1. **Recording defects invalidate W09, W10 and M01.** `W09-N-dark-after.webm`, source frame76 at5.315467s, is black except the pointer: the entire application disappears. `W10-N-light-after.webm` ends with blank frames at2.264933s and2.271700s. `M01-A-light-after.mp4` shows a blank child timeline with Loading messages / Syncing messages through roughly12.7–14.6s; frame41 at13.275711s clearly shows Syncing messages. Frame84 at19.489856s contains overlapping parent/child headers, messages and composer controls during return navigation. `M01-A-dark-after.mp4` also has a blank child timeline around8.04s. These are recording defects under the requested any-frame rule, regardless of whether they are ordinary navigation transitions. They are proof findings, not a diagnosis that this PR caused loading or navigation bugs.
2. **Unrelated pixels invalidate delivered A04/A02/A05/M03 pairs and M01 light pairs.** In `A04-contact-sheet.png`, the web historical, both-disabled and pooled cells change sidebar ages from3h to4h. Both diff panels light up down the sidebar, outside the changed avatar code. This is absent from the report's what-else-changed list. Android A02's three states, A04's four states, and A05's on/off states change the far-right disclosure chevron around physical x1020,y1810 (y1685 in disabled A04 scenes). M03 also changes chevron/copy/fork raster pixels. M01 light has six navigation-handle pixels changed near the bottom of the screen, disclosed in the capture report. Disclosure does not make these avatar-proof pairs valid under the supplied standard. A05's chevron noise was omitted from the report. Rejected `M03-A-dark-expanded-*` additionally differs in the failed child's model label and floating scroll control, as the capturer reported; it remains rejected.
3. **W06/W09/W10 do not show the required delayed tooltip.** Their supplied recordings show focus/hover and navigation, but no popup. The harness root has zIndex9999 and occludes the portal. Files include `W06-W-light-before/after.webm`, `W06-N-dark-before/after.webm`, all six W09 clips and all six W10 clips. The focus ring itself is visible before and after in W06; the ring-survives claim is supported. The script/Clip-timelines description “hover child; delayed tooltip” is not evidence of a visible tooltip. The report correctly acknowledges this limitation. W06's comparison restoration step in Clip-timelines is also not visible at the end of the wide clips, which finish in the actual application.
4. **W08 does not demonstrate its defining transition or the production fallback renderer.** `W08-contact-sheet.png` shows only old Finished/9:00AM notifications. The rejected `W08-W-light-resumed-*` and `W08-N-dark-resumed-*` still retain Finished. Server resume read-backs do not establish live status/elapsed rendering. No W08 recording is linked in the final index. The harness supplies the literal paragraph “Ordinary work entry fallback”; this proves the fallback callback branch, not an ordinary WorkEntry rendering. The capturer discloses this boundary, so it must not be upgraded to full fallback proof.
5. **Required cells and branches remain missing.** No iOS capture or real remote-environment capture exists. W03's new wide-dark pair genuinely has five children, +2 overflow, failure, a no-space path detail and prose detail; wide-light/narrow-light/narrow-dark retain the round1 title-only limitation. Timing, identity/morphology/hydration, pending/failed/fallback handoffs and the iOS inspector gaps are detailed row by row below. H01 proves the real one-source Android handoff and endpoint alerts, but neither plural-source wrapping nor iOS.
6. **The PR body still describes older evidence.** Its evidence columns name before [97e35bcf23](https://github.com/pingdotgg/t3code/commit/97e35bcf23) / after [aebac9ae29](https://github.com/pingdotgg/t3code/commit/aebac9ae29), with a ×32 description. Those old attachments cannot establish this requested head/base. All round2 sheet columns instead match [32ecd71726](https://github.com/pingdotgg/t3code/commit/32ecd71726fee16f335484c1f95cbce126d461f6) / [fd6f84adeb](https://github.com/pingdotgg/t3code/commit/fd6f84adebd5592fff1df47f903ad260c389f2bd) and state ×16/threshold0 and channel>16/255. The broad claim that changed pixels are confined to avatars is contradicted by the newly delivered pairs above. Unbadged overlap reversal itself matches the stated stack-order intent and the code; it is not an extra-scope finding.

Reviewed [#25031](https://git.cafe/pingdotgg/t3code/pull/25031), branch fix/issue-24896, at exact head fd6f84adebd5592fff1df47f903ad260c389f2bd against explicit merge-base32ecd71726fee16f335484c1f95cbce126d461f6. Read-only worktree `/home/ubuntu/wt/pr-25031`; no edits, commits, pushes, reviews or comments. Round1 results are inherited from the supplied [prior audit](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/da248a9bd76fe1ab/audit-report.md), rather than being claimed as newly inspected originals in this pass.

**Combined row table.** W=1440×900 CSS web; N=390×844 CSS web; L/D=light/dark; A/I=Android/iOS. Diff column: R1=previous audit recomputed originals, inherited here; S=round2 contact-sheet panels visually inspected, full-resolution web originals unavailable so their diffs cannot be recomputed; F=round2 full-resolution archived stability PNGs used to recompute both diffs. A row may be insufficient despite useful partial observations. Missing original web PNGs alone are not treated as a rejection, per the user's clarification.

| Row | Combined status | Coverage and exact remaining limitation | Diff check |
|---|---|---|---|
| W01 | SUFFICIENT | R1 W/N × L/D, collapsed/expanded, pointer and keyboard disclosure. | R1 |
| W02 | SUFFICIENT | R2 replaces all spinner pairs: W L/D, N D, collapsed exactly-three clean stacks; own head rebuild matches retained panel. | S; rebuild below |
| W03 | INSUFFICIENT | R1 + R2 W D disclosure/five/failure/details; W L and N L/D actual long path/prose-detail retakes absent. | R1 + S |
| W04 | INSUFFICIENT | No active mixed group, two clock increments/settlement, missing timestamps or inactive hover proof. | None |
| W05 | INSUFFICIENT | R1 W L/D static palette harness only; N D long text and status/active transitions absent. | R1 |
| W06 | INSUFFICIENT | R2 W L/N D standalone actionable/nonactionable, visible focus and actual child open/back; delayed popup absent. Expanded path partly supplied by R1 W01. | S |
| W07 | SUFFICIENT | R2 W L/D all five fixed outcomes, fixed event time, correct dot/no-dot and labels; explicitly HARNESS, no live delivery claim. | S |
| W08 | INSUFFICIENT | Old terminal W L/N D only; resumed rendering remains Finished; no successful transition recording; literal fallback paragraph. | S |
| W09 | INSUFFICIENT | W L/D/N D labelled valid-lineage phased HARNESS; visible queued/running/done/failed, advancing clock, settlement, script unfold/fold and child open/back. Tooltip absent; N D after clip has blank frame. | S |
| W10 | INSUFFICIENT | W L/D/N L unphased/actionable/no-child/short-script/no-group HARNESS; tooltip absent; N L after recording ends blank. | S |
| M01 | INSUFFICIENT | R2 A L/D collapsed/expanded and real navigation; I L/D absent. Android clips contain blank/loading/ghosted frames; L diffs have navigation-handle noise. | F, four pairs |
| M02 | INSUFFICIENT | R1 A L/D three positions readable; I L/D absent. | R1 |
| M03 | INSUFFICIENT | R2 A L collapsed/expanded and D collapsed; D expanded rejected; I L/D absent; accepted native diffs also include control raster changes. Third badge remains visible beside +2 in supplied headers. | F, three indexed + one rejected pair |
| M04 | INSUFFICIENT | No A L/I D live timing, settlement or missing-time controls. | None |
| M05 | INSUFFICIENT control | Singleton/empty-work-log guard observations absent; no changed avatar output. | None |
| M06 | INSUFFICIENT | No supported wide-iOS inspector or phone comparison. | None |
| A01 | INSUFFICIENT | R1 grouped W L/D, N D, A D plus R2 A L; no standalone web specimens or I L/D. New A L no-badge/full-opacity control clean; overlap-only change. | R1 + F, one pair |
| A02 | INSUFFICIENT | R1 W L/D/A D plus R2 A L three real saves; contrast and badge removal visible; I L/D absent; R2 native chevron noise. | R1 + F, three pairs |
| A03 | INSUFFICIENT | Default/custom/Unicode/punctuation label specimens absent. | None |
| A04 | INSUFFICIENT | R2 W L/A D all four settings states and changes visible; I D absent; web sidebar-age drift and native chevron drift invalidate affected pairs. | S web; F native, four pairs |
| A05 | INSUFFICIENT | R1 W L/D real project saves/reversal + R2 A L; I L absent; R2 native on/off chevron changes omitted from what-else list. | R1 + F, two pairs |
| A06 | INSUFFICIENT | ACP same/different/trimmed registry identity and local-instance controls absent. | None |
| A07 | INSUFFICIENT | Required Claude/Codex/SVG/image/fallback morphology across native platforms absent; ACP alone insufficient. | None |
| A08 | INSUFFICIENT | Hydration/null/error/invalid-accent/rename/status-push sequence absent. | None |
| R01 | INSUFFICIENT | No two-environment identity scope or actual remote child open/back. | None |
| R02 | INSUFFICIENT | No absent remote config / fallback / reconnect proof. | None |
| R03 | INSUFFICIENT | No one-account-per-environment pooling/accent control. | None |
| H01 | INSUFFICIENT | R2 real Android L/D C1→VY committed handoff and both alerts/dismissals; all six paired endpoints identical. I L/D and plural-source wrapping absent. | F, six zero-delta pairs |
| H02 | INSUFFICIENT | Pending footer→committed divider and duplicate-footer guard absent. | None |
| H03 | INSUFFICIENT | Failed/legacy/zero/plural-source/raw-model endpoint branches absent. | None |
| H04 | INSUFFICIENT | Missing model/provider/config fallback and alert naming branches absent. | None |
| D01 | SUFFICIENT | R2 real Electron L1440×900, clean collapsed/expanded stills and continuous expand/open/back/collapse before/after. | F, two pairs |
| Z01 | INSUFFICIENT control | Code-traced; R2 W10 no-group guard visible, other surrounding guard observations absent. No dedicated positive capture required. | S for shared W10; remainder none |

**Inspection and recomputation coverage.** Opened all15 indexed contact sheets and every before/after/Diff0/DiffT cell (53 paired scenes,106 diff panels), all104 archived full-resolution stability PNGs through26 four-image review pages, and the four explicitly rejected scene sheets. Decoded every source frame of all40 linked recordings:1965 frames,88 review pages, all visually inspected. Also inspected all133 decoded frames/five pages of the two archived rejected W06 navigation clips; their reload/splash/blank states remain rejected. Recording narration is written steps/timelines, not audio. W03 disclosure/collapse, A02/A04/A05 settings changes, H01 endpoint taps/dismissals and D01 interactions match the visible actions. Missing/contradicted steps are identified above. No linked W08 clip was available, so diagnostic W08 timeline text was not credited as a viewed recording.

All52 native/Electron within-revision stability pairs are pixel-identical. All indexed round2 columns have the exact requested SHAs. Viewports/themes agree with visible captures and delivered rendering metadata. Web Chromium156/Linux/DPR2/font1 uses CSS1440×900 or390×844, exporting1280×800 wide or780×1688 narrow rasters. Wide contact-sheet panels are540×338; narrow panels540×1169. Pixel9/API36 physical1080×2424 corresponds to logical411.43×923.43/DPR2.625/font1, permitted by the matrix's approximate native viewport. Native status bar is hidden and system animations disabled. Electron44.10.0/Chromium152/Xvfb is1440×900/DPR1/font1/reduced-motion reduce. The web preview has reduced-motion=false and relies on stable endpoint frames.

Recomputed Diff0 as saturated abs(RGB_before−RGB_after)×16, threshold0, and DiffT as each RGB channel>16/255. For all25 indexed native/Electron pairs, recomputed original before/after/diff panels resized by the supplied sheet algorithm match the delivered sheet panels exactly (zero mismatching pixels in each of four columns). Recomputed the rejected M03 dark-expanded pair separately. `native-recomputed.json` records full-resolution counts/bounding boxes; `panel-verification.json` records panel equality. This confirms faithful diff construction, not clean proof: unrelated highlighted pixels still invalidate the scenes listed above. H01's six pairs are exactly0 at both thresholds; D01 changes533/1358 pixels at threshold0 and196/667 at>16, confined to avatars. Full-resolution web originals and unredacted Upload-manifest.md are unavailable, so S rows' delivered differences were visually inspected but cannot be independently recomputed. The user explicitly directed continuing with this limitation.

**Independent W02 head rebuild.** Ran the uploaded `rpc-driver.mjs`, with only scratch-root paths changed, from the exact head in the supplied worktree: fresh isolated home, T3CODE_DEV_SEED=0, uploaded frozen-clock.mjs, `vp run dev --home-dir /tmp/lead-25031-r1/proof-audit-3/home-after --port19323`; actual CLI argument was `--port 19323`. Ran seed, settle-parents, state three. Real RPC-created child state and actual settings, no injected projection. Opened proof-group-3, expanded its work log, left the exactly-three group collapsed, moved pointer away and blurred focus.

Rendering environment independently read from the page: Linuxx86_64, Chrome156.0.0.0,390×844CSS,DPR2,zoom1,16px root,dark,reduced-motion=false, frozen Date2026-10-30T11:40:00.000Z. Full saved capture780×1688. Two final rebuilt captures are identical over all1,316,640 pixels. Resizing the stable final capture with the supplied Pillow BICUBIC algorithm to the retained after-panel540×1169 gives **0/631,260 differing pixels, maximum channel delta0,>16 count0, no bounding box**. Both comparison diffs are black. No non-antialiasing delta was observed at the preserved resolution. This cannot establish equality against the unavailable780×1688 delivered original. The first rebuild capture caught a disclosure-chevron transition and was excluded; its raw capture is retained in the bundle to make that choice reviewable.

![W02 delivered head panel versus independent rebuild, followed by both zero-delta diff panels](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/10f0a3d631f18c1c/W02-rebuild-comparison.png)

![A04 before and after sidebar ages, with unrelated sidebar changes in both delivered diff panels](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/600bbdb2104a43fe/A04-sidebar-confirm.png)

![M01 after-recording frame41 at13.275711s: blank child timeline and Syncing messages](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4b97aabad7b5396a/M01-A-light-after-frame41.png)

![W09 after-recording frame76 at5.315467s: entire app blank except pointer](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/546aa579fbea2ef5/W09-N-dark-after-frame76.png)

Owned preview tab closed. All synthetic threads stopped through RPC. Owned dev runner475522 and its remaining server/watch/Vite PIDs475580,475593,475579 stopped explicitly after verifying cwd/command and ownership; backend19323/web8724 no longer listen and those PIDs no longer exist. No emulator or lease started. Worktree clean; lead-owned checkout and parent scratch preserved. Audit scratch is removed only after uploading its sanitized artifacts. Credentials, startup pairing URLs and the fresh home/database are excluded from public artifacts; operational logs/read-backs are sanitized.

FRICTION:
- runbook | Full-resolution web PNGs and the unredacted upload manifest were absent; the capturer scratch was already gone | bundle original PNGs and a recoverable manifest with evidence.

VERDICT: INSUFFICIENT W03–W06, W08–W10, M01–M06, A01–A08, R01–R03, H01–H04, Z01