# Issue 23245 proof audit pass 2 — final

Audited target: /home/ubuntu/wt/issue-23245, HEAD b065ad799c5bfda2027d7d68327bb6bcd1c49666 against merged base 3b95e7831b11b17627dff16843bc087bfca397ec. Checkout clean. Read-only repository audit; no edits, push, comments, PR or merge. Authorized isolated-state/browser reproduction below. Final evidence index and compact PR evidence inspected. This is capture sufficiency review, not implementation or merge approval.

## Initial inventory and audit findings (superseded counts resolved below)

- Delivered provenance is historical before a50cd989aee06ed0762e26027eadf35933eb52ac / after 33e77985c707af61bac392e8adce1e01987ecf14. Neither matches the requested final base/head. The two product files have no differences from historical before to final base or historical after to final head; whole-client equivalence does not follow. Historical-after→final-head changes 78 files (+2930/-348), including web CSS, shared state and server behavior. Captures must retain historical labels.
- Initial manifest snapshot: 120 artifacts, 29 videos, 42 paired screenshot stages. R1/R2/R3 have both sides for all four cells. R4 desktop has both sides; R4-390-light has only baseline and R4-390-dark has neither side. These are inventory facts, not completeness certification.
- All delivered source artifact SHA256 values match files. Independently recomputed all 42 comparison stages: threshold-zero absolute RGB difference multiplied by 8/clipped255; masks max absolute channel delta >16 and >32; pixel counts and all diff PNGs match exactly. No registration, cropping or masking. Comparisons are technically valid for the supplied images.
- All 29 videos have one video stream and zero audio tracks, durations 4.666267–51.8437 seconds. Spoken narration is absent; written narrative/index completeness still pending. No assertion that every video has been watched in full. Sampled one-second frame sheets for desktop light after R1–R4: real app transitions; no visible step captions. Frame sampling cannot certify continuity/no edits for every required substep.
- Desktop delivered PNG1280x800 vs video2880x1800. Narrow PNG/video780x1688. Manifest expressly states historical CSS innerWidth/innerHeight and DPR were not retained; dimensions alone cannot verify CSS viewport. New independent desktop reproduction below establishes the scaled-export mechanism now, but cannot retroactively supply measurements for historical cells.
- At first audit, PR-ready index absent. uploads.json had284 entries, with8 artifact/comparison filenames missing upload mapping (some may be deduplicated aliases); awaiting final index before concluding publication readiness.

## Independent reproduction

Applied test-t3-app skill and its sqlite-fixtures reference; used only T3 collaborative Browser panel. preview_status initially unavailable; preview_open successfully created my own tab. Confirmed localhost7888 process cwd points to target apps/web; dev-runner log confirms isolated /home/ubuntu/wt/issue-23245/.t3 and server15928. Lead stopped its server before writes. Verified port7888 no listener and settings favorites=[]; ran committed `node scripts/seed-issue-23245.mjs /home/ubuntu/wt/issue-23245/.t3 R4`. Restarted `vp run dev` in retained session57323, /tmp/23245-pass2-dev.log. No competing server or direct fixture write while server running. No live install database writes.

Fresh own tab, paired against isolated server; Settings→Providers→Codex displays authenticated/enabled Codex, eight actual reported models, Favorites with one GPT-6.1-Sol and one audit/missing-model-23245. Slug-only row exposes Remove … from favorites and no custom editor/delete badge. Filter N/A with eight reported models. Selected explicit light mode through Appearance UI; returned to Providers. Direct browser measurement innerWidth1440, innerHeight900, devicePixelRatio2, document.fonts.status loaded. PNG export1280x800: raster scaling factors8/9 relative CSS,4/9 relative physical DPR dimensions. No repository edits or arbitrary clock overrides.

Initial reproduction had different scroll and update notification; raw delta271024 pixels(26.4671875%) at threshold0,44550(4.3505859375%) >16,35962(3.5119140625%) >32. Dismissed notification and scrolled the actual settings container to bottom(scrollTop=max276), matching delivered framing, without pixel registration/cropping/masking. Aligned reproduction `/tmp/23245-pass2-R4-1440-light-initial-aligned.png` vs delivered `/tmp/23245-evidence/R4-1440-light-after-initial.png`:199/1024000 pixels(0.01943359375%) threshold0;0 >16;0 >32. Maximum channel delta1; bbox[1277,82,1280,164] at right scrollbar. This independently reproduces one endpoint, not the full R4 recording or the other15 cells.

Retained `/tmp/23245-pass2-reproduction-aligned-delta.json`, `/tmp/23245-pass2-reproduction-aligned-diff0.png`, `/tmp/23245-pass2-video-audit.json`, logs, raw and aligned screenshots. Browser tab remains on Providers with seeded two favorites. Owned dev session57323 stopped with Ctrl-C, exit130. Confirmed no listeners on7888 or15928 after teardown. No other servers stopped. Seeded isolated state and Browser tab retained; reloading requires restarting the server.

## Final delivered package and resolved findings

Final full index: [historical evidence index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/edf6588d9c738958/index.md), local /tmp/23245-evidence/index.md. Compact /tmp/23245-pr-evidence.md is34437bytes, with16cell sections,64embedded images(four columns per cell),32recording links. Both disclose historical SHAs and exact-current-head gaps. No publication performed by this auditor.

Final manifest has132capture artifacts,32recordings,48paired screenshot stages. All16cells now have before/after recordings and screenshots; narrow R4 missing-side finding is RESOLVED. All132 source artifact hashes independently verified. All48 comparison stages independently recomputed; every pixel count and every diff0/threshold16/threshold32 raster matches. Every source/comparison filename has an upload mapping in final uploads.json(328entries). Link mappings inspected locally; remote downloaded bytes not independently rehashed. Written recording-step synopses and ordered historical logs now supplied. All32 clips are silent; no spoken narration. Original step logs include actual wall-clock entries; new recovery logs repeat artifact file times, expressly not within-video timecodes. If user requires spoken narration, all32 clips remain noncompliant; if written narration alongside continuous recordings suffices, written narration is supplied. This audit does not silently upgrade silent clips to narrated audio.

## Exact row disposition

- R1-1440-light, R1-1440-dark, R2-1440-light, R2-1440-dark, R3-1440-light, R3-1440-dark, R4-1440-light, R4-1440-dark: historical before/after media and valid diffs delivered. Original CSS1440x900 and DPR measurements absent; all8desktop cells remain UNVERIFIED for exact original viewport. Independent final-head R4-light endpoint measured/reproduced separately, not a replacement full before/after sequence.
- R1-390-light, R1-390-dark, R2-390-light, R2-390-dark, R3-390-light, R3-390-dark: historical paired media and valid diffs delivered. Original innerWidth/innerHeight/DPR measurements absent; all6original narrow cells remain UNVERIFIED for exact original CSS390x844/DPR. Raster780x1688 is consistent with DPR2 but does not prove it.
- R4-390-light and R4-390-dark: historical paired screenshots and recordings now delivered. Recovery logs explicitly report390x844CSS/DPR2 for all4sides, persisted light/dark, eight actual models, filterN/A, real stars, reload. These fix narrow inventory/provenance gaps for these2historical cells; they remain historical, not exact-final-SHA captures.
- ALL16cells: no complete before/after recording pairs at final merged base3b95e7831b and final headb065ad799c. This remains the decisive unresolved provenance requirement. Equality of the two changed component files is supporting source evidence only. New final-head R4desktop-light screenshot is one independently verified endpoint, not16current-SHA pairs.

## What else changed / interpretation

Index and compact body identify incidental differences: R1popup height/upward placement, hover/provider tooltips and model/reasoning labels; R2sidebar visibility, selection, tooltips and panel placement; R3added Favorites row increases panel height and shifts downstream content, with differing scroll; R4added missing row shifts All/downstream cards and scroll; both new narrow baseline initials show a transient update toast and truncated slugs. Diffs retain the entire frame; no incidental pixels were masked. These notes explain categories, not a pixel-by-pixel causal attribution. Every visible difference remains accessible through full-frame threshold0 and thresholds16/32. No assertion that every incidental difference is implementation-caused or that this audit independently watched every required action in all32clips. Sampled R1–R4desktop-light after frame sheets show real app transitions; all other action-completeness claims remain supplied evidence, not independently certified full playback. Baseline cannot remove a synthetic favorite it never renders; index correctly labels that limitation. R4filterN/A is justified by8reported models. No need for additional broad checks.

The compact body says each mutation capture reloads. R1/R2 picker sequences use close/reopen and navigation; avoid interpreting that sentence as independent proof of full-page reload for every picker substep. Source logs and per-row sequence remain authoritative. Final full index appropriately retains failed attempts/retries and does not count them as passes.

FRICTION
- All16cells retain historical SHA provenance rather than requested final base/head. Partial proof can be published only with explicit unverified labels; it is not complete capture certification.
- Exact historical CSSviewport/DPR remains unverified for8desktop plus6original narrow cells; new R4narrow measurements and the independent final-head desktop endpoint do not retroactively establish the rest.
- All32videos are silent. Written narration supplied, no spoken narration or true per-action timecodes for recovery. Full independent continuity/action playback certification was not performed for all32clips.
- What-else-changed notes now cover named incidental categories and retain all raster changes, but do not establish exhaustive causal equivalence across historical/final clients.
- Repository stayed clean; no edits/push/public comments/PR/merge/broad checks. Only authorized stopped-server isolated seeder and Browser reproduction. Owned server stopped and ports7888/15928 confirmed closed. Logs/screenshots/report retained.

VERDICT: INSUFFICIENT — all R1–R4 ×1440/390 ×light/dark(16cells) lack complete exact-final-base/head before/after recording proof; historical viewport/DPR remains unverified for8desktop and6original narrow cells. No remaining missing historical artifact pairs. Silent-recording narration and all-clip continuity certification remain explicitly limited as above.