# Evidence-only proof audit — pass 1

Target: [#23216](https://github.com/pingdotgg/t3code/pull/23216), GitCafe [#23204](https://git.cafe/pingdotgg/t3code/pulls/23204), branch `fix/issue-16105`.
Reviewed head: [8ac56474a3](https://git.cafe/pingdotgg/t3code/commit/8ac56474a31b840c4b149d58730a0a34cf98e3df).
Actual local merge base against origin/main: [b5a226e49c](https://git.cafe/pingdotgg/t3code/commit/b5a226e49c52ecab4ddc2d678a7012d97e651a6a), matching the canonical before captures. PR metadata also names main and the requested head. Checkout clean and detached at requested head.

## Scope derived independently from the complete merge-base diff

Eight changed files, 333 insertions: three production files select user-only `remark-gfm` singleTilde:false for regular/automation web message bodies and shared user search extraction; two behavior-test files; three projection-fixture/seeder files. No CSS, mobile-native, Electron-shell, provider adapter, contract, transport, or composer changes. Trace: UserMessageRow → regular CollapsibleUserMessageBody / AutomationMessageRow → UserMessageBody → ChatMarkdown. Shared extraction feeds server ThreadFind and its normal/progressive search RPCs. Assistant and proposed-plan processors remain separate. Workflow openings bypass this renderer. Unchanged copy buttons read source text; link/file actions and revert do not acquire new semantics from this parser option, so this audit does not demand an unrelated interaction corpus.

Canonical source: [existing index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6e11b99922250484/index.md), [reproduction bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ab56ed4afdb17972/issue-16105-evidence.tar.gz). Reuse these final-head artifacts. The bundle matrix.md is an earlier planning audit; its concluding INSUFFICIENT verdict predates the supplied recordings, RPC logs and reconstruction. It is not the present evidence verdict.

## Final-diff fixture matrix

| ID | Fixture / expected proof | Required media | Existing coverage / disposition |
|---|---|---|---|
| S1-L1440 | F1 user prose/home paths/parentheses, ~solo~, ~~old~~, bold/link, quote/list/table; F1A unchanged assistant | Before/after + diff0 + diffT, light 1440 | REUSE stills/before-light-1440.png, after-light-1440.png and diffs/light-1440-*; final head |
| S1-D1440 | Same F1/F1A | Before/after + diff0 + diffT, dark 1440 | REUSE dark-1440 equivalents; final head |
| S1-L390 | Same F1/F1A and narrow wrapping | Before/after + diff0 + diffT, light 390 web | REUSE light-390 equivalents; final head |
| S1-D390 | Same F1/F1A and narrow wrapping | Before/after + diff0 + diffT, dark 390 web | REUSE dark-390 equivalents; final head |
| I1 | F2 manual expand/collapse; inline file chips, literal HTML, backtick/tilde fences, literal tail | Before/after recording at dark1440 and light390 | REUSE primary recordings, initial S2 chapters and manual-toggle chapters. Expanded stills supplement; not substitutes |
| I2 | F1 raw ~/folder-a/skills/x and ~solo~ find; old control; active highlight, next/previous, close/reopen | Before/after recording at dark1440 and light390 | REUSE desktop before 7–35s / after 156–200.8s; narrow before 3.5–11.5s / after 8.5–32s |
| I3 | F2 initially collapsed, ~/hidden/a search reveals/highlights hidden tail; manual reverse operation | Before/after recording at dark1440 and light390 | REUSE desktop before 35–52s / after 200.8–218s; narrow before 11.5–15s / after 32–50s |
| I4 | F3 automation initially folded, expand/collapse, body literal tildes, search auto-reveal/highlight | Before/after recording at dark1440 and light390 | REUSE primary chapters; desktop after I3-reveal supplement 9.7–13.54s explicitly supplies folded→revealed result; narrow after 50–66.04s |
| I5 | F4 legacy visible ~/legacy/a searchable; hidden terminal output excluded; F2 context-label exclusion | Before/after desktop recording plus both RPC transcripts | REUSE desktop before 56.5–64.93s and after I4-exclusions supplement 52–80.33s. Empty context record proves label exclusion, not generated chip visual behavior |
| R1 | Normal and progressive server search, all raw-path/single-tilde/hidden-tail/automation/legacy queries and exclusions | Actual authenticated RPC JSON, supplementary to interaction recordings | REUSE rpc/before.json and after.json: both methods, raw queries 0→1; old 9→9; hidden output/terminal label/cross-label query 0→0 |
| C1 | Assistant single-tilde strike and user/assistant double-tilde controls; fenced code literal | Reuse static and recorded controls | COVERED; no separate provider, streaming or plan visual fixture justified by this diff |
| M1 | Exact reported rich-composer input → send → durable rendered user message at final head | Recording | MISSING. Canonical bundle is projection-only; body reports old-head send reproduction, not final-head compose/send recording |
| M2 | A valid UI-generated structured terminal-context record beside literal paired paths | Recording of generation/submission/rendered chip and find exclusion | MISSING disclosed visual/context subcase. F2 records:[] and a context URI are insufficient for actual chip proof. Keep existing extraction proof; do not repeat it |
| M3 | Native Electron running the changed shared web message body | One native recording of send/render plus find/reveal smoke | MISSING native runtime surface; Chrome is sufficient shared-renderer proof but not Electron execution proof. No shell-specific code changed; do not repeat the entire four-pair web matrix |
| P1 | Plain native mobile fixture | Static fixture | ACCEPT maintainer statement: verified with EAS iOS. No mobile-native files changed. Do not manufacture separate Android/iOS required rows; do not claim narrow web proves native mobile or that this PR fixes its separate renderer |

## Exactly what to capture next

1. **M1, required core runtime gap:** at final head record the exact reported prompt in the real rich composer, submit it, hold the sent user bubble showing both literal tildes and no accidental strike, reload/reopen and show durable text. Include double-tilde control and a raw-path find result in the same recording if convenient. One theme/width is enough; existing before reproduction and static four-pair matrix remain reusable. Provider permutations are unnecessary because adapters/protocols did not change. This proves the actual trigger without treating projection seeding as send proof.
2. **M2, disclosed regression-control gap:** use a valid composer-generated terminal-context record, alongside visible paired paths; record submission, actual chip rendering and visible-path/hidden-output or chip-label find behavior. One desktop-width capture is enough. This is not a second claim that exclusion RPC evidence is missing. If maintained as an explicit unverified subcase, it remains a coverage limitation rather than a demonstrated parser defect.
3. **M3, disclosed native surface gap:** one Electron smoke recording using the shared message renderer, literal home paths and find/reveal. Can combine M1/M2 into it and avoid duplicate work. It is a runtime coverage limit, not evidence of an Electron-specific bug.

No missing static web row; no missing existing web expand/find/automation/exclusion recording row. No Android-specific proof requirement, provider-send matrix, event replay, relay/tunnel permutation, settings/palette flow, or scheduler run follows from this diff. All affected interactions listed above require recordings; static screenshots/RPC logs alone do not close M1–M3.

## Audit performed and limits

Read the complete merge-base diff and relevant render/search callers; fetched actual GitCafe PR body and public index/bundle. Verified all 89 bundle checksums. Read revision metadata, fixture, capture recipe, manifest, RPC logs, diff accounting and media chapters. Inspected comparison sheets, representative recorded frames and supplied media-audit frames. Final-head metadata and merge-base align with source. No browser, server, tests or provider turn launched. Sampled frame inspection is not a claim that every video frame was watched. No new app runtime verification was performed.

Existing desktop still output is 1280×889 for a declared 1440×1000 CSS viewport; desktop recordings retain 2880×2000 app pixels. Narrow output is 780×2000 for 390×1000 CSS viewport/DPR2. This encoding is disclosed and is not a missing 1440-CSS-viewport row. Diff0 is raw absolute RGB (display x8); diffT threshold is max channel delta >16. Reflow and translucent-composer effects are documented, not masked.

FRICTION: Existing PR/index metadata still says native mobile unverified; the maintainer now reports EAS iOS plain-static verification. Accept that fact, but its artifact URL/head was not supplied in this audit. Update evidence provenance in a later authorized editing pass; no recapture demanded here.
FRICTION: Bundle matrix.md retains a stale pre-capture INSUFFICIENT verdict. Use this pass and final index to distinguish completed rows from historical plans.
FRICTION: Original desktop recording includes long tool/media-clock delays. Seek using encoded media seconds and supplements, not tool elapsed times; no app-latency conclusion is justified.
FRICTION: No implementation, comments, PR body edits, commits, pushes, watching or polling. All scratch is under /tmp/pr-23216-audit; only this synthetic report is uploaded.

Verdict: REUSE existing final-head web artifacts. Proof remains incomplete only for M1 and explicitly disclosed M2/M3 runtime limitations. This is an evidence audit, not an independent code correctness or merge verdict.