# Read-only proof audit, pass 3

Final-head smoke: NOT VERIFIED. Checkout is clean at [e8058c32](https://git.cafe/pingdotgg/t3code/commit/e8058c32bac198a471de72f678e570efc0ba38e5). Initial listener inspection saw web8018/server16058; the processes disappeared before /proc cwd inspection and the HTTP request returned connection refused (HTTP000). A subsequent listener check was empty. No runtime was started, stopped, reconfigured or driven by this audit. Existing runtime presence is not a passing smoke.

## Confirmed evidence defects first

1. **Minor, nonblocking comparison confounds:** some pairs differ in panel-button hover/highlight or horizontal-scrollbar visibility. These are visible in the full-frame diff maps and disclosed by the supplement. Counts include those pixels, surrounding reflow and, for narrow PR, underlying chat. They are not target-link-only change metrics and do not demonstrate a wrapping regression. No masking or registration was used.
2. **Revision qualification:** all supplemental endpoints predate the merge head. They compare [d4d5ad61](https://git.cafe/pingdotgg/t3code/commit/d4d5ad61be36dc32ef9e1934d64b3bd6cd4fe54e) with [eea98953](https://git.cafe/pingdotgg/t3code/commit/eea989536849e5b5049987ee24318b5a77326542), not the final merged runtime. The merge parents are exactly d4d5ad61 and [3cc31ed2](https://git.cafe/pingdotgg/t3code/commit/3cc31ed27c8956b36cee35a50509fb7c6869592b). Against the latter, the cumulative diff remains one file/four added lines. No changes between original candidate and final head were found in ChatMarkdown, chat/files/pullRequest component directories or index.css. This strongly supports carrying forward the renderer evidence; it does not manufacture a final-head runtime smoke.

No false passing claim, corrupted media, incorrect diff calculation, or confirmed code regression was found. The labeled synthetic PR is valid renderer evidence, not a defect or a claim of successful live retrieval.

## Audit method and evidence

Read the [original matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1f399b349c05c692/issue-24147-pass1-matrix.md), prior local pass2 audit, [prior index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9d94cc8314b1fd1e/issue24147-index.md), and [supplement](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/86ac7aa5be58eb47/issue24147-supplement-index.md). The public supplement matches the local index byte for byte. Earlier 19-pair/38-diff independent rebuild results are retained, not repeated or newly claimed as pass3 execution.

Visually inspected every supplemental pair, repeat and Diff0/DiffT map in all ten labeled sheets: attachment1440/390, file1440/390, PR1440/390, assistant-token390, user-token390, user-leading1440 and user-shapes1440. Enlarged narrow candidate originals for Files, attachment and PR. Independently recomputed all ten pixel-count pairs, all20 diff maps exactly, and all20 endpoint/repeat comparisons (zero delta). Diff0 is absolute RGB difference amplified8/clipped255; DiffT uses strict max-channel >16. Downloaded and SHA256-verified all68 public PNG/MP4 media against the manifest and local files.

Inspected all eight raw recordings through independently extracted 2fps sequences spanning their full durations and enlarged selected interaction/endpoint strips. Durations: base/candidate file6.713/6.586s, attachment4.688/4.661s, PR4.924/4.853s, token-pan7.920/7.942s. The supplied 20-frame sheets therefore cover these clips without duration truncation. This is sampled visual inspection, not a frame-by-frame playback/flicker measurement. Inspection strips are navigation aids, not new screenshot proof.

## What the supplement establishes

- **S4 N/O and C10/P5 material gaps resolved for the samples.** Both narrow chat roles show the long token; the user header tail is initially clipped on both revisions and is reachable during panning. Real Files rendered markdown and real unsent attachment previews cover no-chat-lane layouts at both widths. Tables already scroll horizontally on base and candidate; prose/list/heading retain wrapping. No new whole-markdown horizontal overflow is established. The wide competitive user table reallocates width toward the token and grows taller; that is an observed global effect, not a newly confirmed unacceptable regression.
- **C11/P6 actual production renderer gap resolved, with an integration limit.** The actual PR summary component/plugin chain renders explicitly labeled synthetic cached detail. Explicit versus generated reference/mention/commit anchors are visible. Captured DOM metrics report generated anchors bypassing helper wbr insertion on both revisions, while explicit labels lose wbr nodes. This corroborates routing; it is not live PR, comments or media-segment proof. The activity-read failure remains visible.
- **Interaction proof added.** Files: pan right/left, collapse/expand, source/rendered return. Attachment: pan right/left and source/rendered return. PR: pan right/left and collapse/expand. Both chat roles: token-table panning and return. Asynchronous source loading is visible; endpoints return to rendered mode. No provider stream was invoked.

## Row statuses

“Pass sampled” is bounded to the retained fixtures; “partial” is not a code failure.

| Rows | Pass3 status | Remaining qualification |
|---|---|---|
| S1 | Pass core, retained | Both roles/themes/W/N, th/td; prior independent rebuild retained |
| S2 | Partial | ASCII samples pass; Unicode remains tofu/unverified glyph proof |
| S3 | Pass sampled | Native word breaks; added user W and actual file prose/list; not every narrow role permutation |
| S4 | Pass sampled; material gap closed | W/N/O now covered; wide token column grows, narrow tables pan; no promise of fit without scroll |
| S5 | Pass sampled | Hyphen/underscore/hash behavior retained; added actual no-lane text |
| S6 | Pass sampled control, retained | Suffix dot/slash rule remains; no change claimed |
| S7 | Partial control | Explicit URLs covered; GFM autolink control absent |
| S8 | Pass sampled | Added user W leading dot/slash versus suffix-dot control |
| S9 | Partial control | Bold/code retained; image+text not verified |
| S10 | Partial control | Globe sampled; loaded favicon/brand transition not verified |
| S11 | Pass cached production-renderer sample | Generated reference/mention/commit controls; live/comments/media not verified |
| S12 | Partial control | Fenced control retained; full bypass-control suite absent |
| S13 | Optional, not run | No provider turn required |
| C1 | Core pass, composite partial | Added narrow token proof |
| C2 | Core pass, composite partial | Added user W shapes/leading and N token; not all variants in every role |
| C3 | Exercised, retained | Expanded work-entry body |
| C4 | Unverified distinct affected wrapper | Actual inline compact reasoning/thoughtLine with truncate/line-clamp and thought-preview plugin not demonstrated |
| C5 | Pass sampled, retained | Plan preview/full/preview |
| C6 | Representative equivalence only | C3 covers inset typography risk; actual event detail not exercised |
| C7 | Unverified | Actual small italic inspector not exercised; named limitation per original matrix |
| C8 | Inventory-only | Pending-question caller not exercised; original matrix permits inventory-only |
| C9 | Unverified | Actual review/context comment not exercised; named limitation per original matrix |
| C10 | Pass sampled; material gap closed | Actual Files and unsent attachment, both widths/O; source/rendered returns |
| C11 | Pass cached renderer sample | Actual production PR body; live fetch/comments/media remain unverified |
| C12 | Unaffected | Fenced workflow script; no separate capture needed |
| P1 | Core pass, composite partial | Improved user shapes; no exhaustive Cartesian claim |
| P2 | Pass primary, retained | Light primary both roles |
| P3 | Pass sampled | Narrow primary retained plus long-token panning both roles |
| P4 | Improved, partial | Actual heading/list/paragraph and narrow inset added; missing unaffected controls are not fixes |
| P5 | Pass sampled; gap closed | Actual no-lane Files/attachment |
| P6 | Pass renderer sample | Cached production PR; live integration qualification retained |
| P7 | Partial; distinct C4 requirement still open | Expanded body retained; compact reasoning absent; C7/C9 also unverified |
| R1 | Pass sampled, retained/extended | Prior primary round trips plus new panel interactions; no streaming claim |
| R2 | Pass sampled with prior limit | 576/575px rather than exact35.99rem; prior scaled video qualification retained |
| R3 | Plan portion passes; reasoning portion unverified | No actual compact reasoning reverse round trip |

## Verdict rationale

The pass2 material S4 N/O, C10/P5 and actual C11/P6 renderer gaps are resolved for the named samples. The primary short-label fix remains convincingly demonstrated, with no confirmed regression. Do not request code changes or additional unaffected-control captures for S7/S9/S10/S12 merely to fill rows. Live forge retrieval is a transparent integration limitation for this text-layout change, not a prerequisite to accept the cached renderer proof. C6/C8 do not require redundant captures under the original equivalence/inventory rules; C7/C9 remain honestly unverified.

However, the original matrix explicitly singled out C4 compact inline reasoning as a distinct wrapper requiring a short/long-link sample. Source confirms it renders ChatMarkdown inside truncate or line-clamp with a thought-preview plugin; block/table/no-lane proof does not demonstrate its clipping behavior. Therefore the original proof matrix cannot yet be closed: C4/P7 and the reasoning portion of R3 remain the precise outstanding proof obligation. This is missing evidence, not a confirmed defect or an instruction to edit source. Final-head runtime smoke is separately unverified.

FRICTION: Web reader rejected public index URLs; curl worked. Concurrent listeners vanished before HTTP smoke; no server restarted and no browser opened. Local inspection used only retained public/synthetic artifacts and read-only source/git/runtime checks. No source edits, checkout switches, app-state writes, provider calls, forge commands, push, comment, merge or watch. Report/inspection aids only under /tmp; report uploaded using ghdrop. Reduced-motion media emulation, native mobile/Electron/remote modes, live streaming and live PR integration remain unverified. Audit by GPT-6.1-sol through Codex in T3 Code.
VERDICT: INSUFFICIENT to close the original proof matrix: C4/P7 and R3 reasoning remain unverified; S4 N/O, C10/P5 and C11/P6 renderer gaps are resolved. No confirmed regression.