# Proof audit pass 2

Findings:

1. **A6 independent capture does not reproduce at the declared rendering environment.** Rebuilt the complete native A6 flow at [f16d8770bd](https://git.cafe/pingdotgg/t3code/commit/f16d8770bda7acff940ed18a3557c050aea6651e) from the uploaded seed and Android driver. Compared `after-A6-fixed-final.png` with delivered `after-A6-v2-fixed-final.png`: **95,413 / 2,617,920 pixels (3.6446%) differ**, 11,915 with luminance delta >16, maximum channel delta 213. Of these, 95,407 are above y=400; six are gesture-pill antialiasing below. The status bar, header text/icons, and top edge of the content surface have different vertical placement. Relaunching through the documented explicit Metro URL leaves 95,514 changed pixels. This is a reproducibility finding, not evidence that this PR changes header styling. An unrecorded native inset/status-bar condition is a possible explanation, not established. Record/pin that condition or provide a reproducible replacement pair before accepting the pixel proof.

   The reported environment JSON matches **exactly**: Android 16/API 36, OS fingerprint `google/sdk_gphone64_x86_64/emu64xa:16/BE2A.250530.026.F3/13894323:userdebug/dev-keys`, RN 0.88.0, logical 411.4286×923.4286, DPR 2.625, font scale 1; physical PNG 1080×2424, light. Cached native client fingerprint is `971743703b097b8870b39ff842d130f50b115eeb`; agent-device 0.21.32, pooled Pixel 9, real Gboard, server clock 2026-11-02T20:00:00Z, T3CODE_DEV_SEED=0. Rebuild received `{"type":"extension_ui_response","id":"editor-1","value":""}` and its doubled final screenshots are identical. The behavior passes; the independent rendering comparison fails.

   [Full delivered/rebuild pair](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/10686f1ffbd38dec/rebuild-A6-pair.png), [zero-threshold difference](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/13336c0dcd6b5ec3/rebuild-A6-diff0.png), [recording](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/74321a0d660c4a6e/after-A6-fixed.mp4), [environment](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/253908de3421272f/android-environment.json), [pixel measurements](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/16f34926f280966a/comparison.json).

   ![Delivered A6 at left and independent rebuild at right: the header and content edge have different vertical placement](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/63574e4838b4b81f/rebuild-A6-header-comparison.png)

2. **The live PR description remains stale.** [#29420](https://git.cafe/pingdotgg/t3code/pull/29420) still presents before [701e13cebb](https://git.cafe/pingdotgg/t3code/commit/701e13cebb040d8433cd9aefb803ddd0d58c41d6), after [e799b6a1eb](https://git.cafe/pingdotgg/t3code/commit/e799b6a1eb2228c2e4080f9d70329c1c91761d7e), earlier F rows, and says scripts were not uploaded. It does not link the supplied new index. Its narrower “only clocks, caret phase, key previews and timing” list does not describe the new pulse/shadow/summary differences. Replace this old proof block with the current index; old-revision attachments are not accepted current-head evidence.

3. **Minor factual correction:** the new index says A1 composer controls differ at three pixels. Recalculation finds **six RGB pixels**, all tiny gesture-pill differences and none above luminance threshold 16. This does not invalidate A1 behavior.

Reviewed [head f16d8770bd](https://git.cafe/pingdotgg/t3code/commit/f16d8770bda7acff940ed18a3557c050aea6651e) against [merge-base ac8dce0cc3](https://git.cafe/pingdotgg/t3code/commit/ac8dce0cc3abd8a337f2d806b5082a6834b9b81a), using the [new evidence index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/92b26a8bbddc5026/evidence.md) and [matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f1a3ac51a745c78b/fixture-matrix.md). The worktree and live PR metadata match the requested head; the worktree remains clean. Only QuestionAttachments text synchronization changes; no style/layout edit.

| Row | Status | What the accepted recordings show |
|---|---|---|
| A1 | SUFFICIENT | Three baseline question corruptions; all three head trials and both composer controls finish `Teh cat sat on the mat and went hoem`. Native key taps and autocorrection are visible. Slower measured cadence is disclosed and does not invalidate the reproduced failure. |
| A2 | SUFFICIENT | Both establish `Hello world`; baseline middle edit becomes `Hyzxello world`, head `Hxyzello world`. Prediction acceptance followed by xyz remains ordered on head. Final adjacent prediction/Submit taps resolve to the exact visible/provider values `Jell-O cuz world` and `Jell-O cuz I world`. |
| A3 | SUFFICIENT | First and Second each clear the focused input; typing deselects the option. Retyping with Hello suggestion yields baseline `Hello ct`, head `Hello cat`. Both reverse transitions are visible. |
| A4 | SUFFICIENT | Accepted baseline v2 and head chunks show PNG add/remove with draft intact, focused collapse/expand, suggestion/continuation, another thread and restoration, resolution, then a distinct empty request. Picker readiness retries do not hide the successful round trip. |
| A5 | SUFFICIENT | Both begin with multiline prefill. Head accepts suggestion, inserts newline, survives collapse/expand and continues. Its exact receipt is `"  Propo Helo catend\nsed commit\n\nDetail  "`; the baseline visibly displaces insertion and receives `"  Propo helosedcatend\n  commit\n\nDetail  "`. Spaces/blank line are preserved. |
| A6 | INSUFFICIENT — independent rendering comparison | Delivered before/after videos and receipts do show typing, prediction, newline, Select all/Delete, empty collapse/expand, and `value:""` rather than cancellation. The independently rebuilt final frame has the material header-placement delta in finding 1. |
| I1 | SUFFICIENT | Apple prediction, continuation, First clear, retype, retained-card collapse/expand, continued typing and immediate Submit are visible. Both authoritative/provider answers equal final `Help cat`. |
| I2 | SUFFICIENT | Mid-text prediction/newline, retained-card collapse/expand, continued edit and exact receipt `"  Proposed help cat\nLinecommit\nEnd\nDetail  "` on both sides. A subsequent distinct editor-2 starts empty, is typed/deleted through native keys, stays empty and submits `value:""`. |

All eight accepted rows have before/after columns labelled with the requested SHAs. Android is light portrait 1080×2424; iOS recordings are native 1206×2622, stills normalized to the declared logical 402×874, light. All 22 standalone selected source PNGs equal their corresponding sheet halves. I visually inspected all 11 selected sheets, all 22 diff images, and ffmpeg frames across all 18 accepted recording segments, with denser 2–3 fps extraction around immediate-submit/clear transitions. Accepted narration matches the visible steps and receipt values. No selected still has ghosted elements, stale overlay, spinner/skeleton, blank/unrendered answer, or unexplained changed region.

All 22 diffs regenerate pixel-for-pixel using the stated RGB absolute ×255 and luminance >16 methods. The new “what else changed” list accounts for text/caret/predictions, summary pulse, A2 summary width and adjacent prompt/time movement, keyboard/gesture antialiasing and iOS battery/shadow. A5/A6 delivered final pairs have zero changed pixels. I1/I2 have no differences over threshold 16. These statements concern the supplied pair, not the failed independent A6 comparison.

I also downloaded all supplementary archives and looked at 149 additional unique raw images and frames from 19 additional unique videos, plus the three directly linked failed A4 chunks. Those include the disclosed invalid Metro overlay, incomplete Pi sessions, early A2 setups, and A4 picker retries. They are excluded from accepted row selection; replacements above carry the behavior proof. Sampling spans each recording; it is not a claim of examining every encoded frame. Old PR attachments at earlier revisions are excluded from current-head acceptance.

Independent rebuild used the uploaded seed/environment/fake providers, inspector/navigation helpers and Android driver, with scratch paths/session changed. The only functional driver adjustment aliases multiline Gboard `Enter` where the script hardcodes `Done`; early attempts stopped before recording. No application source edits, direct input injection, live database access, pushes, comments, or PR mutations. Own isolated server and Metro were stopped by confirmed port-owner PIDs 569764 and 569782; their sessions exited, both ports are clear, and pool session `pr29420-proof-audit-2` was closed. Lead worktree and parent scratch were left intact. Own scratch is removed after preservation upload.

FRICTION:
- tool | Python urllib received HTTP 403 for public artifact URLs; curl downloads worked | use curl for the evidence downloader.
- runbook | supplied driver assumes Gboard Done while the multiline editor exposes Enter; two attempts aborted before capture | derive an Enter/Done alias from the native keyboard snapshot.

VERDICT: INSUFFICIENT A6