# Independent A1 proof-audit addendum

Android proof sufficient at [d3de380d](https://git.cafe/pingdotgg/t3code/commit/d3de380d7a5f39986844467f53b2da9165e834ee). Independently rebuilt the head development JS bundle with React Compiler and --clear, using the cached native fingerprint and explicit isolated home. No native build, tests, source changes, or checkout switch.

A fourth real sleep30/Done turn in the indexed target completed 08:15:13.663 UTC. Raw recording shows Home with expanded Working, Done under Working, then automatic return above the same older control and removal of Working. No taps after expansion (telemetry36.268s). Sampled transition is Done by~44s, returned by~46.25s, consistent with grace. Home remains visible through the stable tail. Still captured >17s after persisted completion. Raw recording81.220789s at1080×2424; no fabricated frames or inserted holds. Video PTS is not synchronized precisely to server UTC.

Compared independent settled still against delivered A1-after-grace: 2,746/2,617,920 pixels differ (0.104892%); 2,448 exceed max-channel16 (0.093509%); mean absolute channel delta0.115481/255. All differences are confined to clock (927 pixels), status-bar cellular icon (132), and older-control status (1,687). Excluding these three regions gives zero differing pixels. No non-AA unexpected layout difference. Clock8:15 vs7:43; control1h vsDone follows its later visit (07:13:18.185 after07:06:41.138 completion), clearing unseen completion. Target transcript has one additional real turn, not fabricated content; Home title/layout unchanged.

Also inspected newly supplied zero-amplified/threshold16 images and make-diffs.py: genuine RGB absolute subtraction,8× saturation, max-channel>16 mask. Supplied base/head pair has36,722 nonzero pixels,33,207 above16, mean channel delta1.622836; changes confined to status bar and y300–800 intended row/header reorder. No changes belowy800. Subtraction ghosts are displaced text, not runtime duplicates.

Setup was difficult: daemon unavailable, stale binding/release errors, worker disconnect, cold bundle/blank screen, shared-content/local storage warnings, and initially delayed target transcript. These precede the decisive run; existing transcript, running Working state, automatic return, and persisted ordinal4 establish the tested behavior. Local persistence is not verified by this run. Cached development-client evidence is not release-build proof. iOS remains unverified.

FRICTION category|what|avoidance

- Device pool|daemon unavailable/stale binding/worker disconnected; recovered same session|serial reconnect/retry, retain errors, require usable snapshot before sending.
- Startup|cold bundle/blank screen and early pairing|wait for compiled bundle and loaded UI before pairing/capture.
- Storage|shared-content dialog and local persistence warnings|dismiss and confirm live thread data; do not claim persistence coverage.
- Selectors|one malformed Navigate up selector; delayed target load|quote whole selectors and require existing transcript before Send.
- Comparison|clock, cellular indicator, later control visit, extra genuine turn|quantify exact regions and explain fixture history rather than label all differences AA.

Teardown: pool close returned Closed; retained backend/Metro stopped, ports13773/8081 empty. Clean requested head preserved. No forge writes/push/merge/watch or tests. Scratch only under /tmp. Model/harness: gpt-6.1-sol through Codex in T3 Code.

VERDICT: SUFFICIENT