# Current-head evidence audit pass4 — findings first

**INSUFFICIENT for the full CURRENT-HEAD evidence contract.** No implementation defect established. Completed portions are useful, but neither old-head source equality nor old R4 proof fills missing current-head media.

1. **Three R4 cells lack complete current-head evidence:** R4-1440-dark has only an interrupted initial screenshot and no accepted continuous clip; R4-390-light and R4-390-dark were not captured. All three require the original full R4 chain and paired stages; the dark initial alone is not a passing cell.
2. **R2-1440-light lacks a settled final endpoint.** Real star/Favorites/select/unstar/stale fallback is present in its continuous clip, but the final screenshot catches the outgoing popup transition: 28,333 pixels differ at threshold0,16 at max-channel>16. This is disclosed and not a product defect. The strict settled/matched endpoint requirement remains unmet; retain/repeat the full continuous sequence through a settled final result.
3. **The required independent CURRENT-HEAD R4 desktop-light rebuild is blocked and was not performed.** At handoff, the authorized worktree had no Git metadata/source, and its current-head userdata had no SQLite database or model manifest. The capturer reports external deletion mid-run at10:11UTC, with `ENOENT: process.cwd failed ... uv_cwd`. I independently confirmed the missing paths. No source/state was recreated or borrowed from another workspace. A fresh blank native preview tab was prepared at1440×900, but no server was started and no replacement screenshot/diff is claimed.
4. **Final source/state revalidation is unavailable after deletion.** Before capture, I independently verified the new checkout resolved to the target full SHA with clean tracked state. Complete five-file diff and merge-base were read and saved before deletion. New artifacts carry target SHA/capture times and ordered native events, but final Git cleanliness and pinned-seeder/runtime-state inspection cannot be repeated. Preserve this qualification.
5. **Copied catalog narration needs correction:** new narrow R3 video visibly reports10 genuine OpenCode models; inherited prose says11. Big Pickle remains genuine and is starred before disable, and both sides' compared disabled endpoints have0 reported models. This does not require an invented catalog or extra permutations. Use the observed count for new clips.

TARGET: [377dd251be](https://git.cafe/pingdotgg/t3code/commit/377dd251be942831904f75c29a63d85c6b22d0cd), base [3b95e7831b](https://git.cafe/pingdotgg/t3code/commit/3b95e7831b11b17627dff16843bc087bfca397ec), [#23479](https://git.cafe/pingdotgg/t3code/pulls/23479), [GitHub mirror #23479](https://github.com/pingdotgg/t3code/pull/23479).

The complete five-file diff was inspected before new captures. [Current matrix decision](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bc5b863cc2c7c9bb/23479-current-matrix.md): R1–R4 ×1440/390×light/dark still applies, with separate R2 empty controls. No new permutations. Web only; no Electron/native mobile changes or verification. Seeder parsing/mkdir and Knip are the only intervening changes, but source equality does not certify new media.

Preserve valid historical [pass3 proof](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/03429e68ef503627/rebuild-R4-1440-light.png) as OLD HEAD b065ad799c supporting evidence. Its independent rebuild was zero-delta; it is not current-head proof. Old after screenshots/videos were not relabeled.

Parent alone preserves the original body and appends evidence; current head/CI reporting belongs to parent. This audit made no implementation edits, commits, branches, pushes, comments, PR edits, watches or PR polling. App association only was registered as required. No tests or broad checks were run. GPT-6.1-Sol through Codex/T3 Code; test-t3-app and ghdrop skills.

## Row table

PASS below means the delivered cell behavior and endpoint evidence passed this bounded audit. It does not certify the missing independent rebuild or erase the overall INSUFFICIENT verdict. All before sides remain exact BASE; all accepted new after sides are target head. Numbers are independently recomputed changed pixels at threshold0/>16.

| Cell | Current-head cell status | Paired stages (0/>16) | Behavior / exact gap |
|---|---|---|---|
| R1-1440-light | PASS | initial 69871/19437; reopened 70122/19280; result 0/0 | Fallback, selection, reopen, search/clear and keyboard. |
| R1-1440-dark | PASS | initial 86022/3488; reopened 84562/6077; result 0/0 | Fallback, selection, reopen, search/clear and keyboard. |
| R1-390-light | PASS | initial 375350/122323; reopened 359517/101313; result 0/0 | Fallback, selection, reopen, search/clear and keyboard. |
| R1-390-dark | PASS | initial 360788/58071; reopened 344995/61947; result 0/0 | Fallback, selection, reopen, search/clear and keyboard. |
| R2-1440-light | INCOMPLETE | initial 0/0; stale-reopened 69871/19437; result 28333/16 | Functional chain present; final settled/matched endpoint missing. Separate empty control passes. |
| R2-1440-dark | PASS | initial 0/0; stale-reopened 86022/3488; result 0/0 | Genuine star; Favorites once; select/unstar; stale-only reopen. Separate empty control passes. |
| R2-390-light | PASS | initial 0/0; stale-reopened 375420/122393; result 0/0 | Genuine star; Favorites once; select/unstar; stale-only reopen. Separate empty control passes. |
| R2-390-dark | PASS | initial 0/0; stale-reopened 360788/58071; result 0/0 | Genuine star; Favorites once; select/unstar; stale-only reopen. Separate empty control passes. |
| R3-1440-light | PASS | initial 94005/16523; result 0/0; picker-recovery 0/0 | Real Big Pickle star→disable;0models; missing unstar→reload/revisit→usable picker. BASE removal BLOCKED. |
| R3-1440-dark | PASS | initial 92641/17063; result 0/0; picker-recovery 0/0 | Real Big Pickle star→disable;0models; missing unstar→reload/revisit→usable picker. BASE removal BLOCKED. |
| R3-390-light | PASS | initial 149155/42577; result 0/0; picker-recovery 4644/71 | Real Big Pickle star→disable;0models; missing unstar→reload/revisit→usable picker. BASE removal BLOCKED. |
| R3-390-dark | PASS | initial 147559/42020; result 0/0; picker-recovery 4612/71 | Real Big Pickle star→disable;0models; missing unstar→reload/revisit→usable picker. BASE removal BLOCKED. |
| R4-1440-light | PASS | initial 41677/10881; removed 684/174; result 0/0 | Missing+reported once; no missing custom CRUD; missing unstar; real unstar/restar; reload. Filter N/A8models; BASE missing removal BLOCKED. |
| R4-1440-dark | MISSING/INCOMPLETE | No accepted paired stages | Initial screenshot only; complete continuous chain and stages missing. |
| R4-390-light | MISSING/INCOMPLETE | No accepted paired stages | No current-head capture; full chain and stages missing. |
| R4-390-dark | MISSING/INCOMPLETE | No accepted paired stages | No current-head capture; full chain and stages missing. |

## Independent checks and qualifications

The complete current diff is limited to the two web components, committed fixture/seeder and Knip manual-seeder entry. The earlier R1–R4 matrix still covers it. Web composer and Settings paths apply; Electron/native mobile are outside this explicitly web-only audit. No server/contracts/provider-adapter feature expansion or extra fixture zoo was invented.

The completed PARTIAL delivery was audited only after its handoff report, `/tmp/23479-current-head-report.md`. It contains13 complete main after clips plus4 empty-control after clips, not16 complete main cells. All17 new clips were independently decoded, their full-clip sampled contact sheets visually inspected, and short R1 search/clear phases inspected at8fps. R4 control sequence was additionally decoded at8fps. This is sampled playback review, not a claim that every frame was manually watched. The initial incomplete dark R4 screenshot is excluded from passing comparisons. Its presence cannot certify a recording.

All69 declared new artifact hashes and viewport metadata check;69 includes the incomplete dark R4 initial screenshot. All47 BASE endpoint copies match the historical exact-base artifact hashes. All20 reused BASE recording hashes match. No old AFTER clip is reused as new-head proof. All47 delivered full-frame pair diffs independently reproduce absRGB×8 clipped255 and binary maxRGB>16, with matching counts/hash bytes. There is no all56-stage claim. No masks, cropping, alignment or normalization were applied to comparisons. Contact-sheet resizing/cropping is only an inspection aid, never the comparison input. All13 available main-cell paired-stage sheets and the four supplemental control pairs were inspected.

Metadata measures CSS1440×900 or390×844, DPR2, root/body16px, zoom1, visualViewport scale1, explicit theme/root class. Desktop PNG1280×800 and narrow780×1688 exports are distinguished from CSS dimensions. Chat sidebar/details are closed, same blank server/New thread composer, correct per-stage Sol/Low or Astra/Medium framing, neutral endpoint pointer20,80. Endpoint window scroll matches. Actual Settings overflow-element scroll observations are retained in the diff archive: R3 initial added row changes card height and natural bottom-clamped inner scroll (desktop initial248 versus after-removal226), as expressly disclosed. Persisted R3 result pairs are0/0; recovery desktop pairs0/0; narrow recovery retains a caret plus shadow/raster deltas,71pixels>16 on each theme. This is accounted for, not silently excluded. R4 desktop-light removed-stage scrollbar/shadow paint is retained; persisted result is0/0. R2 empty controls are independent stopped-server resets, not spliced into the main clips; stale preference removal is crosscovered by R3.

The original body remains byte-identical locally:36085bytes, SHA256 `6d7a4faecdc583bf6f9438836552e432765ce16de8bd8f32f45201be1781dc9f`. No remote PR body/CI polling or editing was performed. Native app PR association is registered and not watched. Parent alone appends an original-preserving evidence body and reports current head/CI.

[Actual decoded playback supplement](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9048095193762b3a/playback-supplement.md) · [Original-frame PTS JSON](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7cd4206982b70a6d/playback-pts.json) · [All47 full-frame diffs/masks + scroll/reuse records](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/79a6994166f718d5/full-frame-diffs.zip) · [Recomputed metrics](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f35994bd42cb2380/recompute.json) · [Rebuild blocker record](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/cf159b081450b399/rebuild-status.json) · [278 published URLs checked](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/09612885db864f26/url-audit.json). All278 loaded; applicable media hashes matched.

Timing is supplemented without recapture: actual decoded frame PTS samples for all17 NEW clips are in the linked table; untouched BASE clips retain only their qualified BASE rows from pass3's supplement. No old AFTER timing is reused. Samples are visible-state timestamps, not exact click/onset times or UTC conversion. Missing clips remain missing.

## FRICTION

- WORKSPACE / EXTERNAL DELETION: authorized source and isolated database/model-manifest disappeared mid-capture. Independently observed; prevents faithful current-head rebuild and final source-clean check. No guesses about the actor or cause.
- CAPTURE COMPLETENESS: three R4 cells missing/incomplete; R2 desktop-light settled final endpoint absent. Exact gaps above; no extra permutations requested.
- TIMING: wall-clock offsets are not media seeks. Repaired for available new clips using decoded original-frame PTS; no timing-only recapture requested.
- NARRATION / DATA: inherited11model OpenCode prose differs from new narrow10model catalog. Preserve actual observation; zero-model disabled endpoints and genuine Big Pickle action still verified.
- RENDERING: R3 changed height/clamped scroll and narrow caret/shadow; R4 scrollbar/shadow; R2 outgoing transition. Quantified and disclosed; only the unsettled R2 final violates the settled-endpoint requirement.
- RUNTIME / TEARDOWN: capturer's own8595/16635 ports independently confirmed closed. Auditor started no server, seeded no state and captured no new app endpoint; older6401/7564 and other environments untouched. The fresh blank native preview tab remains blank. No live userdata used.
- TOOLING: new partial manifest differs from prior schema; audit normalization lives only under `/tmp/23479-audit4`, not in delivered media or source. No implementation edits or package rewrites.

**Final verdict: INSUFFICIENT.** Required remedies are exactly three complete current-head R4 cells, a settled R2-1440-light final sequence, and one independent current-head clean R4 desktop-light rebuild after restoring authorized exact-head source and faithful isolated state. Final source/fixture provenance must be revalidated then. Preserve old R4 proof as supporting OLD HEAD evidence. No implementation remedy is established by this evidence audit.