# Proof audit pass 2

Read-only audit of [#13706](https://git.cafe/pingdotgg/t3code/issues/13706), branch `fix/issue-13706`, [b5a226e49c](https://git.cafe/pingdotgg/t3code/commit/b5a226e49c52ecab4ddc2d678a7012d97e651a6a) → [e84990bae5](https://git.cafe/pingdotgg/t3code/commit/e84990bae57049749dc7ddb0835e15452eb336ab). No PR exists for this work. Audited canonical fixture, materializer, full source diff, v2 manifest/metrics/traces/stills, recording contact sheets, control traces, and final evidence report. No tracked edits, public writes, uploads, pushes, server launches, DB access, or provider requests. Own preview tab closed; existing servers left running.

## Decision

The essential A1 dependency race is sufficiently demonstrated. The full pass-1 matrix is not complete. These are evidence gaps, not newly established production correctness defects. Opening a narrowly worded PR can follow the maintainer's stated hard stop; this verdict does not silently waive unverified matrix entries or grant merge approval. Claims should remain confined to the controlled dependency race and separately identified integrated smoke. Do not claim a deterministic product reproduction or all-platform verification.

## Independent rebuild required by the brief

Reused real vp dev on port7050, own tab opened with reuseExistingTab:false. HEAD independently resolved to e84990bae57049749dc7ddb0835e15452eb336ab. Canonical fixture and existing generated TSX both SHA256 d3153cb5e491fbce893a1180dcba4bf9f595f04b84003c5dd7ed66437958c1a4; materializer ce49853d7c74f7418f8a603eaa82783657246c16676104de228a2d1181869364. Existing materialization was byte-identical to committed source, so no rematerialization was necessary. Rebuilt the affected state independently by running the committed prepare/run driver, not by reopening a delivered still.

1440×900 CSS, light, DPR2, Chromium154 Linux, visual scale1, loaded16px/20px monospace, reduced motionfalse; environment matches manifest. prepare() returned150 rows,total135000,DOM/internal offset5000,viewport800. Started recording at2026-10-22T00:42:39.861Z; ran window.repro.run() with defaults. At151.8ms active index149/viewPosition1/itemSize900/targetId149; at187.6ms removal produced149 rows,targetIdnull,total134100. At1890.1ms settled DOM/internal offset133300,enddistance0,Row148 label439.5..459.5 visible. Two subsequent animation frames and later sample at20010.3ms retained identical geometry. Recording includes idle tool-call time after settle; timestamps are trace-relative, not a promise of identical video PTS.

Saved [rebuild trace](/tmp/13706-proof-rebuild-trace.json), [recording](/tmp/13706-proof-rebuild.mp4), [still](/tmp/13706-proof-rebuild-settled.png), [metrics](/tmp/13706-proof-rebuild-pixels.json), [raw diff](/tmp/13706-proof-rebuild-diff1x.png), [8x diff](/tmp/13706-proof-rebuild-diff8x.png). Comparing rebuilt still against v2-A1-1440-light-after-settled.png at exported1280×800 raster: changed pixels threshold0=0,threshold8=0,max channel delta0. This is an actual exact pixel match, not antialias tolerance or masked comparison. Wide screenshot export scaling is disclosed separately from1440×900 CSS/DPR2 layout. No independent font-file assertion beyond computed font and matching raster.

## Provenance and A1 evidence

All manifest-listed artifact SHA256 hashes independently matched. Manifest declares before b5a226e49c and after e84990bae5, same final fixture on both. Verified fixture/materializer against worktree bytes. Historical runtime SHAs are supported by manifest and bundle hashes rather than independent time travel; current head was checked directly. Display clock is a fixed label; animation and performance clocks remain live.

Every A1 width/theme has both recordings, ready/settled stills, traces, unamplified1x and amplified8x diffs, binary threshold8 and paired panels. Threshold0 means any RGB channel differs; threshold8 means maximum channel delta>8. Independently recomputed comparisons match delivered metrics. Ready pairs are all exactly0. Settled changed>0/changed>8/maxdelta: wide light3985/3923/238; wide dark4045/3961/215; narrow light6948/6726/238; narrow dark7152/6854/215. Expected differences are final row identity/label placement and scroll/end readout, including wrapping. No unrelated visual change established. White header-only recording contact sheets are cropped diagnostics, not evidence that full rows are absent; the full rebuilt still visibly contains Row148.

Scripted steps for all A1 recordings: await fonts/readiness; prepare public size warm-up150×900; require5000/135000; record; animated scrollToEnd; remove final row requested150ms; observe stable geometry>=200ms and>=3 frames after minimum1200ms; screenshot; observe two more frames; stop. Every before has total90 immediately after mutation and ultimately lands at0/enddistance133300; every after retains134100 and lands133300/enddistance0. Active index149/viewPosition1 exists at mutation. Four pairs are four tested runs, not a statistical reliability study. Wall-clock scheduling differs and video frames can miss transient states; traces establish collapse, recordings show visible landing.

| A1 recording pair | before mutation start/end ms; settle ms | after mutation start/end ms; settle ms |
|---|---|---|
| 1440 light |160/204.2;1612.9|180.9/185.9;2524.6|
| 1440 dark |174.1/187.2;1796.2|151.3/156;1860.9|
| 390 light |195.6/205;1482|154.1/174.4;1578.8|
| 390 dark |151.1/158.2;1496.1|154.6/158.1;1584.6|

## Entire matrix disposition

| Row | Audit status and required evidence still missing |
|---|---|
| A1 | PASS controlled dependency race, four web viewport/theme pairs; independent rebuild also PASS. Narrow web is not native. |
| A2 | UNVERIFIED size-only callback. data=false,size=true plus shrink props/traces do not prove actual size-only prepareMVCP invocation. No accepted direct callback proof or focused size-only branch test. Exploratory material cannot close this row. |
| C1 | INCONCLUSIVE. Paired resize recordings/traces exist; retained target149/itemSize900→1000, but total135100→135000 means final resized correctness is not proven. Extra public setItemSize timer at200ms and actual measurement complicate control. Do not call preserved resize behavior PASS. |
| C2 | PARTIAL. Paired recordings/traces exist for MVCP-disabled,viewPosition0,nonanimated controls; use fresh-page verified runs only. Missing pending-end-before-target,ordinary-anchor restore,explicit data-only and complete selected-branch proof. Controls have scripted options, but no complete timestamp-indexed acceptance package for remaining phases. |
| P1 | UNVERIFIED deterministic paired product shrink/finalization race. Negative long-thread report cannot substitute. No paired replay recording/data-key/active-target proof. |
| P2 | PARTIAL smoke only. Real short/long Claude,busy-JS reportedly negative. Final actual-app OK smoke ended at0.5px enddistance; recording covers pending,not completion. Integrated report labels older runtime12b1ac53eb,not final e84990bae5; production guard unchanged, but it is not a final-head four-pair replay. No committed sanitized product transcript replay or deterministic busy duration. |
| P3 | UNVERIFIED. No required product empty/short,remembered-middle,find/citation,reduced-motion recordings and timestamped scripted steps. Dependency nonanimated control is not equivalent. |
| P4 | UNVERIFIED. No nested tool-group append/replace/remove and collapse/reopen recordings; native work-log counterpart also unverified. |
| N1 | UNVERIFIED iOS and Android feed/keyboard/insets. No actual native client recording/replay. Hard-stop routing may proceed separately;390 web does not close this. |
| D1 | UNVERIFIED actual Electron scroll behavior. Shared React bundle proof does not close shell integration. Maintainer hard stop permits PR preparation with this disclosed. |
| U1 | UNVERIFIED representative ancillary scroll-API controls and native catalogs; no recordings/timestamped scripts. No positive null-target exposure established. |

Control trace timestamps (before-mutation / after-mutation / settled ms), paired recordings under `/tmp/13706-evidence/`: C1 before196.9/198.9/2449.6,after194/196.5/2697.7; C2 MVCP-disabled before161.2/166.4/2796.9,after181.6/186.2/2787; viewPosition0 before153.9/166.7/2416.8,after165/168.5/3470.4; nonanimated before176.7/181.9/2408.4,after159.1/162.5/2485.2. Script: fresh page,prepare,record,run respective options `{action:"resize"}`,`{data:false,size:false}`,`{viewPosition:0}`,`{animated:false}`,wait settle,stop; C1 additionally requests public size1000 at200ms. These existing control artifacts predate final label/stability fixture and are not final-head v2 captures.

## Source and scope

Only production behavior added against base is targetId!==null before positive-viewPosition target size correction, identically across six React/native/RN-web CJS/ESM bundles. getId returns null outside data. Guard prevents getItemSize from estimating/caching a missing target and corrupting total. Valid target path remains unchanged by inspection; C1 integration remains inconclusive. No concrete defect found in this narrow guard. Lockfile selects updated patch hash; other patch body behavior predates this change; full-index metadata changes are formatting.

Focused test adds real bundle-function extraction and total-size regression across react.js/mjs and react-native.js/mjs. Reported red then14pass is credible and source unchanged after fixture-only updates, but not independently rerun here. Test does not prove scroll landing or native runtime, and RN-web direct parity is untested. These limits must not be promoted to all-platform coverage.

Committed fixture/materializer are explicitly required by user, overriding generic scratch prohibition. No defect for committing them. Repro conditions beyond reporter include border-box900px rows,100px header,fixed display clock/theme,public size warm-up,flushSync mutation,read-only fiber trace,centered labels and stability completion. All are shared by before/after; none establishes natural product reachability. No DB snapshot applies to synthetic rows. Generated HTML/TSX were untracked at initial inspection; final git status is clean, consistent with primary cleanup. Audit made no repository edits.

## FRICTION

- tool: First trace parser assumed array; artifacts wrap trace in an object. Corrected parser and verified all eight traces; no evidence discarded or invented.
- brief: Broad matrix remains unverified despite essential A1 success; status disclosure is not a waiver.
- infra: Variable frame delivery/timers prevent identical elapsed-time alignment. Preserve requested and actual times; judge semantic phases and stable geometry.
- policy: No public upload because brief forbids public writes. Report and independent evidence remain local; primary can publish within its authorization.

VERDICT: INSUFFICIENT A2 C1 C2 P1 P2 P3 P4 N1 D1 U1