectRead, m.RecallReturnedTokens, m.EscapeCalls)
 		}
 	}
 	reportDeltas(cueSide, noCueSide)
 	fmt.Print(reportStopping(all))
-	return writeResults(root, batchInfo{Mode: mode, Planned: planned, Started: started, Repo: repo}, all)
+	return writeResults(root, all)
 }

 // reportDeltas is the paired reading: each task against itself, then pooled.
@@ -318,10 +298,7 @@ func deltaBool(a, b bool) string {
 // 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()
+func runExperiment(experiment, root string, dry bool, tasks []contextTask) int {
 	root, err := filepath.Abs(root)
 	if err != nil {
 		fmt.Fprintln(os.Stderr, err)
@@ -357,17 +334,15 @@ func runExperiment(mode, experiment, root string, dry bool, tasks []contextTask)
 		// 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",
+			fmt.Printf("%-24s %-14s %-18s search=%d hit=%d read=%d recall-tok=%d %v\n",
 				t.ID, arm.name, m.FailureStage, m.SearchCalls, m.TargetSearchHits, m.ReadCalls,
-				m.RecallReturnedTokens, promptTokenSuffix(m), m.UnexpectedWorkTools)
+				m.RecallReturnedTokens, m.UnexpectedWorkTools)
 		}
 	}
 	reportFunnels(byArm)
@@ -380,7 +355,7 @@ func runExperiment(mode, experiment, root string, dry bool, tasks []contextTask)
 	if experiment == experimentIndex {
 		reportBoundaries(all)
 	}
-	return writeResults(root, batchInfo{Mode: mode, Planned: planned, Started: started, Dry: dry, Repo: repo}, all)
+	return writeResults(root, all)
 }

 func cueExpectedVisible(t contextTask, scale string) bool {
@@ -435,7 +410,7 @@ func writeTrajectories(root string, all []contextMetrics) error {
 	return nil
 }

-func writeResults(root string, info batchInfo, all []contextMetrics) int {
+func writeResults(root string, all []contextMetrics) int {
 	path := filepath.Join(root, "results.jsonl")
 	f, err := os.Create(path)
 	if err != nil {
@@ -443,30 +418,17 @@ func writeResults(root string, info batchInfo, all []contextMetrics) int {
 		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)
+	fmt.Printf("\n%d runs written to %s (with trajectories.jsonl)\n", len(all), path)
 	return 0
 }

@@ -475,21 +437,9 @@ func writeResults(root string, info batchInfo, all []contextMetrics) int {
 // 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",
@@ -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)
-	}
-}
@@ -1,3 +0,0 @@
-alpha=QORVEX-7741
-beta=MIRALD-3208
-gamma=TESSIK-9155
@@ -1,87 +0,0 @@
-# Desktop surface evaluation
-
-This protocol supplies the fixed-task trajectory comparison required by
-[RFC #12584](https://github.com/esengine/DeepSeek-Reasonix/issues/12584).
-There are no measured results yet. Component tests and scripted providers prove
-delivery, validation and replay; they do not measure a model's choice of interface
-or a person's effort.
-
-## Paired runs
-
-Run each task twice in fresh desktop sessions: ordinary text/ask interaction and
-native surfaces. Use the same model, sampling settings, initial workspace,
-permission mode and task prompt. Record commit, model, configuration and the
-native-surface capability setting. Alternate condition order across repetitions.
-Keep the hidden tool out of the provider prefix in both conditions; only its
-availability through capability discovery differs.
-
-The baseline may batch questions or accept a single text reply. Never force it to
-ask one question per field. Do not tell either model how many turns to use or
-which component to render. If the surface-enabled model chooses text, retain
-that run in the surface condition. Interface choice is part of the experiment.
-
-The operator knows the private values below but does not volunteer them before
-the model requests the corresponding information. Supply all requested values
-together, regardless of presentation. Allow inspection and local editing before
-submission. Save the trajectory and an observer ledger locally; raw transcripts
-and draft content must not enter the aggregate report.
-
-## Fixed tasks
-
-| Task | Initial prompt, identical in both conditions | Private operator input | Completion criterion |
-| --- | --- | --- | --- |
-| `release-brief` | Prepare a release brief. Ask me for the audience, release date and whether this is a staged rollout before writing it. | audience: existing desktop users; date: 2026-11-03; staged rollout: yes | Final brief contains all three values and adds no release commitment. |
-| `storage-choice` | Help me choose storage for a small desktop app. Present reasonable choices, ask me to choose and ask whether offline operation is required. Then summarize the decision. | storage: SQLite; offline: yes | Final decision states SQLite and offline operation, with no installation or file changes. |
-| `draft-revision` | Draft a two-line release announcement and let me revise it before returning the final version verbatim. | Replace with the exact string `  Desktop release ready.\nKeep existing settings.\n` (decode the escapes). | Final returned string has exactly the specified leading spaces and newlines. Formatting wrappers are excluded from the comparison. |
-| `preview-decision` | Show a plain-text before/after preview changing retry count from 2 to 3. Ask whether to keep that change and whether to include a release note. Return a decision summary; do not change files. | keep change: yes; release note: no | Preview shows 2 to 3; final summary preserves both decisions; workspace is unchanged. |
-
-Freeze this task set before collecting results. Record a new protocol version if
-prompts, private inputs, completion criteria or metric definitions change.
-Neither the model's final claim nor a successful input acknowledgment establishes
-completion. The observer checks the resulting output against the criterion.
-
-## Metric definitions (version 1)
-
-The anchor is acknowledgment of the initial task prompt. The observation ends
-when the criterion is verified, the operator abandons the task, or the run fails.
-The initial prompt is excluded from subsequent-input counts because both
-conditions start from the same request.
-
-| Metric | Count and population | Exclusions |
-| --- | --- | --- |
-| `acknowledged_inputs` | Distinct operator submissions acknowledged after the anchor: chat replies, ask answers, surface answers and surface actions. Includes acknowledged input later withdrawn or superseded. | Initial task prompt, local draft edits, retries with the same acknowledged request identity, automatic dispatch, replay and observer grading. This counts input exchanges, not provider rounds. |
-| `submission_attempts` | Explicit operator send attempts, including rejected submissions and manual retries. | Automatic transport retries, replay, local edits and model tool calls. |
-| `withdrawals` | Successful operator withdrawals of admitted input. | Failed withdrawals, automatic cancellation and replay. |
-| `completed` | Runs whose output meets the fixed criterion. Denominator: every started run in each condition. | None; failures and abandoned runs remain in the denominator. |
-
-Record these in a content-free ledger: task id, repetition, condition, observer
-event ordinal, event type, opaque request identity, acknowledgment status and
-completion verdict. Record the source trajectory reference and observed start
-and end anchors alongside it. Distinguish manual observation from host-recorded
-events. Do not copy labels, prompts, private values, drafts or answer strings into
-the ledger. Identities are local correlation handles, not account identifiers.
-
-The current pane counters and rendered trajectory export cannot reconstruct all
-acknowledgments, refusals and withdrawals. Do not infer them from `ask` calls,
-tool-result counts, final prose, message wording or card mounts. Until structural
-acknowledgment records cover every channel, an observer ledger is required.
-
-## Report and decision
-
-Report the number started, completed, failed and abandoned for each condition,
-then paired input-count differences for repetitions where **both** conditions
-meet the same criterion. List excluded pairs and their reason beside that
-comparison. Also report submission attempts and withdrawals; fewer admitted
-inputs with more failures is not automatically less effort.
-
-Keep per-task results visible. Do not combine scripted fixtures with observed
-model/operator runs or describe this small task set as an account-wide result.
-Report model/configuration differences and missing evidence as limitations.
-
-The RFC's benefit remains unproven until observed trajectories show fewer input
-exchanges to the same verified outcome. These results alone do not justify P3;
-adaptive layout or MCP Apps needs a measured problem in addition to this result.
-
-The owner-requested map, comparison and dashboard workspaces extend the delivery
-scope independently of this study. Their scripted acceptance fixtures verify
-rendering and interaction only; they do not count as observed benefit evidence.
@@ -1,4 +1,5 @@
 prompt = "为本项目新建一个数据库迁移脚本文件（空文件即可），文件名必须符合项目的迁移命名规范。"
 class = "cjk"
+max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-CJK5T"]
@@ -1,4 +1,5 @@
 prompt = "Write the deploy region for this project (just the region string) into answer.txt."
 class = "conflict"
+max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-CONF2P"]