# Final proof audit round3 — findings first

**SUFFICIENT**, for the corrected final delivery **together with this audit's [decoded playback supplement](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5c666bec5c0268c8/playback-supplement.md)**. All16 functional cells pass. No implementation bug found in the complete four-file diff. No additional recapture is required. Parent should append/link this supplement alongside the corrected index; the capture timing document alone contains UTC/wall-clock elapsed values, not media seek positions. This is the only remaining publication handoff, not missing evidence.

Audited [#23479](https://git.cafe/pingdotgg/t3code/pulls/23479), [GitHub mirror #23479](https://github.com/pingdotgg/t3code/pull/23479), base [3b95e783](https://git.cafe/pingdotgg/t3code/commit/3b95e7831b11b17627dff16843bc087bfca397ec), head [b065ad79](https://git.cafe/pingdotgg/t3code/commit/b065ad799c5bfda2027d7d68327bb6bcd1c49666). Exact local head rechecked, clean tracked worktree. This is an evidence audit, not a fresh CI/merge-readiness verdict; pending/current checks belong to parent and were not polled.

The original description is preserved: `/tmp/23479-original-body.md`, SHA256 `6d7a4faecdc583bf6f9438836552e432765ce16de8bd8f32f45201be1781dc9f`,36085bytes. I made no PR/body/comment/source/branch/commit/push mutation. Old R1/R2/R3 media remain historical; claims below use corrected replacements only. Retained R4 evidence remains current and immutable.

1. **Background confound resolved.** Every corrected R1/R2 chat endpoint and R3 picker-recovery endpoint has sidebar and details closed on both sides. The same blank server/New thread draft, project label, per-stage composer model/reasoning, mode, pointer and scroll are visible. No project counts/highlights/relative ages or clock remain. R1 initial Sol/Low and reopened/result Astra/Medium match within each pair. Desktop Settings retains its256CSS-pixel navigation column.
2. **Timing repaired without recapture.** Independently decoded all40 final MP4s, checked their hashes/durations, and visually reviewed each full-clip contact sheet. Supplement provides actual playback phase samples for all40, including separate empty controls and retained R4. R1 was sampled at8fps; others2fps, with R4's control interval additionally8fps. These are visible-state positions with stated sampling resolution, never fabricated exact click times or a UTC offset conversion. Original UTC remains separately available.
3. **Nonzero differences disclosed.** Narrow R2 initial70-pixel caret differences; R3 card-height/clamped-scroll effects, light-desktop256-pixel separator and narrow recovery shadow differences; R4 removed desktop scrollbar/separator/shadow differences are expressly disclosed. I inspected all16 full-frame pair sheets and independently recomputed every56paired-stage threshold0 amplified diff and maxRGB>16 mask. All counts, bounding boxes and hashes match. No crop, image registration, masking or pixel exclusion was used for comparisons. Zero claims apply only to computed zero pairs.
4. **Independent clean rebuild passes exactly.** One new R4 head endpoint using the committed seeder/fixture and isolated stopped-server state matches the retained desktop-light head initial screenshot with0pixels at threshold0,0at>16, maxRGBdelta0. No provider-update toast was present before capture. Historical independent R4 endpoint credit (199delta1pixels,0>16) remains valid and qualified; this fresh exact match supplements it. The prior pass2 toast mismatch is not substituted as proof.

Corrected [index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6fdd8bf5732f247e/index.md), [appendix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0f741240dd1618a7/appendix.md), [manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/705f4dfee9c11d6a/manifest.json), [WHAT ELSE CHANGED](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a2ea794abf306cc1/what-else-changed.md), [separate UTC/wall-clock timings](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/05f6168a9d43f237/timings.md), and [preserved R4 endpoint proof](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/89c0fa54012076db/23479-preserved-r4-links.md).

## Row-by-row matrix result

Each row includes both before/after continuous clips and all named paired endpoints. Numbers below are independently verified changed pixels at threshold0/>16, in stage order. Full images and per-cell measurements/data provenance are in the manifest/index; playback positions are in the supplement.

| Cell | Status | Paired stage differences (0/>16) | Behavior/control qualification |
|---|---|---|---|
| R1-1440-light | SUFFICIENT | initial 69871/19437; reopened 70122/19280; result 0/0 | Initial/reopened stale-only Favorites: baseline empty, head Codex fallback. Search, clear, keyboard selection and reopen recorded; result matched. |
| R1-1440-dark | SUFFICIENT | initial 86022/3488; reopened 84562/6077; result 0/0 | Initial/reopened stale-only Favorites: baseline empty, head Codex fallback. Search, clear, keyboard selection and reopen recorded; result matched. |
| R1-390-light | SUFFICIENT | initial 375350/122323; reopened 359517/101313; result 0/0 | Initial/reopened stale-only Favorites: baseline empty, head Codex fallback. Search, clear, keyboard selection and reopen recorded; result matched. |
| R1-390-dark | SUFFICIENT | initial 360788/58071; reopened 344995/61947; result 0/0 | Initial/reopened stale-only Favorites: baseline empty, head Codex fallback. Search, clear, keyboard selection and reopen recorded; result matched. |
| R2-1440-light | SUFFICIENT | initial 0/0; stale-reopened 69871/19437; result 0/0; empty-control-initial 0/0; empty-control-result 0/0 | Fresh real favorite wins; real unstar returns stale-only case. Empty reset control is a separate continuous clip on each side; stale removal crosscovered by R3. |
| R2-1440-dark | SUFFICIENT | initial 0/0; stale-reopened 86022/3488; result 0/0; empty-control-initial 0/0; empty-control-result 0/0 | Fresh real favorite wins; real unstar returns stale-only case. Empty reset control is a separate continuous clip on each side; stale removal crosscovered by R3. |
| R2-390-light | SUFFICIENT | initial 70/70; stale-reopened 375420/122393; result 0/0; empty-control-initial 0/0; empty-control-result 0/0 | Fresh real favorite wins; real unstar returns stale-only case. Empty reset control is a separate continuous clip on each side; stale removal crosscovered by R3. |
| R2-390-dark | SUFFICIENT | initial 70/70; stale-reopened 360788/58071; result 0/0; empty-control-initial 0/0; empty-control-result 0/0 | Fresh real favorite wins; real unstar returns stale-only case. Empty reset control is a separate continuous clip on each side; stale removal crosscovered by R3. |
| R3-1440-light | SUFFICIENT | initial 91295/16522; result 256/256; picker-recovery 0/0 | Genuine OpenCode enable11models → real Big Pickle star → disable0models. Head removes saved row, reloads/revisits. Baseline removal BLOCKED; manual Codex recovery explicitly qualified. |
| R3-1440-dark | SUFFICIENT | initial 92636/14793; result 0/0; picker-recovery 0/0 | Genuine OpenCode enable11models → real Big Pickle star → disable0models. Head removes saved row, reloads/revisits. Baseline removal BLOCKED; manual Codex recovery explicitly qualified. |
| R3-390-light | SUFFICIENT | initial 149155/42577; result 0/0; picker-recovery 4574/1 | Genuine OpenCode enable11models → real Big Pickle star → disable0models. Head removes saved row, reloads/revisits. Baseline removal BLOCKED; manual Codex recovery explicitly qualified. |
| R3-390-dark | SUFFICIENT | initial 147559/42020; result 0/0; picker-recovery 4542/1 | Genuine OpenCode enable11models → real Big Pickle star → disable0models. Head removes saved row, reloads/revisits. Baseline removal BLOCKED; manual Codex recovery explicitly qualified. |
| R4-1440-light | SUFFICIENT | initial 9448/5106; removed 4045/1956; result 0/0 | Eight genuine Codex models, missing slug + real favorite once. Head missing unstar; real-model unstar/restar both; reload/revisit. Baseline missing removal BLOCKED; filter N/A8models. |
| R4-1440-dark | SUFFICIENT | initial 11773/5020; removed 3883/1622; result 0/0 | Eight genuine Codex models, missing slug + real favorite once. Head missing unstar; real-model unstar/restar both; reload/revisit. Baseline missing removal BLOCKED; filter N/A8models. |
| R4-390-light | SUFFICIENT | initial 13836/11244; removed 0/0; result 0/0 | Eight genuine Codex models, missing slug + real favorite once. Head missing unstar; real-model unstar/restar both; reload/revisit. Baseline missing removal BLOCKED; filter N/A8models. |
| R4-390-dark | SUFFICIENT | initial 13747/11261; removed 0/0; result 0/0 | Eight genuine Codex models, missing slug + real favorite once. Head missing unstar; real-model unstar/restar both; reload/revisit. Baseline missing removal BLOCKED; filter N/A8models. |

## Independent verification and provenance

Reviewed original matrix, pass1, entire pass2 + published links, original body, original capture report/index, preserved endpoint proof, and complete four-file diff before corrected delivery. Final delivery was declared ready in `/tmp/23479-recapture-report.md`; no capturer output was modified. No additional permutations were invented: unchanged locked/unavailable guards were not reimplemented or native-tested.

All124new screenshot/recording artifacts have matching hashes and measured per-cell metadata; retained32R4 screenshot/recording artifacts also match their original hashes/metadata. All40clip hashes pass. All56 full-frame paired-stage diff/mask bytes and metrics pass independent recomputation: [machine-readable result](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8b095bace143a23a/recompute.json). All332distinct published references loaded successfully; applicable media bytes match manifest hashes: [URL audit](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c777ba121926de18/url-audit.json). Published index/appendix/manifest/timings/whatelse bytes were additionally compared directly to the files audited.

Every cell records1440×900CSS or390×844CSS, DPR2, root/body16px, zoom1, visualViewport scale1, explicit theme/root class, source SHAs, isolated source data and actual endpoint scroll observations. Desktop screenshot transport is1280×800; narrow780×1688; these raster sizes are not mislabeled CSS dimensions. Settings scroll observations include actual overflow elements, not only window.scrollY. R3 differing clamped scroll follows changed content height and is disclosed; chat endpoints are consistently unscrolled. Both recapture states were copied from the same read-only snapshot, with independent state/server lifecycles. Fixture/seeder hashes match committed head. No source changes in desktop/mobile, so only responsive WEB applied; no native verification required. No provider turn was started.

The code diff is limited to the picker choosing Favorites only when a favorite is currently eligible; provider models appending noncustom synthetic missing saved-favorite rows; the fixture JSON and seeder. Existing real rows remain unique and settings unstar follows the normal persistence path. No implementation/source change was made during this audit and no broad checks were run.

## One independent rebuild this round

Original head ports7958/15998 were closed before seeding. From `/home/ubuntu/wt/pr-23479`, with its own isolated `.t3` home and server stopped, ran exactly `node scripts/seed-issue-23245.mjs /home/ubuntu/wt/pr-23479/.t3 R4`. Seeder SHA256 `a0165479cbd2436423b035bcdcd72b01d04f6419e6f8c3520717ef45cea96a6d`; fixture SHA256 `e0c1f803a44cadc4948e041e3f0c079aa66230a8b307050a0e8801e4f47c003e`.

Started own `vp run dev` session14711 at09:55:33UTC. Fresh owned native-preview tab, normal startup pairing; UI navigation Settings→Providers→Codex. CSS1440×900/light/DPR2/font16/zoom1/scale1. Actual Settings scroll array276and593 matched chosen retained artifact. Waited actual document fonts and paint, neutral pointer20,80, verified notifications empty. No masking, alignment, CSS injection, UI source edit or provider turn. Captured one saved endpoint; compared fullRGB against retained `R4-1440-light-after-initial.png`. Raster1280×800, threshold0=0,>16=0,maxdelta0,bboxnull, SHA256 `5697555c70bdba2bb2ce23eae77c50fb3ff31a95a02daf4f5c5518dbdc4ad0de`.

[Exact delta](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a937610daa68caaa/rebuild-delta.json), [full-frame amplified diff0](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/cfff7fea3439281f/rebuild-diff0.png), [full-frame >16 mask](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0a1f62dc74ed7bc8/rebuild-threshold16.png).

![Independent R4 head rebuild, light1440CSS, exactly matching retained initial endpoint](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/03429e68ef503627/rebuild-R4-1440-light.png)

Stopped only own session14711 withCtrl-C at09:58:48UTC; ports7958/15998confirmed closed. No original base/capturer/older6401or7564server was touched. No live userdata database write or server against live state occurred; browser screenshot storage is the harness artifact directory, not application state.

## FRICTION

- Evidence formatting: capture index's honest UTC/wall-clock timeline still needed decoded media positions. Resolved with the40clip supplemental table; no recapture.
- Data/framing: old sidebar confounds required corrected chat endpoints. Resolved by closed sidebar/details with matched blank draft; originals remain historical.
- Rendering: carets, separator/overlay-scrollbar paint, shadows and changed content's natural clamped scroll are nonzero and disclosed. They are not called zero or attributed to unrelated model data.
- Browser tool: one Settings locator used unsupported `exact`attribute; corrected immediately to supported snapshot locator. Fresh native preview then worked. No alternate browser or native surface used.
- Publication tooling: Python urllib received403for direct document byte checking; curl uses the working supported transfer path. No evidence URL was changed.
- Runtime: provider update notification absent at independent endpoint; no masking/dismissal fiction. Capturer clips retain real loading/notification transitions and endpoint cleanup.

Verdict **SUFFICIENT** for corrected delivery plus this playback supplement. Precise handoff: preserve original body and qualified original R4 endpoint, append corrected index + this audit/supplement links; parent handles pending CI. No further implementation or evidence remedy remains. Audit performed with GPT-6.1-Sol through Codex/T3 Code, using test-t3-app and ghdrop skills.