# Issue 16455 proof audit PASS2

**Findings first**

1. **Rebuild has a real non-AA delta; no pixel-identity claim is justified.** Rebuilt the HEAD 1440×900 light linked-panel worktree case in the existing isolated head environment, using the committed `fixture.json` and exact `seed-app.js`. Initial frame versus delivered HEAD: 131,858/5,184,000 pixels differ at threshold 0 (2.543557%); 1,166 exceed max-channel 16/255 (0.022492%). Settled refusal versus delivered HEAD: 1,952,219 differ at threshold 0 (37.658546%); 11,069 exceed 16/255 (0.213522%). Pointer visibility and composer caret/focus differ, beyond anti-aliasing. Flat-region/glyph noise also differs between separately encoded videos. Layout, fixture rows, model label, refusal text and wrapping match on visual inspection. This is a capture reproducibility finding, with no observed additional implementation regression. The seed retains client preferences and existing records rather than resetting an entire browser profile.
2. **C3 proves close-control clicks; it does not show a cleared final toast.** The original supplemental clip's last decoded frame, PTS 20.4997s, still contains the retry refusal toast. The first local-refusal close click is followed promptly by reload/reopening; a settled toast-free state before that reload is likewise not independently demonstrated. The lead has narrowed the current local PR body/index to close-control clicks and explicitly disclosed the final clearing limit. Required local/refusal/reopen/worktree-retry actions and both errors are present in continuous footage. A stronger claim that dismissal visibly completed would remain unsupported.
3. **Fresh app creation remains unvalidated.** Committed Python seeding succeeded independently at a new `/tmp/issue-16455-pass2-seed-check` root. Exact committed browser seeding succeeded against existing correct isolated projects/thread/link. That exercises reuse and relinking, not fresh `project.create` or `thread.create`. Rebuilding against the existing correct fixture is expressly permitted by this audit brief; this limitation does not invalidate the reproduced refusal.

**Verdict scope:** sufficient for the agreed four paired C1 recordings and head-only C2/C3 interaction sequence, with the narrowed close-click claim and disclosed reproducibility differences. No claim of pixel identity, visibly completed final dismissal, fresh app seeding, native or Electron coverage is approved.

Before [e075957](https://git.cafe/pingdotgg/t3code/commit/e07595714a467e18c750107137d2c2e50f0c3530); after [060ffe0](https://git.cafe/pingdotgg/t3code/commit/060ffe0849efaa64cf523c93f355a20891d4b14f). Read-only source audit; worktree remained clean at the pinned HEAD. Only isolated service seeding/relinking, browser-local state, new temporary fixture repositories and audit files were written. No source edits, pushes, comments, merges, servers or children were started.

**Row status**

| Scoped row | Status | What is actually supported |
|---|---|---|
| C1 1440×900 light | Sufficient | Continuous baseline/head checkout menu → worktree choice → unrelated draft → baseline materialization failure versus HEAD repository refusal. Approximate elapsed milestones agree with sampled decoded frames. |
| C1 1440×900 dark | Sufficient | Same sequence and terminal contrast. HEAD also shows the slow-request warning during the longer wait; no service-latency claim is warranted. |
| C1 390×844 light | Sufficient | Continuous paired sequence; full wrapped refusal and close/copy controls visible at HEAD. Baseline fails after the unsafe preparation attempt. |
| C1 390×844 dark | Sufficient | Baseline green Checked out result and changed worktree/branch controls versus HEAD refusal. Baseline Astra / HEAD Sol is a real, disclosed downstream visible difference, not an unexplained frontend-code change. |
| C2 local checkout, HEAD 1440 light | Sufficient | Original supplemental recording shows selecting local mode, unrelated draft and mismatch refusal. Runtime transcript independently rejects local mode. |
| C3 close/reopen/worktree retry, HEAD 1440 light | Sufficient for narrowed claim | Close-control interaction, reload/reopened linked panel, worktree retry and second refusal appear in one continuous recording. Final toast clearing is explicitly unverified. |
| Committed-fixture rebuild, HEAD 1440 light | Sufficient with finding | Same refusal reproduced using exact committed browser seed and existing correct isolated fixture; full-resolution pixel results above. No fresh-app-creation or pixel-identity claim. |
| Other callers | Source-traced; not client-tested | Panel Fix/Resolve, linked-row handoffs, prompt-link route, branch dialog and native outbox reach the shared preparation command/service. No new mandatory Cartesian visual rows inferred. |
| Electron; native iOS/Android; relay/tunnel/multiple environments; provider/MCP/scheduled flows | Unverified, disclosed | Responsive Chromium is browser proof only. No provider turn was started. |

**Evidence and independent checks**

Reviewed the final eight-entry manifest, nine original videos by sequential decoding and timestamped frame sampling, extracted initial/terminal images, raw diff assets, runtime/snapshot JSON, the corrected local index and reviewable PR-body draft. Older unfrozen clips, old failed index and superseded matrices were excluded. Missing native actionTimeline entries are not missing continuous recordings: frame-observed elapsed narration is acceptable when labelled approximate, rather than fabricated exact UTC click times. The supplied frame-observed C1 milestones checked against decoded PTS samples show the stated menus, preparation, drafts and terminal outcomes.

Independently recomputed threshold-0 and max-channel >16 statistics for all eight before/after milestones; every supplied count matched. The ×4 operation is saturated absolute RGB difference; threshold16 is a binary mask; no masks or geometric registration were applied. Display sheets resize panels, while comparisons use the full source-frame resolutions, 2880×1800 and 780×1688.

| Pair | Initial >0 / >16 | Terminal >0 / >16 |
|---|---|---|
| 1440 light | 0.020351% / 0.000212% | 11.310127% / 1.718904% |
| 1440 dark | 0.063850% / 0.001312% | 8.159587% / 0.464236% |
| 390 light | 3.445285% / 0.003038% | 63.618757% / 7.060396% |
| 390 dark | 1.047363% / 0.000608% | 26.603020% / 2.375744% |

The written visible-change inventory matches inspected images: error copy, wrapping, toast height/shadows/copy control, preparing duration and slow-request warning vary; the successful dark-phone baseline also changes branch/worktree controls and model label. Both sides retain empty drafts. No frontend source file changed in this patch.

Runtime detail reads succeed before and after refused local/worktree mutations. Independently compared all four C1 HEAD before/after JSON pairs, runtime before/after, and this audit's rebuild before/after: exact equality for recorded local config, refs, HEAD, worktree list, status and untracked listing. These observations are not an object-store comparison, a complete command/network trace, or an executed setup-script assertion. No setup script was configured. Source places the repository guard before local/worktree preparation mutation paths; WS status refresh is success-only. Native outbox does not proceed to startTurn after failed preparation. Existing client failure paths account for the draft/navigation/toast behavior observed.

Inspected the actual red log: baseline regression receives Success where Failure is expected. Inspected the actual complete green log: 109/109 tests passed, 76.83s. Verified production/test files are unchanged between the tested integration commit and final fixture-only HEAD. Did not mislabel the old skipped-test log as full-green proof and did not rerun tests or perform another code-review round.

**Rebuild procedure and capture limitation**

Own fresh T3 preview tab only, head web8017, existing server16057; exact pinned source and committed fixture. Paired normally, ran the seed through existing authenticated app APIs, opened the existing stopped app thread and linked PR panel, used light appearance, CSS1440×900, DPR2, visual scale1, loaded system font16px and the supplied fixed client Date 2026-10-22T04:30:00Z. Scroll was reset to top-left; no source/network mock or hidden UI was injected. An ordinary provider-update notification was dismissed before capture.

The first locator attempt failed before any checkout; it was excluded from the rebuilt action evidence. The corrected unique menu locator performed the one rebuilt worktree action successfully at 04:45:36–39 UTC. A text wait saw the refusal, but immediate recording stop retained a preparing frame as the video's last compositor frame. That pending frame is not passed off as the settled rebuild. A separate 04:46:34–35 UTC observation recording, without repeating the action, supplies the full-resolution settled refusal image used for the settled pixel comparison. Consequently this audit does not claim one uninterrupted rebuilt video showing terminal paint; the delivered C1 originals supply that interaction proof. The initial comparison uses the rebuilt first frame, visibly before Checkout, against the delivered third decoded initial frame. Neither comparison was registered, resized or masked.

[Initial comparison](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/95195d525b2473ae/rebuild-initial-pair.png) · [Settled comparison](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/70f1aba8a3e38360/rebuild-settled-pair.png) · [Settled ×4 raw difference](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2f2cb933163dd0de/rebuild-settled-diff4.png) · [Settled >16 mask](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/38e42e4186f0b1d0/rebuild-settled-threshold16.png) · [Numeric result](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2e2bc576e45568b0/rebuild-settled-stats.json) · [C3 frame observations](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dfe122eae6771277/retry-dismiss-check.jpg).

![Left: delivered HEAD refusal; right: rebuilt HEAD settled refusal, same layout and text with pointer/caret and encoding differences](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/70f1aba8a3e38360/rebuild-settled-pair.png)

**FRICTION**

- Fresh preview profile required normal pairing. No credentials or pairing material are included in this report or uploaded audit images.
- Exact-match locator strings must include the menu item's description; abbreviated names timed out. No alternate browser was used.
- Video frame-count/FPS metadata is unreliable for seeking these VFR clips; independent inspection used sequential decoding and actual frame PTS. Snapshot export downsizes desktop images, so full-resolution video frames were compared instead.
- Browser recording stop can precede a compositor frame carrying an already-observed terminal DOM state. This affected the auditor's rebuild and was explicitly corrected with a separate settled observation, not by relabelling the pending frame.
- Native UTC click logs are absent for several clips. Approximate observed elapsed milestones now accompany the continuous recordings; exact click latency remains unverified.
- The uploaded index supplied at handoff predates the last C3 wording correction. The lead owns re-uploading the corrected index/PR-body wording; the corrected local files are the final claim set assessed here.

**Fixture handback**

No further app actions are pending. `/tmp/issue-16455-audit-pass2/handoff.json` records the final fixture repository state; it equals the rebuild-before snapshot. Unrelated fixture remains origin-only, clean master at 5afa3f44d1115ab61425b2999daafa1b492b898d, one worktree, no untracked/setup marker files. No provider was run. Only an empty browser-local draft remains from the refused action. The new independent Python fixture stays separately under `/tmp/issue-16455-pass2-seed-check`. Both servers remain running and owned by the lead; this auditor started/stopped neither. No live userdata database was opened for writing.

VERDICT: SUFFICIENT