# Proof audit pass 2

Findings, in order of remaining proof impact:

1. **Independent recreation has a non-antialiasing pixel delta.** Rebuilt the committed eight-thread fixture in a fresh isolated home and ran the finished after artifact on a separately owned iPhone17e/iOS26.5 session, 390×844 points, 1170×2532 pixels, DPR3, Large Dynamic Type, light t3-code palette, normalized status bar. The recreated F1 Git menu has enabled View PR and Clean · PR42, but differs from the delivered after image at **157,364 pixels above zero; 42,899 pixels (1.4480968%) at max RGB-channel delta >8/255**, without masks or alignment. Menu content/edge is approximately four physical pixels left; battery/glass pixels also differ. Behavior reproduced; pixel equivalence did not. This is a finding under the requested independent-rebuild criterion, not evidence of an implementation regression.
2. **Required matrix coverage remains incomplete:** S6 wide controls/measurements/truncation; S7 Android states; R1 final return; R2 after-Cscope row selection and Android routing; R3 Android width/compact transitions. Disclosure makes the evidence honest but does not complete those rows.
3. **Misleading narration was corrected by superseding it.** The original captioned index overstated several timestamps, themes and baseline steps. The audited corrected index uses ten raw clips with accurate written checkpoints and explicitly withdraws the baseline early-state sequence. The old baked-caption clips have not passed this audit and must remain superseded. Underlying observed routing/list behavior is reported separately below.

Audited final evidence: [corrected index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b4d1f6999779670f/index-corrected.md). Downloaded public Markdown is byte-identical to `/tmp/17716-native-proof/index-corrected.md`. Its 28 four-column pairs are unchanged from the original final set. Independently recomputed every Diff0 (absolute RGB ×4) and DiffT (max channel >8/255 binary mask), with no masks/alignment: all images/counts match. Downloaded all 112 public four-column images and all ten public raw videos; every file byte-matches its corresponding local asset. No captioned-video link remains in the corrected index.

| Matrix row | Status | Checked result / missing part |
|---|---|---|
| S1 | Captured; recreation finding above | F1 thread Git menu and overview, light/dark: PR42 status, enabled View PR, retained L42; before lacks fallback. Correct F1 underlay on both sources. |
| S2 | Captured | Separate Review Git menu, light: disabled Push becomes View PR42. |
| S3 | Captured | F2 pending/Waiting for host state; F3c red closed; F3m purple merged; F0 no links. Terminal disabled Create PR verified in raw recording at 00:34 and 01:01–01:02. |
| S4 | Captured | F4 checkout PR3 wins; linked42 stays visible. Both baseline and after reach native3. Focused identity test passes. |
| S5 | Captured | Ccap both list callers light/dark: false empty/omitted becomes four rows with correct count and no mutation controls. Genuine read-only Cscope retains both lists and after View PR43. |
| S6 | Partial | iPad inspector/list light/dark captured at 1032×1376 points. Missing F1 C+ wide control, instrumented pane/header widths, exact1440 and known-long-title wide truncation proof. Captured F1 wide uses Ccap/known-open. |
| S7 | Unverified | Android linked/no-link states, accessible header/overflow action, light/dark. |
| R1 | Partial | All three changed F1 Git callers route42; F4 routes3. Earlier returns recorded; final return after Review's last42 destination absent. |
| R2 | Partial | Ccap list/context/row42 selection; Cscope list/context/Git View PR43; F6 stack42→43. Missing after-Cscope list-row selection and Android single/multiple/stack routing. |
| R3 | Unverified | Android measured <600/≥600dp header widths, keyboard/compact/maximal-sibling overflow transitions. |
| R4 | Captured | After raw starts F1 at00:00–00:02, F2 menu by00:03, later terminal/F0/F4 switches with checkout3 route. Baseline later F0/F4 switching/route3 shown; individual baseline states supplied by stills. No timing claim. |

Six rows captured, three partial, two unavailable. Pixel recreation is a separate finding.

Authoritative narration checkpoints below were checked by decoding from the beginning against media timestamps, sampling raw clips twice per second. Times are rounded stable visual checkpoints, not exact action-start/latency measurements. Fast input seeking on these variable-frame-rate files returned stale frames and was excluded. Full ten-file recommendations, including every raw absolute path, are retained in [narration recommendations](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/16159c9959ecff59/17716-proof2-narration-recommendation.md), local `/tmp/17716-proof2-narration-recommendation.md`.

| Absolute raw local path | Decoded observation correcting the old narration |
|---|---|
| `/tmp/17716-native-proof/videos/after-core-light.mp4` | 00:15 View PR with initial Clean; 00:40 settled Clean · PR42. 02:15/02:20 Review menu closed; 02:25 menu open; 02:35 native42. Ends at destination; no final return. |
| `/tmp/17716-native-proof/videos/after-scope-stack-light.mp4` | 00:15–00:25 Copy/Open-only context; 00:30 Git View PR menu; 00:32–00:34 native43; 00:37 Threads; 00:40/00:45 stack. The route is not between00:15–00:30 and is not a list-row selection. |
| `/tmp/17716-native-proof/videos/after-denied-list-light.mp4` | 00:10 dark overview; 00:15/00:20 light dedicated list; 00:25 dark list; 00:30/00:40 light list; 00:45/00:55 context; 01:00 native42 loading; 01:05 error. Filename does not mean exclusively light. |
| `/tmp/17716-native-proof/videos/after-denied-light.mp4` | 00:00–00:04 Threads, not Git menu; 00:06 F5 thread; 00:09 disabled Push/Create PR overview; 00:12/00:15/00:25 four retained rows. |
| `/tmp/17716-native-proof/videos/before-states-light.mp4` | Early F1 idle then F3M idle, not the claimed early F1→F2/closed/merged sheets; long pause. 03:25 F0 overview; 03:45 F4 menu/checkout3; 04:00 native3. |
| `/tmp/17716-native-proof/videos/before-core-menu-light.mp4` | Supplemental read-only baseline, not matched C+. 00:00–00:07 disabled Push; about00:10 read-only overview. Entire clip approximately14.6s, not50s. |
| `/tmp/17716-native-proof/videos/after-states-light.mp4` | Closed disabled Create PR00:34; merged01:01–01:02. Retry banners occur; no uninterrupted-connectivity or latency proof. |

The corrected index incorporates these checkpoints accurately. Before/after native destinations identify synthetic `example/fixture` and the number; their expected detail-load errors do not verify loaded remote PR content.

Visible changes checked against all static pairs: PR42 status suffix; View PR replacing disabled Push/Create PR; disappearing no-local-commits subtitle; enabled icon/text contrast; shorter native menu with following rows shifted upward; retained four-row Ccap sections replacing omitted/empty states, with counts/icons/titles/separators and changed occupied height. Cscope preserves links/context actions and gains readable View PR43. F2 pending, terminal colors, F0 absence, checkout3 priority and F6 stack shape remain. Full-frame differences include glass/antialiasing, scroll/label/underlay position variation, plus transient retry banners in video. They cannot all be attributed to the patch. F5 Ccap does **not** prove View PR: latest pending43 is selected and its fallback is rejected with capabilityfalse. Known-open42 does not override it.

Sources are before [8acbefefc5](https://git.cafe/pingdotgg/t3code/commit/8acbefefc51b4d281dcd04108c82baa1ab902b0e) and after app [9c3a6ee0fe](https://git.cafe/pingdotgg/t3code/commit/9c3a6ee0fefce393799bc616d193747e2ddb2208). Independently read FINISHED build metadata matching those SHAs: before9bc0517a-6c8d-4231-9c40-64288a247b56 and afterb740cbf8-b29c-4c28-a30c-f1e26a70a568. Later fixture cleanup [a5a3390a93](https://git.cafe/pingdotgg/t3code/commit/a5a3390a93b876737abc4f554a3ba61f0f6efb23) and knip entry registration [af614de1fb](https://git.cafe/pingdotgg/t3code/commit/af614de1fbd1f58b5b54d101dcfbcffd359bda69) do not change captured app source. Current source checkout is clean. Source review and focused `use-thread-pr.test.ts` run: 12 tests passed. No concrete exploitable vulnerability or implementation regression identified by this audit; proof insufficiency is not a recommendation to expand the fix's implementation scope.

Coverage limits: real native iOS26.5 phone1170×2532px/390×844pt DPR3 and tablet2064×2752px/1032×1376pt DPR2; phone font categoryLarge, numeric scale not exposed; tablet font category not separately read. Inspector width approximately282pt is pixel-derived, not instrumented. Stills use normalized9:41; raw recordings show actual changing clock. Android, exact1440pt, older-iOS nonglass toolbar and loaded remote PR content remain unverified. Fnull is a logic control, not an omitted native timing demand. Watch/Unlink/Link mutations were intentionally not executed. Web/desktop/provider/server/contracts paths are unchanged.

Independent recreation used committed fixture scripts in fresh `/tmp/17716-proof2-runtime/.t3/repro`, synthetic read-only database backup plus reseeding, independent ports/proxy/tunnel and finished after app. Never used live `~/.t3` state. Owned EAS session01a161cb-25e3-7a7a-b6ef-062fe8d0faa3 returned STOPPED. Tracked backend1405410/proxy1406560/tunnel1406561 process groups stopped and absent; managed simulator credentials cleared. No new runtime/captures after cleanup.

![Left: delivered after F1 menu. Right: independent recreation, same native geometry.](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ec82eeb80ebe6390/recreation-pair.png)

[Recreation Diff0 ×4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/23d11fdb82016b21/recreation-diff0-x4.png) · [Recreation DiffT >8/255](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/84464c1929e5636b/recreation-difft-8.png) · [Pixel statistics](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3747a65c64a21303/recreation-stats.json) · [28-pair recomputation](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3d663bb8b1846b47/17716-proof2-diff-verification.json).

FRICTION:

tool | Fast seeking on raw VFR media returned stale frames; original coarse contact-sheet labels were offset. | Decode sequentially against media timestamps before writing narration.
brief | Evidence index changed during audit and old captions overstated visible checkpoints. | Freeze the manifest and write only checkpoints observed in raw media.
infra | Android/exact1440/older-iOS paths were unavailable. | Declare concrete native coverage limits before capture planning.

VERDICT: INSUFFICIENT S1 independent pixel recreation; S6 wide controls/measurements/truncation; S7 Android states; R1 final return; R2 after-Cscope row selection and Android routing; R3 Android width/compact transitions