# Independent proof audit pass 2, correction round 3 — [#19909](https://git.cafe/pingdotgg/t3code/pulls/19909)

## Findings first

No remaining proof blocker or implementation defect identified. The prior INSUFFICIENT finding was valid for superseded captures: a persistent thread details panel, not a transient Git popup, remained behind the native dialog and after dismissal. Every newly delivered native pair and both native recordings now exclude that panel, including after Close and Escape. No code change was needed to correct the fixture.

The original correction-round report had a real completeness defect: its main still table and geometry-base/head.json retained only the final W800-light selection. This did not mean seven pairs were missing: all assets were delivered. During this audit the lead recovered the other geometry from original capture result files and published the [complete eight-row report](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8e67a2da0b9607dc/capture-report-complete.md), [complete baseline geometry](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e77f194c4c7808bd/geometry-complete-base.json), and [complete head geometry](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8846c0847bfb1fd2/geometry-complete-head.json). I checked all eight rows in each reconstructed file against the original native/browser result files; every geometry object matches. Treat those supplements as final claims. The original archives still contain the truncated geometry files; these supplements are required alongside them.

Linux at actual T=40 already clears native controls on baseline. Evidence establishes improved clearance, not a reproduced collision. Windows remains unverified and accepted; native 390px remains unsupported at minimum 840px and accepted. Neither is a hard stop. The synthetic 52px geometry regression remains authorized and is not represented as Windows proof.

## Identity and independent method

Read applicable worktree AGENTS.md. Read-only checkout /home/ubuntu/wt/pr-19909 remained clean on takeover/pr-19909 at head [9a5d15e3c6](https://git.cafe/pingdotgg/t3code/commit/9a5d15e3c6c0c217ddd1fb3265a6bfe746e381a9), actual mergebase [3cc31ed27c](https://git.cafe/pingdotgg/t3code/commit/3cc31ed27c8956b36cee35a50509fb7c6869592b). These match environment.json and the corrected claims. Feature diff has exactly three files: height expression, real browser geometry regression, comment-only ZoomableImage update. Original brief/history preserved: original fork [d0c7c3a8d5](https://git.cafe/pingdotgg/t3code/commit/d0c7c3a8d58e686fea7ffaba15dcd72c5661ca0f) and earlier pass-2 evidence are historical, not current takeover proof. No fork access or repository modification.

Downloaded the delivered public proofbundle and proof-source archives independently into /tmp/ev-19909/audit-r3. All 145 manifest entries in the proofbundle match SHA-256. All manifest-covered source archive members match. The corrected report was also fetched from its published URL. Fresh full-matrix checks and fresh extraction of representative frames were performed against the downloaded MP4s, not inherited audit sheets or capturer assertions.

Personally inspected all 16 primary stills and all 16 differences in labeled full-image comparison sheets, all 32 pre-open/dismissed images, every one of the eight narrated steps in each of four recordings via newly extracted frames, native control crops at original resolution, and full-resolution native Close/Escape frames. These show open, 125%, reset 100%, bottom marker, HTML Close, reopen, Escape, and completed state. Native recordings retain the window and controls after HTML Close. No details panel or transient menu persists after either dismissal. Ordinary focus/hover tooltips on the reopen control are visible in some supporting frames; these are not the previously rejected panel. Some pre-open tall supporting frames show the fixture-title tooltip; primary stills are settled and clean.

Independently recomputed every Diff0 as clipped abs(RGB delta)×8, and every DiffT as RGB [255,60,120] exactly where any channel differs by >16, black elsewhere. All 16 delivered diffs match. All 16 stability companions match primary images exactly. All four browser pairs are exact-zero RGB differences. Native changes are shorter pane, displaced Close/footer/content, shadow/ring/gradient boundaries, and newly exposed backdrop; window controls and horizontal geometry stay fixed. No unrelated changing region found. This agrees with the corrected report's what-else-changed account.

## Row statuses

Counts are independently recomputed changed pixels: threshold zero / any-channel >16. Each static row includes baseline, head, Diff0, DiffT, stability companions, pre-open and dismissed states. Native outer/content sizes match; scale=1. Native T=40; browser has no desktop bridge/titlebar. Both themes were checked individually.

| Row | Viewport | Status | Geometry and pixel findings |
| --- | --- | --- | --- |
| L800 light | 1440×800 | SUFFICIENT | Close.top 42→62; Zoom-in.bottom 758→738; 371683 / 75705; clean native backdrop/dismissal |
| L800 dark | 1440×800 | SUFFICIENT | Same geometry; 302174 / 13103; clean; fresh independent rebuild exact-zero delta |
| L1200 light | 1440×1200 | SUFFICIENT | Close.top 46→63.203125; Zoom-in.bottom 1154→1136.796875; 743183 / 67983; proportional branch; clean |
| L1200 dark | 1440×1200 | SUFFICIENT | Same geometry; 556842 / 12954; clean |
| W800 light | 1440×800 | SUFFICIENT | Close.top=22, Zoom-in.bottom=738 unchanged; 0 / 0 |
| W800 dark | 1440×800 | SUFFICIENT | Same unchanged geometry; 0 / 0 |
| N800 light | 390×800 | SUFFICIENT | Close.top=20, Zoom-in.bottom=740 unchanged; 0 / 0; long title truncates; controls fit |
| N800 dark | 390×800 | SUFFICIENT | Same unchanged geometry; 0 / 0; controls fit |
| I-L baseline | 1440×800 dark native | SUFFICIENT | Every narrated step visible; 100→125→100%; bottom marker; clean Close/reopen/Escape/completion |
| I-L head | 1440×800 dark native | SUFFICIENT | Same sequence; native controls remain distinct; clean dismissal states |
| I-N baseline | 390×800 light browser | SUFFICIENT | Every narrated step visible; clean dismissal states; footer accessible |
| I-N head | 390×800 light browser | SUFFICIENT | Same successful sequence and clean states |
| L390 both themes | requested native width | ACCEPTED UNAVAILABLE | Product minimum [840,620]; no synthetic renderer width claimed |
| WIN | actual Windows | ACCEPTED UNVERIFIED | No Windows run; explicitly accepted platform gap |

Native 800 and 1200 exercise the subtraction and proportional branches. Plain wide/narrow browser controls exercise T=0. Browser recordings preserve the disclosed 500×900 outer window, 390×800 content, and unused white strip without scaling. Captions occupy appended space below application pixels. Linux Electron 44.9.3/Chromium 152.0.7977.167, browser Chrome 148.0.7778.96, fixed clock, reduced motion and source identities agree with delivered provenance.

The production inline HTML→fullscreen route is covered. Desktop uses the same component. Settings/palette/keybinding do not provide another HTML dialog caller; native mobile uses separate UI. Providers supply the same HTML reference, so no provider matrix is required for this CSS-only change. Remote/tunnel/macOS/mobile were not executed; no claim otherwise. Fixture seeding uses the disclosed isolated project/stopped thread, real published HTML attachment, and completed projection row. It proves visible UI behavior, not provider execution or event sourcing. No live ~/.t3 state was accessed.

## Independent native rebuild

I personally executed current REBUILD.md/rebuild-one.sh with the retained isolated /tmp/ev-19909/desktop-state at the exact latest head. Started owned Vite on 6361 and the documented headless-desktop driver, actual Electron/backend on 14402, Xvfb :193. No browser pairing, CSS titlebar simulation, code/config/auth edit, direct component harness, or live state was used.

The newly produced L800-dark-auditor-head.png compared against the freshly downloaded delivered L800-dark-head.png has **0 changed pixels, maximum channel delta 0, mean channel delta 0**. This is my fresh run, not the capture owner's earlier zero-delta claim. Fresh stability frames match. Runtime outer/content=1440×800, scale=1, real T=40, Close.top=62, Zoom-in.bottom=738. Personally inspected the new full image and full dismissed image: clean native backdrop, no persistent details panel, and no transient menu. Results and extracted audit material remain under /tmp/ev-19909/audit-r3.

![Independent rebuild: head native HTML Close clears window controls](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5a579a7509fe8f85/L800-dark-auditor-head.png)

![Independent rebuild: clean state after HTML Close](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0f725b88e15f14f2/L800-dark-auditor-head-dismissed.png)

Stopped the owned driver by freshly captured PID 1314236; its handler closed Electron 1314271/backend and Xvfb 1314260. Stopped Vite through retained terminal session. Verified all three captured PID paths absent, driver session exited, and ports 6361/14402/23399 clear. No browser/recorder was launched by this audit. Repository still clean at the requested head. No commits, pushes, PR creation, comments, watch, merge or fork actions. Only the required app-local PR registration was performed.

## Friction

FRICTION: evidence | original report and geometry JSON retained only the last selection despite complete assets | accumulate row results or construct final tables from immutable per-capture outputs; corrected supplements now provide all eight rows.

FRICTION: none during independent native rebuild, pixel verification, recording inspection, teardown or report upload.

VERDICT: SUFFICIENT L800-light, L800-dark, L1200-light, L1200-dark, W800-light, W800-dark, N800-light, N800-dark, I-L-base, I-L-head, I-N-base, I-N-head; L390 accepted unavailable; WIN accepted unverified.