# Issue 19307 fresh proof audit pass 2 — frozen 66366

## Findings

**No unresolved blocking proof findings. All 24 recording rows are SUFFICIENT for the agreed synthetic web proof scope.** Both responsive V2 notice-details endpoints have a disclosed 112-raster-pixel position mismatch; neither establishes exact layout equivalence. Its proposed anchor/compositor cause remains inference. These limitations do not erase the demonstrated Show navigation and control behavior.

Read-only proof audit for [#24336](https://github.com/pingdotgg/t3code/pull/24336) / [#19307](https://git.cafe/pingdotgg/t3code/issues/19307), not a fourth whole-PR review. Before [ff1f6fd884](https://git.cafe/pingdotgg/t3code/commit/ff1f6fd8849cdd2e0a07096eac4fd4440d1ad989); After [66366dff35](https://git.cafe/pingdotgg/t3code/commit/66366dff35fb0307402cb534e1da74a1c7d56326). Checkout remained clean at full After SHA. Against the previously audited a94, only the fixture and seeder changed; app/package product files are byte-identical. The prior audit's passing After artifacts were not used to infer the fresh result.

## Frozen publication and byte audit

Inspected [final index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/35e93eb1074f9790/19307-66366-final-evidence.md) and [manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/71ee731a59bc06e9/19307-66366-final-manifest.json). Independently downloaded every **500 public artifact, 23,631,825 bytes**, matching SHA256 and byte count. Both top-level downloads are byte-identical to local files. All 468 artifact URLs in the index map to the manifest; its additional top-level manifest link is expected. All 501 local frozen files recorded at validator start remain unchanged.

Index SHA256 `cb46d1612e715a6dc9ea86ab9cc3c9066a90c7d82fac9a8df5e497e6b571a3e7` (119,237 bytes); manifest SHA256 `a9e31d9890872d1f78c2f39f5334f361a5275ef40f88f961c18d5120ba62b89e` (1,025,304 bytes). Fixture module SHA256 `c46db702527facc8382e091cc51175a66f3811fa644a040c7a8c8aa537ec4c17`.

All 130 retained Before media/still files are byte-identical to the prior frozen evidence. All 12 new After raw recordings differ from the old After recordings. Historical Before fixture provenance is disclosed; Before is not relabeled as final Date-wrapper capture. Fresh After artifacts have fresh public uploads even where endpoint bytes match old pixels.

## Independent native browser rebuild and fixture checks

Read test-t3-app and ghdrop. Opened an owned native T3 preview tab, minted a private pairing credential against the retained isolated backend, then loaded the real persisted parent route. No other agent's cookies or tab were used. Imported the committed fixture source over Vite, built and schema-decoded all 36 combinations, and independently ran the actual web derivation over all 36: F0=0 tasks, F1=1 prior command, F2=1 linked child, F2-watch=1 monitor without Show, F3=10 canonical mixed tasks, F4=2 prior roster command/monitor. Null projection returns no tasks. Native foreground identities remain excluded across phases.

Reinstalled the committed browser installer and checked preservation of the first native Date constructor, replacement of the wrapper, callable Date(), zero/one/seven-argument construction, invalid dates, parse/UTC, real wall-clock and performance advancement, and stable pinned time. All checks passed. Independently generated CLI installer is exactly 206,677 bytes and byte-identical to the owner's installer, SHA256 `243462d2738d9c7346530b3bc13da2256216ab118c10d6a81bc5bd7eb7b86a8a`.

Installed F4/running then F1/running, collapsed sidebar/details through UI, pinned light/clock, awaited fonts, bottom-scrolled, blurred, moved pointer to (100,400), and allowed two RAF. Compared against the NEW final V1-1440-light-after-running PNG. **Full-surface RGBA Diff0=0, T16=0, maximum channel delta=0 across 1,024,000 pixels.** Initial immediate sidebar raster settle had 27 changed pixels (one T16, max20); another committed pin/two-RAF settle produced exact zero without changing content.

![Independent final After rebuild, full surface, exact zero RGBA delta](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6a4e69a2c8f9e6c5/independent-66366-rebuild.png)

CSS 1440×900, DPR2; PNG1280×800 is the complete surface at 8/9 scale. Responsive CSS390×844 exports PNG780×1688. A separate read-only persisted F0 semantic/index/schema check passes with only the parent's lastVisitedAt change. Fresh strict pre/post-startup assertion files are nonempty and pass before browser visits; parent and child persist. No active-work recovery mutation is normalized.

## Recording rows

Every row passes metadata, viewport/DPR/computed-theme, phase/control/route milestones, raw/export availability, frame order, exact PTS and duration checks. Independently ffprobed **2,125 frames: 1,171 Before +954 fresh After**. Raw/export frame counts, every PTS, VFR gaps, duration and final hold match exactly.

| Recording | Status | Frames raw/export | Duration raw/export (s) | Max PTS / duration delta |
|---|---|---:|---:|---:|
| V1-1440-dark-after | SUFFICIENT | 47/47 | 10.790867/10.790867 | 0 / 0 |
| V1-1440-dark-before | SUFFICIENT | 78/78 | 23.593200/23.593200 | 0 / 0 |
| V1-1440-light-after | SUFFICIENT | 49/49 | 11.004800/11.004800 | 0 / 0 |
| V1-1440-light-before | SUFFICIENT | 78/78 | 23.477900/23.477900 | 0 / 0 |
| V1-390-dark-after | SUFFICIENT | 59/59 | 9.752533/9.752533 | 0 / 0 |
| V1-390-dark-before | SUFFICIENT | 107/107 | 22.451433/22.451433 | 0 / 0 |
| V1-390-light-after | SUFFICIENT | 59/59 | 9.860800/9.860800 | 0 / 0 |
| V1-390-light-before | SUFFICIENT | 109/109 | 23.488200/23.488200 | 0 / 0 |
| V2-1440-dark-after | SUFFICIENT | 123/123 | 16.096400/16.096400 | 0 / 0 |
| V2-1440-dark-before | SUFFICIENT | 111/111 | 24.090167/24.090167 | 0 / 0 |
| V2-1440-light-after | SUFFICIENT | 128/128 | 16.610033/16.610033 | 0 / 0 |
| V2-1440-light-before | SUFFICIENT | 115/115 | 27.179133/27.179133 | 0 / 0 |
| V2-390-dark-after | SUFFICIENT | 208/208 | 16.786433/16.786433 | 0 / 0 |
| V2-390-dark-before | SUFFICIENT | 206/206 | 22.481367/22.481367 | 0 / 0 |
| V2-390-light-after | SUFFICIENT | 195/195 | 15.357567/15.357567 | 0 / 0 |
| V2-390-light-before | SUFFICIENT | 207/207 | 21.964233/21.964233 | 0 / 0 |
| V3-1440-dark-after | SUFFICIENT | 16/16 | 3.497967/3.497967 | 0 / 0 |
| V3-1440-dark-before | SUFFICIENT | 33/33 | 6.808467/6.808467 | 0 / 0 |
| V3-1440-light-after | SUFFICIENT | 18/18 | 3.939533/3.939533 | 0 / 0 |
| V3-1440-light-before | SUFFICIENT | 29/29 | 7.400433/7.400433 | 0 / 0 |
| V3-390-dark-after | SUFFICIENT | 26/26 | 2.982300/2.982300 | 0 / 0 |
| V3-390-dark-before | SUFFICIENT | 50/50 | 6.384900/6.384900 | 0 / 0 |
| V3-390-light-after | SUFFICIENT | 26/26 | 2.886233/2.886233 | 0 / 0 |
| V3-390-light-before | SUFFICIENT | 48/48 | 6.421400/6.421400 | 0 / 0 |

Independently decoded and compared every raw/export app-region frame in order at 288×180. All 2,125 frames correspond, maximum per-frame mean absolute RGB delta 1.995094 from lossy compression. Five exact-frame playback samples of each of all 24 exports were visually inspected; this is not continuous human viewing of every frame. The footer clock is encoded frame PTS and holds during gaps. Ordered guides are static, not action-synchronized captions. The final index puts host action timelines ABOVE endpoint comparisons; host times include API latency and are not proportionally mapped to media time.

V1 includes idle, queued, queued with selected waiting runtime, preparing, starting, running, waiting, idle-return, actual Stop click with held transport, simulated acceptance retaining the task, and separate F0 removal. Fresh Stop is disabled during connecting/nonwaiting foreground work, enabled for selected waiting/idle, disabled Stopping while held, and enabled again after synthetic acceptance until removal. Before missing controls are documented.

V2 includes actual idle Show navigation to persisted child and parent return, running/waiting, fresh After running Show navigation, and watch-only idle/running/waiting without Show. Responsive navigation uses notice details. Before running Show is honestly unavailable. V3 includes roster-only idle/running/waiting with exactly the prior command and monitor and excludes foreground work. F0/F3 remain logic-only; no V4/new native-child fixture was added.

## Full-surface pixel audit and collateral

All **90 full RGBA pair counts, T16 counts, percentages, bounding boxes, maxima and BOTH actual diff PNGs per pair reproduce independently**. Diff0 is saturated absolute RGB delta amplified4; T16 marks maximum channel delta>16. Every source pixel is alpha255, making the original RGB and independently recomputed RGBA counts equivalent. No masks or pixel editing. All24 same-revision repeats compare exact zero. Of90 pairs,24 are exact zero and66 nonzero.

Active-state banner/Stop/queued displacement and composer layout are intended. Settled collateral was inspected against source images and is explicitly disclosed:

- Four required V1 responsive idle/idle-return pairs: Before hovered Stop versus neutral After; light4,809/max11, dark4,781/max9, allT16=0, bbox608–703/1418–1473.
- Other Stop hover: desktop light788/max12, dark773/max10; responsive light4,809/max11, dark4,781/max9; allT16=0 where isolated. These include V1 idle/return/accepted and V3 idle/waiting as allocated.
- Desktop V2 linked idle/returned/waiting Show hover: light193/T16159, dark200/T16183.
- Desktop child Git context: Before Select ref / Commit / +0−0; fresh After fix branch / Push / +813−14. Light1,102/T16932; dark1,100/T16953. Both actual panels are unmasked.
- Responsive V2 notice-details: same Show popover displaced112 raster pixels, Before top1246 versus After1358. Light13,852/T161,750; dark13,872/T161,848, bbox516–628/1246–1423. Other settled idle/child/return surfaces match. This is disclosed capture-position mismatch; no automatic reposition or causal compositor explanation is established.

V3 responsive-light idle's first endpoint had an additional124-pixel dot-opacity residual (T16121). The owner replaced it with a fresh stabilized replay, not pixel editing, leaving only4,809/T160 historical Stop hover. Frozen row carries captureNote; discarded initial PNG is outside the frozen set; raw recording unchanged. Its final pixels and all90/24 metrics were independently checked after replacement.

## Source reachability and claim boundaries

Confirmed the narrowly relevant finding6 source guard: provider-native subagent requires subagent lineage AND creationSource=provider; ChatView replaces its composer with ProviderSubagentBar, which has no Stop. Composer background strips cannot expose the claimed child Stop. Shared derivation with no latestRun returns only PR-watch tasks before provider roster enumeration. Finding6 is source-backed false positive and needs no fabricated child Stop capture or fixture extension.

Runtime/client projections and Stop acceptance are simulated. These artifacts do not establish provider termination, live provider execution, native mobile, Electron shell/IPC, remote/relay/tunnel synchronization, or multi-device behavior. F4 is Claude-shaped metadata. The final index and refreshed PR draft retain these limits.

Quick-checked the refreshed PR draft after root steer: all74 public links and all12 running tables map to correct final assets; base/head,2125frames,90pairs,24repeats, both popover inference, actual Git context, timeline placement and simulation limits agree. Accepted exact Fixes line is unchanged. PROOF_AUDIT_PENDING is the expected replacement placeholder. Canonical91tests/32packages and17Date-test suite counts are root's execution attestation; this proof auditor did not re-run package suites or recertify historical suite counts.

## Independent machine-readable audit data

- [frozen-independent-timing.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/846b448adb1b3095/frozen-independent-timing.json)
- [independent-pixels.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2146c5a1de58eb4a/independent-pixels.json)
- [public-hash-verification.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a1a9a63185d7abc9/public-hash-verification.json)
- [independent-row-status.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/546da7acd2e7c1ca/independent-row-status.json)
- [frame-content.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8bb8c01267db91a2/frame-content.json)
- [date-browser-checks.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9e890365ba2713f1/date-browser-checks.json)
- [independent-logic-results.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/28adfd2d9cbd1af6/independent-logic-results.json)
- [rebuild-comparison-final.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/38207a44f5ab5990/rebuild-comparison-final.json)

## FRICTION

- Supplied sibling ID omitted leading thread:; both initial forms returned thread_not_found. Live catalog revealed exact ID and direct coordination succeeded.
- Native preview initially closed; own preview_open succeeded. Private pairing completed without exposing a credential. No startup log was published.
- Initial browser import treated browserInstaller as a function; it is a string. Evaluated the committed string and proceeded. Sidebar selector rejected once; refreshed snapshot ref succeeded.
- Initial direct raw shared-helper call with queued latestRun correctly returned zero; actual web wrapper selects prior eligible runs. Independent36 results use the same web wrapper as ChatView.
- Immediate sidebar raster settle and V3 owner endpoint replay are disclosed above. Both popover mismatches remain honestly nonzero.
- A first exact float-equality assertion on computed percentages was too strict for floating arithmetic; independent recomputation matches within1e-12 and integer counts/bounds/maxima match exactly.
- Native decode-to-rawvideo emitted temporary muxer DTS rounding warnings; source ffprobe timestamps and frame counts remain exact. No media rewritten by auditor.
- No product/fixture/repository edits, pushes, commits, forge writes, merges, agents, heavy checks, server starts/stops, process signals or live T3 database writes. Only owned browser/auth state and temporary audit reports were changed. Authorized ghdrop uploads contain synthetic evidence and audit metadata only. Capture owner cleaned up its retained servers after our early rebuild success under root authorization; isolated DB and artifacts remain.

VERDICT: SUFFICIENT