# Proof audit pass 2

## Findings first

No blocking proof gap remains for the scoped A–D representative Codex web happy path on exact final73664cd38654950538122c56f40c04429ad8a0c8. Accepted original baselinea46f80ee28 remains valid; base d53873b6a4 is not being misrepresented as that original before capture. Worktree is clean. Independently recomputed adapter/service/orchestrator hashes match head736-content.sha256. This is an evidence audit, not another code review. Reported97/97 focused regression results were not rerun here.

Nonblocking qualifications:
- Independent same-state rebuild was attempted in a fresh own default-profile1280x800 tab using test-t3-app. Both origin and supplied exact thread route redirected to /pair. No credentials minted, no server/tab mutation, no turn seeded. The server subsequently stopped. Therefore no independent rebuilt screenshot or cross-client pixel delta exists. To complete that additional check: provide an authenticated auditor-owned tab while the same isolated server/state is available, load the exact thread route, capture1280 light, and compare with after-head736.png. This optional-if-feasible check does not negate recorded D reload/restoration.
- Final stability pair differs16 pixels at bbox(321,656)-(322,672), exactly the composer caret. No zero-delta claim. No anti-aliasing/timestamp delta was measured for an independent rebuild; those must not be invented. Earlier1440 light pair differs30 caret pixels; other inspected supporting pairs compare identically.
- Final index B wall-clock interval spans37.608s, while its actual video duration is30.434s; AB is49.620s (A19.219s). Treat UTC ranges as operator wall-time narration, not a frame-exact duration mapping. Actual required tool/result behavior is visibly retained; this discrepancy is not a missing-action gap.
- Older B loading-only claim was inaccurate: its18.8s frame already showed output. Final package corrects it. C-request.png is an earlier thinking still; actual request is visible in the recording, not that filename.

## A–D results

| Row | Status | Visual and runtime corroboration |
|---|---|---|
| A | PASS exact final | A-head736 / AB0–19.2s: loaded READY, one new queued notification, QUEUED, settled controls. Run23 preserves exact Reply with the word QUEUED.; original native turn01a16e17-f72f-7fc2-b713-800a278400e1 starts13:27:06.820, completes12.760. No adoption turn/start in complete outgoing allowlist. Prior identical history distinguished by run IDs. |
| B | PASS exact final | B-head736 / AB after19.2s: command lifecycle, final DONE_TOOL_FINAL_23919, group/detail expansion, resolved TOOL_FINAL_23919 output held alongside final reply (AB45–48s and B-head736.png). Run25, native turn01a16e18-6c61-7fe1-a552-e08db9b71af7, one command exit0 at13:27:39.788 before final/completion43.358. No adoption turn/start. |
| C | PASS exact final | CD roughly14–18s pending native question/options, BETA selection, card clears/answered row;20–21s resumed final DONE_REQUEST_FINAL_HEAD_23919 BETA_FINAL_HEAD_23919. Run27/native turn01a16e19-f300-73f0-b832-86e6895c3d20. Native request13:29:21.950; outgoing answer26.071; resolved26.074; completion28.925. One persisted resolved request with selected answer; no second response/start. |
| D | PASS exact final | CD32–39s ordinary composer NEXT, one reply, settled;45–47s splash;49–53s restored answered request/final reply/NEXT. after-head736.png confirms settled restoration. Run28/native turn01a16e1a-4ff3-77e0-a31c-0451d09e46c1; one normal outgoing turn/start13:29:40.593, completion43.994. Provider thread idle; no pending request. |

Baseline before.png and before-runtime.txt agree: same native thread completed external QUEUED12:38:15.821 while baseline UI remained READY only. Different richer after history means diff0 is behavioral evidence, not a controlled isolated renderer pixel regression. No new render components exist.

## Inspected artifacts

Opened/looked all four delivered older main recordings via independently regenerated timestamped ffmpeg contact sheets, plus subsecond B/C frames; also all final delivered main AB-head736.mp4 and CD-head736.mp4 through new sheets covering their full duration (including CD tail), and original A/B736 recordings. Read both evidence indexes, matrix, baseline/runtime-complete/head736 runtime exports, and hash manifests.

Images opened: before.png; after.png; diff0.png; diff0-amplified.png; diffT.png; B-last.png; C-request.png; C-last.png;1440-light.png;1440-dark.png;390-light.png;390-dark.png; B-head736.png;after-head736.png;diff0-head736.png;diffT-head736.png; generated audit contacts/selected full-resolution frames and supporting delta. Compared the supplied stability pairs with PIL independently.

## Explicitly unverified dimensions

Representative real A–D flow is1280x800 light only. Prior1440/390 light/dark settled-history stills support earlier revision layout only, not final-head responsive flows. Matching before1440/390 light/dark,1280 dark before/after, every size/theme interaction, native reduced-motion emulation, approval request family, explicit late approval/cancellation runtime, Electron/native mobile, alternate providers, relay/tunnel are unverified. No full provider/platform matrix is required for this unchanged renderer/contracts scope. Focused cancellation regressions are reported97/97; do not relabel them as live cancellation proof. Seed helpers remain outside the repository per user hygiene override.

## FRICTION category | happened | avoidance

| Category | Happened | Avoidance |
|---|---|---|
| Authentication | Fresh auditor tab redirected to /pair during read-only server window | Arrange authenticated own tab before the capture window; retain server until auditor confirms capture. |
| Temporal sampling | Coarse contact sheets missed old C pending card | Extract subsecond frames around correlated native request time. |
| Evidence narration | Old B loading-only and zero-delta claims contradicted pixels | Read actual frames and measure each pair; final index corrects these. |
| Timing | B wall interval differs from media duration | Distinguish tool-operation wall time from encoded video time. |
| Revision | Earlier A–C98b/D dirty did not prove final | Fresh exact736 captures and matching hashes now close this gap. |

VERDICT: SUFFICIENT for scoped A–D proof, with explicit limitations above.