# Proof audit PASS2 — issue24745

Findings first: no blocking implementation or proof defect found. C/Q/L/N/R have sufficient frontend evidence. H remains explicitly unverified at runtime; its transient typing hold clears on blur, so real terminal selection cannot preserve that state. A is appropriately source-only. This verdict does not extend to native mobile, Electron shell, relay/tunnel, or provider/backend lifecycle behavior.

Audited clean head [4e080eb004](https://git.cafe/pingdotgg/t3code/commit/4e080eb004ead1a450214326e52bec98c9f52dba) against [c79ef8c454](https://git.cafe/pingdotgg/t3code/commit/c79ef8c454e1268ff79da5a9dcb74572e86f7861). One changed file, ChatComposer.tsx, 11 additions/2 deletions. The active-thread return precedes the toast; shown-question guard precedes snapshot/context/draft mutation; activePendingProgress remains in the imperative handle dependencies. Normal insertion/focus code is unchanged. Capture SHA provenance is recorded by the capture author/index, not cryptographically embedded in individual screenshots; the independently rebuilt head capture agrees with the supplied after image.

Nonblocking evidence qualifications:

- The narrow C claim of zero pixels is threshold-dependent: 32,950 pixels outside masks differ by only 1–2/255 at zero tolerance. None exceed 8/255. Desktop C/Q/L are exactly identical outside the disclosed masks even at zero tolerance.
- The N tolerance was selected during capture review. I independently checked its 125 RGB pixels above 8/255: age label, composer caret and thin sidebar raster; no unexplained content difference. The permissive number alone was not used as approval.
- Toast masks cover expected overlay/shadow and the underlying timeline/diff text it occludes, not the question/answer/choice. I inspected all five diff images and all matched before/after captures. The extra light390 still has stacked notifications from repeated genuine actions; text remains readable and answer is unobscured.
- Bundle index differs from the current public/local index only by the later-added bundle self-link. All other locally present bundled file bytes match. Public index fetched using curl with a browser User-Agent is byte-identical to the supplied local index. Python urllib's default request received HTTP403.

Row statuses:

- C desktop 1440x900 light — PASS. Base recording shows actual hello selection/Add to chat silently; hidden-draft state records answer-before plus terminal context. Fix shows info toast and preserved answer, draft-before/zero contexts. Independently rebuilt through supplied driver and real UI, with same result.
- C responsive 390x844 dark — PASS. Both recordings show terminal selection/menu. Fix toast wraps within viewport, question and answer remain visible; recorded state says no horizontal overflow. 1440 dark and 390 light stills complete four width/theme combinations.
- Q desktop — PASS. Fixture explicitly allowCustomAnswer:false, multiSelect:true, no initialAnswer. Beta remains selected through real terminal action. Base silent; fix toast, unchanged draft/zero contexts. Multi-select is a disclosed fixture choice to avoid single-select auto-submit.
- L desktop — PASS. Fixture explicitly initialAnswer="literal @test `hello`", allowCustomAnswer:true. Before/after recordings retain the literal, with fix toast and zero draft-context delta.
- R desktop — PASS. beforeRecovery includes selection/action and real dismissal exposing leaked answer-before/reference. afterQLRN closes questions, visibly shows draft-before with no terminal reference before retry. State observations confirm zero contexts. Independent C dismissal also revealed draft-before/zero contexts.
- N desktop — PASS. beforeNormal and afterQLRN show normal selection/action adds one terminal reference while preserving draft-before; no rejection toast attributable to retry. The continuous after clip has earlier rejection toasts until they naturally expire; these are not new N rejections. State records contain exactly one hello context after retry. Focus behavior is supported by caret and unchanged source, not an explicit recorded activeElement assertion.
- H — NOT RUNTIME-VERIFIED, accepted limitation. useComposerTypingGuard/composerTypingGuard clear held requests on blur or after 1500ms; composer blur calls that guard; held input normalizes activePendingProgress to null. No persistent hold/defer UI is supplied or invented.
- A — SOURCE ONLY, accepted. Unchanged no-active-thread return occurs before all new behavior; terminal UI is thread-bound.

Artifact inspection scope: all 14 indexed full captures, all 5 saved diff images, and all 13 local recordings decoded and inspected as eight ordered frames each (start through end). This includes the two superseded/background recordings beforeQLR and lead-before-bug, which were not substituted for the stronger final evidence. No audio narration exists in these screen recordings; their claimed steps were checked against visual sequence and state observations. I also inspected full-size narrow theme captures and Q/L states for legibility and literal/choice preservation.

Dimensions verified: CSS 1440x900 / 390x844, DPR2. Desktop recordings 2880x1800; narrow recordings 780x1688. Saved desktop PNGs are 1280x800, narrow PNGs 780x1688. Background lead-before-bug is 2560x1600 and is not used as a matched comparison. No native mobile coverage claimed.

Independent rebuild:

- Reused worktree, head, isolated .t3, and ports 6294/14334. Stopped the old runner and exact verified port owners before seeding; no concurrent replacement server. Supplied seed.py was copied outside the repo with only its output directory changed to /tmp/issue-24745-pass2; seeded C only, leaving original evidence files intact. No event-log mutation.
- Own incognito tab; used supplied cache clearing only there, paired privately, real buttons to match layout, supplied terminal printf, answer-before, and independent draft-before fixture. Supplied capture function used real terminal double-click selection and Add to chat; no direct composer handle or toast manipulation.
- Rebuilt image versus supplied afterC: 136/1,024,000 RGB pixels differ at zero tolerance (0.01328125%); 135 at >8/255 (0.01318359375%). Bounding box [529,571,537,588), exclusively terminal cursor. Zero changed pixels outside the original disclosed terminal cursor mask [528,570,538,589), even at zero tolerance. No toast mask needed for this after/after comparison.
- rebuilt-C.png, rebuilt-C.mp4, rebuilt-C-diff.png, rebuilt-state.json, independent-diff.json retained in /tmp/issue-24745-pass2. Diff image is 5x amplified RGB delta for visibility.

Friction / handoff:

- Original session50783 was not accessible through this harness. Port-owner cwd verification: old 723532 -> worktree/apps/web, 723596 -> worktree/apps/server. Runner 723287 was identified through those owners' ancestry and verified cwd; SIGINT ended runner but left children, so exact verified port PIDs were terminated before offline seeding. No pattern kill.
- Restarted same runner: owned session10113, runner PID735395, web735447, server735460. Remains running for lead inspection/cleanup. Private log /tmp/issue-24745-pass2/dev-private.log must not be uploaded (startup pairing material). Final state restored to all questions dismissed, draft-before, zero contexts, hello terminal output, 1440 light. Own tab tab_4_4e75d5a9-0bf1-4234-b878-be4537fd7733 remains available.
- Pair page initially retained stale renderer after server restart; reload restored authentication. No alternate browser automation system used. Initial type/click timeouts did not affect evidence capture.
- Read-only repository audit: no source edits, checkout changes, commits, pushes, PRs, comments, merges, or watches. Only authorized isolated fixture/browser/terminal state changed. No direct live-home database writes. Temporary files retained for lead cleanup; no new provider turn launched.

Original [evidence index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/763baf7c0c13c16a/index.md) and [replay bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a359bf979f41e351/issue-24745-replay.tar.gz).

VERDICT SUFFICIENT