# Proof audit, pass 2

No blocking proof findings. The delivered evidence is sufficient for all eight affected rows in the supplied matrix. The four unchanged rows retain their explicit waivers.

Reviewed [#28056](https://git.cafe/pingdotgg/t3code/pull/28056) at [d3e3620310](https://github.com/pingdotgg/t3code/commit/d3e3620310e8489e1cee5776e79d64ad7173bc63), against [083c903950](https://github.com/pingdotgg/t3code/commit/083c9039505fc9207b88343e19d0a3de4a5e622d). The head checkout is clean, and its merge-base with origin/main is the requested baseline. The local main branch is older; the explicit baseline was used throughout.

The capturer's G3/G4 scope note concerns the title, not an undisclosed change. The PR body already says that excerpts of added or deleted files get no expansion loader. G3/G4 show precisely that: the compact added/removed line, counts, caption and Open in Code control remain, expansion controls disappear, and requests stay at zero. G3 stops expanding unrelated added lines; G4 also avoids the baseline inline null-line renderer error. These changes are within the PR's described scope.

| Row | Status | Observed proof |
| --- | --- | --- |
| D1 | SUFFICIENT | Real app; five before/after pairs. Matching unified/split expansion, long-line scroll/wrap interaction, ignored whitespace without final newline, and immutable commit after a dirty edit. All selected pairs have zero pixel delta. |
| D2 | SUFFICIENT | Real app; fourteen pairs. Truncation, append, same-length context edit, and branch-range repeat; stale rejection and refresh/re-expansion. Additional 1440/390 light/dark pairs have correct dimensions/themes. Baseline error-boundary comparisons and recovery steps are explicitly labelled. |
| P1 | SUFFICIENT | Labelled production PullRequestCodeTab harness, actual shared loader, controlled query/RPC output, commit:null; expanded unified and split views. Two zero-delta pairs. |
| P2 | SUFFICIENT | Labelled production GuideStop/GuideFileDiff harness. Recording shows other-file disclosure and expansion, then excerpt expansion. Two zero-delta pairs. The other file moves below the viewport after the excerpt grows; the preceding state and continuous recording show its expansion. |
| G1 | SUFFICIENT | Four zero-delta pairs: modified-source new/old excerpts, addition-only and deletion-only snippets; surrounding lines render, old side says removed code, caption present/absent variants. Later length-changing source hunk is included. |
| G2 | SUFFICIENT | Stale source outside the excerpt is rejected by the actual loader; compact patch persists. Source/cache refresh restores expansion. Recovery differs by twelve pixels at one RGB level, zero above 16. |
| G3 | SUFFICIENT | Genuine wholly added source; initial and attempted-expansion pairs, zero requests, code/header retained. Independently rebuilt both after stills with zero pixel delta. |
| G4 | SUFFICIENT | Genuine wholly deleted source; initial and attempted-expansion pairs, zero requests, removed-code header and compact deletion retained. Baseline inline renderer error disappears. |
| U1 | WAIVED | Unchanged multi-hunk excerpt guard, as scoped by the supplied matrix. |
| U2 | WAIVED | Unchanged loading branch. |
| U3 | WAIVED | Unchanged missing-range fallback. |
| U4 | WAIVED | Unchanged annotations/commenting behavior. |

Inspected all 141 delivered PNGs: 66 captures, 66 diff images, eight labelled sheets and the D2 overview. Each of the 33 pair sets was opened in an unscaled 2x2 view; native detail was checked where fine text needed it. No selected still contains ghosted/duplicated UI, a stale overlay, loading placeholders, missing code, or a mismatched viewport/theme. Normal horizontal/vertical scrolling is visible, not hidden rendering failure.

Opened and visually inspected all 34 delivered recording files through decoded timelines (primary MP4 at one frame per second, raw/supplemental WebM at one frame per two seconds), together with native interaction stills and transcripts. Each published narrated step agrees with the observed sequence. Startup and baseline crash-reload frames contain disclosed loading transitions; those are not used as settled proof. Timestamp notes occur after actions/waits as disclosed. Open in Code is a callback check in the harness, not proof of hosted navigation.

Independently recomputed both amplified diffs for all 33 pairs. All 66 images and manifest counts match exactly: maximum RGB channel delta >0 in magenta, >16/255 in orange, on the stated faded after image. The changed areas agree with the code and the evidence's complete what-else-changed list; no unrelated visual change was found. Expected stale-content rejections are distinguished from baseline renderer failures. The pre-existing pierre-dark worker error is disclosed, so no zero-console-error claim is inferred.

## Independent rebuild

Downloaded replay.md, seed.py, capture-app.mjs, capture-harness.mjs, harness entries/HTML and the driver archive. Extracted the archive into its own new directory after checking paths and rejecting links. Ran the published seed and G3 harness driver at the requested head. The driver's only change was ROOT pointing to this step's scratch subtree; production sources were untouched. Started dev with T3CODE_DEV_SEED=0 and an explicit scratch home; real ports from dev-runner were web 5760 / server 13800. No live userdata was used.

Rendering environment matched: fs-11, Ubuntu Linux x86_64, Playwright headless Chromium 148.0.7778.96, no device emulation, DPR 1, 100% zoom/font scale, 1440x1000 CSS px, light theme, reduced motion, Date frozen at 2026-11-03T12:00:00Z.

| Reproduced still | Pixels >0 | Pixels >16 | Maximum channel delta |
| --- | ---: | ---: | ---: |
| G3-initial.png vs delivered G3-after-G3-initial.png | 0 | 0 | 0 |
| G3-attempted.png vs delivered G3-after-G3-attempted.png | 0 | 0 | 0 |

The no-loader G3 harness does not need live RPC authentication. This independent run used an empty auth storage state and logged an ambient WebSocket 401 plus the known theme warning; both are retained in its transcript. The settled images are pixel-identical. It made zero contents requests and invoked the Open in Code callback.

Hosted authentication, guide generation, persistence and routing remain unverified, as the evidence and matrix explicitly permit for harness rows. Native mobile, Electron IPC, relay and tunnel have no changed branch here and were not separately exercised. No broader merge-readiness claim or rerun of code tests is made by this proof audit.

## Preservation and cleanup

The reproduction archive contains the independent stills/video/transcript, driver copies, replay stdout, sanitized dev-server log, per-pair validation and pixel comparison. The inspection archive preserves derived visual inspection material. Original delivered artifacts were already public and remain at the supplied links. Private auth files and startup credentials are excluded.

The dev process started for this audit was stopped through its captured session (exit 143); neither 5760 nor 13800 has a listener. The browser tab opened by this step was closed. No baseline worktree was needed or created. No source edits, commits, pushes, comments, PR-body edits, provider turns, emulators or leases occurred. Remove only this step's scratch subtree after uploading; preserve the lead checkout and evidence-capture directory. Budget was not exhausted.

FRICTION:
- tool | Python urllib received HTTP 403 for artifact downloads; curl downloaded every file successfully | Use curl for this artifact host.
- tool | Preview status initially unavailable and explicit about:blank rejected as an invalid preview URL; opening with the URL omitted worked | Use an omitted URL for a blank shared preview tab.
- tool | The requested DejaVu contact-sheet font was absent; installed Liberation was used | Select an installed font.
- model | Initial montage rendering obscured fine text and caused two premature suspicions; native inspection disproved them and the commentary was corrected | Inspect native detail before classifying fine-text defects.

VERDICT: SUFFICIENT