# Proof audit pass 2 — issue 24934

VERDICT INSUFFICIENT for the complete proof matrix. The delivered index honestly identifies the remaining rows. This is not a finding of an implementation defect, and does not revoke the user's permission to prepare a PR with explicit unverified rows. Return those gates to the dispatcher; this pass did not repair or expand capture coverage.

Audited [3027afd797](https://git.cafe/pingdotgg/t3code/commit/3027afd79720618974d965e7076e77e8aac1d38b) against [e3592efe7f](https://git.cafe/pingdotgg/t3code/commit/e3592efe7f5dbd3a869e46e608d8f6815e1bb7f6), in `/home/ubuntu/wt/issue-24934`. HEAD matches and source remained clean. The seven-file diff matches the stated scope: current/legacy single menus, header shared builder, both bulk builders and filtered dispatch loops. No native mobile feature or transport change. Predicate classification agrees with source, including queued/null-run-id and waiting. Tests were read, not run.

| Row | Audit result | Remaining gate |
|---|---|---|
| W1, 1440 light | PARTIAL: five caller paths and menu presentation supported | Accepted recording omits positive actions. Earlier sanitized F1-only trace is supplementary; fresh legacy read→unread transition was not independently established. |
| W2, 1440 dark | COVERED for projection-fixture menu presentation | Does not establish persisted mutation or native desktop. |
| W3, 390 light | QUALIFIED: menus, omission and mixed count supported | Chat/composer layout differs between base/head; full-page comparison is not omission-only. Synthetic modifier/context events do not prove touch long-press. |
| W4, 390 dark | QUALIFIED, same limits as W3 | Same incidental layout and input-method limits. |
| I1 | MISSING | Keyboard open/navigation/Escape sequence and real active→finished transition not delivered. Fixture status replacement is not a real run transition. |
| D1 | UNVERIFIED | Headless desktop infrastructure unavailable; Electron native menus not exercised. |

Inspected all 96 accepted endpoint images and 96 diff images through the delivered contact sheets, with individual representative/rebuild inspection. Inspected sampled frames from all eight recordings, not continuous playback or every frame. The supplied W1 head video sheet covers only the early portion of its 93.34-second recording; independently decoded 3-second samples through approximately 90 seconds and inspected its remaining sheets. Samples support the described menu sequence; they cannot establish precise keyboard actions or absence of every transient defect. Recordings contain projection changes, not provider execution.

All 96 PNG dimensions match the manifest. Wide CSS viewport is 1440×900 at DPR 2, saved PNG 1280×800; narrow is 390×844, PNG 780×1688. ffprobe confirms all wide recordings 2880×1800 and narrow 780×1688. Light/dark appearances match labels. Recalculated all 48 pair pixel counts, all 48 amplified RGB diffs, and all 48 threshold-16 overlays: zero mismatches. Archive entries match corresponding local bytes; no rejected directory is included. Base/head attribution is supported by the index and checkout reflog, not an immutable SHA embedded in each recording. No independent base rebuild was performed.

Visible extra changes: narrow chat bubbles/composer wrapping and vertical placement differ substantially, including header comparisons after the drawer closes. Wide finished controls also have tiny nonzero raw differences outside menus (hover/border/icon raster differences; dark finished controls have zero pixels above threshold 16). These do not contradict omission behavior, but avoid a literal “only changed pixels are the omitted item” claim. Animation disabling is injected CSS; native reduced-motion emulation remains unverified as disclosed.

Independent rebuild used the published projection fixture and the affected current-active sequence from the capture driver at HEAD, existing isolated server 7374, own T3 preview tab, explicit tabId on every subsequent call, 1440×900/light/DPR 2. Only sandbox browser projection/preferences were changed; pairing recovery used `node apps/server/src/bin.ts pair` privately. No provider, database write, source edit, new server, server stop, forge mutation, push, merge or watch.

Compared decoded RGB against delivered `head-1440-light-current-active.png`: **387/1,024,000 pixels changed (0.037793%)**, 60 above channel threshold 16, mean absolute channel delta 0.004616/255. Interior menu ROI `[106,491,289,796]` is **pixel-identical (0 changed)** and omits Mark unread without a gap. Full-frame differences are primarily F1 elapsed label (35m versus 9m), header hover chevron and fine sidebar/border pixels; the published Date override does not force a perfectly identical elapsed label in a fresh session.

[Rebuilt capture](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a9c88cf022001603/rebuild-head-1440-light-current-active.png) · [RGB difference ×8](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8f41091951d0495d/rebuild-diff-x8.png) · [Pixel metrics](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ca4f093aaf3820a7/rebuild-delta.json)

FRICTION: fresh pairing required manual token submission after link navigation; no credential published. preview_evaluate rejects leading/trailing whitespace, so the published fixture required trimming. Updating preferences can refresh shell state, requiring fixture reapplication in the sandbox. Delivered W1 head contact sheet truncates the long recording; full-duration sampling was generated outside the worktree. These were audit/reproduction friction, not product regression findings.

Skills used: test-t3-app for authorized browser verification and ghdrop for artifact/report publication. No other people were messaged or comments posted.