# Proof audit pass 2 — issue 24896

Audit outcome: **INSUFFICIENT for the full requested matrix.** The inspected web badge geometry is convincing, including an independent reconstruction of one final capture. This is a proof-coverage verdict, not a new code blocker.

Scope: [#24896](https://git.cafe/pingdotgg/t3code/issues/24896), final [aafc777fe6](https://git.cafe/pingdotgg/t3code/commit/aafc777fe6acf80de1eb55ca1408d4df4e4e52de), merge base [e3592efe7f](https://git.cafe/pingdotgg/t3code/commit/e3592efe7f5dbd3a869e46e608d8f6815e1bb7f6). Source remained read-only; no push, forge comment, merge, or watcher. The lead reported the final changed checks passed; this audit did not rerun or independently certify them.

[Public evidence index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/de7e118451457cc7/index.md) · [Matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/79aad3cd95015bac/issue-24896-matrix.md) · [Manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/da48223428f458c6/processed-manifest.json) · [Diff metrics](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ff3a02c6a6ecc1f1/diff-metrics.json) · [Handoff](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4ed3fcc1b091ded8/handoff.md).

## Bounded row findings

W grid means CSS 1440×900 light/dark and CSS 390×900 light/dark.

| Row | Actual visual proof | Status / remaining gate |
|---|---|---|
| W1 | All four cells, collapsed and expanded; eight before/after recordings. CL and VY initials remain readable through the three-avatar stack and beside +1. Expanded completed/failed dots are top-left, badges bottom-right. Long title truncates without colliding with failure/time labels. Independent fresh-load reconstruction matches the published final timeline exactly. | **PARTIAL.** Synthetic projection fixtures in the actual app prove geometry and disclosure. No comparable organic Hello-task before/after pair. Full-frame composer differences prohibit an isolated whole-screen regression claim. |
| W2 | Yellow singleton has readable VY badge and opposite dot in all four cells. Plain singleton has no badge and retains bottom-right failed dot; every plain before/after pair is exact-zero. | **PARTIAL.** Sole-enabled settings were applied with restart, but no runtime provider-registry snapshot establishes effective account eligibility. The matrix's singleton group-header request is unreachable: the event renderer only mounts the group for more than one member; singleton lifecycle captures are the correct reachable path. |
| W3 | Yellow and neutral workflows, all four cells. Phased clickable completed member and unphased noninteractive failed member show readable badges, separate dots, and no narrow title/detail/time collision. | **PARTIAL.** Members are settled, but full captures visibly retain a “Starting” child footer. They do not satisfy the requested fully settled/no-loading screen criterion. |
| W4 | Missing account with known Claude driver and updated notification, all four cells. Provider mark retained, no invented badge; update row has no status dot. All four before/after pairs are exact-zero. | **PARTIAL.** Absent-driver Bot fallback has no runtime visual proof. Capturer reports the current scene schema requires driver and did not forge invalid wire data; no claim that this branch passed. |
| M1 | No fresh native before/after captures inspected. | **INSUFFICIENT.** No completed current-source EAS build/session proving multi-account overlap, theme separator, +1, or historical missing account. |
| M2 | No fresh native before/after captures inspected. | **INSUFFICIENT.** Sole plain-account and accent-alone native branches unverified. Prior builds are not evidence for final HEAD. |

No provider-generic ACP visual claim is supported by this evidence; the supplementary distinct-ACP-agent runtime check remains unverified. This audit does not reopen the completed code reviews.

## Provenance, viewport, and diff validity

The actual baseline captures use [01fe8d9188](https://git.cafe/pingdotgg/t3code/commit/01fe8d91887af5e0669a0378b77a97c9114c43c6), **not** the merge base. I independently ran the baseline/merge-base comparison: all three affected renderer files are identical (git diff exit 0). The full baseline-to-merge-base file list contains server/provider/install changes, not web/mobile rendering changes. Capturer identifies final captures as exact aafc777fe6; the independent reconstruction verified that SHA immediately before and after its run.

DOM metadata records the labelled CSS viewport. Desktop PNGs are 1280×800, an 8/9 raster reduction of CSS 1440×900; narrow PNGs are 780×1800, 2× CSS 390×900. Desktop recordings are 2880×1800, also 2×. Thus raster dimensions do not contradict the CSS-width labels. Every compared still pair has matching raster dimensions.

I visually inspected full-context sheets, enlarged relevant avatar/row crops, all changed-region sheets, amplified/mask sheets, and sampled frames from all eight W1 recordings. I independently recomputed all 120 manifest hashes and all 84 diff pairs: 112 PNGs, eight MP4s, 28 before/after comparisons, 56 stability comparisons. No hash, pixel-count, or amplified-image mismatch occurred. All 56 stability comparisons are exactly zero. All four plain-singleton and all four missing-account before/after comparisons are exactly zero.

For **every pair**, the index supplies both a zero-threshold amplified difference and a threshold mask, including zero pairs. Zero threshold counts any RGB-channel difference; amplification multiplies each absolute channel delta by 8 and clips at 255. The threshold mask counts pixels whose maximum absolute RGB-channel delta is **greater than 16/255**. No pixels were ignored in these published full-frame metrics. Black masks were inspected, rather than inferring “unchanged” from matching labels.

W1 needs a narrower statement: in each cell, timeline pixels outside the avatar gutter are exact-zero before/after when considering the top 600 CSS pixels (top 1200 raster rows narrow; top 530 raster rows desktop), excluding only the avatar gutter. The composer region differs substantially at 390px and slightly on desktop. The full-frame diff remains valid and includes those differences; it is not clean isolated patch proof. The collapsed header's muted appearance is the existing opacity-55 style, not a loading ghost. Captured metadata reports fonts loaded and zero toasts. W3's visible Starting footer remains explicitly disqualified from a no-loading claim.

## W1 recording inspection timestamps

These are sampled visible milestones, not exact click-dispatch times. Each clip begins collapsed, shows the four settled rows expanded, then ends collapsed. No user message or provider turn was submitted.

| Cell | Before: collapsed / expanded / collapsed | Final: collapsed / expanded / collapsed |
|---|---|---|
| 1440 light | 0.00s / 3.48s / 4.87s | 0.00s / 3.30s / 4.61s |
| 1440 dark | 0.00s / 3.40s / 4.76s | 0.00s / 3.28s / 4.59s |
| 390 light | 0.00s / 1.89–2.94s / 4.11s | 0.00s / 1.98–3.08s / 4.31s |
| 390 dark | 0.00s / 1.87–2.91s / 4.06s | 0.00s / 1.89–2.94s / 4.10s |

The public index links all eight recordings by phase, width, and theme. The before/after stills provide sharper avatar inspection than the reduced contact frames.

## Independent reconstruction

I downloaded the public [seed](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2596c4ac5c4ba7e7/seed.py) and [driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/121f8c3336836bd4/capture-driver.js) with curl, compared them byte-for-byte to the handoff copies, and reran the seed only while my fixture server was stopped. Seed SHA256: d8bbef9f6ba12e9b277081e1496a59ab28bc26aa57ba06918cba3b22b3791df9. Driver SHA256: 650ac0402c335c9ed5181c670242579132d0dd09c951931f1568b5f908b4e6fb.

The preserved checkout `/home/ubuntu/wt/issue-24896-evidence-base` was clean at final HEAD. I used its `.t3`, never the primary checkout's `.t3` or live `~/.t3/userdata` database. The seed relies on the handed-over copied database containing the original parent records; public scripts alone do not recreate that private starting database from nothing.

I started `vp run dev --no-seed`, paired a fresh incognito Browser panel, and ran the downloaded driver with only its width/theme loops restricted to 390/light. The same production renderer, route, account configuration, disclosure steps, DOM metadata, twins, and recording calls were retained. No source or screenshot-only avatar markup was introduced.

Rebuilt W1 expanded: CSS 390×900, raster 780×1800, fonts loaded, zero toasts. Against the downloaded [published final still](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/70a3f0c08c29fda3/head-aafc777fe6-W1-390-light-expanded-twin1.png):

- Full-frame changed pixels: **113,715** at zero threshold; **22,582** at >16.
- Difference bounding box: **[21,1483,759,1773]**, entirely in the composer region.
- Entire top 1200 raster rows, including every avatar, badge, dot, title, timestamp, and surrounding timeline pixel: **0** at both thresholds.
- Rebuilt collapsed twins and expanded twins: **0**.

I inspected the rebuilt full image and its amplified/mask images visually. The final avatar geometry reproduces; the whole frame does not. Rebuild PNG SHA256: ae267f1b20818d0de2f9ce9b655d5f53221aa760bf36b8252ef9dbae487deedd. Original published PNG SHA256: f016b5df45123edda4687128944716c8590d939ac3167feaacfd7c2a579a2828.

Local reconstruction evidence is retained under `/tmp/issue-24896-audit-rebuild/`: manifest.json, metrics.json, rebuilt.png, amplified8.png, threshold16.png, and timeline-comparison.png. The manifest also identifies the four original rebuilt stills and recording. Auditor public-write authorization was absent, so these new artifacts await lead publication.

Both auditor-owned server sessions (55077 and 38823) were stopped through their tracked terminal sessions. Final `ss` confirmed ports 5821 and 13861 closed; source/index remained clean. The primary server and primary state were never used.

## FRICTION

- Fresh profile mattered: the first browser profile displayed stale/syncing projection state; fresh incognito reproduced the fixture. No loading frame from that attempt was accepted.
- Preview navigation to an existing pairing route did not consume the hash until a full reload; pairing was completed before capture. No pairing credential is included here.
- Python urllib received HTTP 403 for the public seed; curl downloaded the exact assets successfully.
- The fresh final-source rebuild retained a different composer layout from the published final still. The timeline matched exactly; this narrows the claim and does not identify or fix the unrelated composer cause.
- Effective sole-account runtime config, organic before/after, absent-driver fallback, fully settled workflow screen, and native proof remain explicit pending gates. Settings files, old builds, source inspection, and stable screenshots do not substitute for those runtime claims.

Model/harness: GPT-6.1 Sol through Codex in T3 Code. Read-only proof audit; no source edits or public writes.

VERDICT: INSUFFICIENT W1, W2, W3, W4, M1, M2