# Proof audit pass 2

## Findings

1. **Full proof is insufficient.** Required rows 03, 08–15 and 17–21 have no executed coverage. Rows 02, 04–07 and 16 are partial. Shared dependency changes affect citation/fullscreen/restore pins, nested lists and six generic list callers, so the central reproduction cannot discharge those obligations.
2. **The central issue proof is sound for the demonstrated fixture and Electron environment.** Final-head four-variant held native drag, 90 downward wheel ticks, >800ms pause/resume, release and real clipboard copy preserve the test12 anchor and contiguous independently enumerated fixture prefix. Wide copy contains 137 markers through test80; narrow contains 103 through test63. All four baseline variants lose the original anchor and fail prefix checks. Exact complete Reply80 selection is not required for this narrower claim and is not established.
3. **Unchanged visual proof is insufficient.** Final unselected base/head comparisons exceed tolerance; wide composer focus/compact geometry, workspace action/change totals, minimap, scrollbar and hover/time controls differ. Rapid-scroll comparisons terminate on different rows. These are rejected as matched unchanged passes, rather than excused as rasterization noise.
4. **Cleanup is substantially improved, not exact.** Actual text-bearing rows fall from 146 wide / 110 narrow while selected to 27 / 19 after collapse and far-end scrolling, compared with fresh unselected controls of 18 in all variants. Bounded excess +9/+1 remains. This proves release of most selected-span retention, not exact resource recovery or all row04 transitions.
5. **No additional confirmed source bug asserted by this proof audit.** Missing thread lifecycle, rich endpoints, nested tool selection and other pin consumers remain material untested risks. Contact sheets show transient blank content during very rapid scrolling in controls/cleanup; settled captures recover. No quantitative stall/frame-time claim is warranted.

## Provenance and scope

Read-only source: /home/ubuntu/wt/issue-12943. Base [909e4c9482](https://git.cafe/pingdotgg/t3code/commit/909e4c9482), final head [fb7048ced5](https://git.cafe/pingdotgg/t3code/commit/fb7048ced5436e66373eb132a3ae66453b8b8eba). HEAD independently read and checkout clean after audit. The base is a capturer-documented byte-for-byte relevant-source restoration with old installed dependency patch, not an independently rebuilt base checkout by this reviewer. Superseded head captures do not prove final head.

[Final evidence index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f31b31c3fe7e904c/evidence-index.md). [Matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/583cf2b90ef8c54a/matrix.md). Synthetic 100-turn/200-message SQL and capture scripts intentionally live outside the PR; this follows the current hygiene instruction.

Electron 44.9.3, display :103, primary-owned existing runner 1042382/session 43378, isolated .t3, ports 6243/14283. WL/WD=1440×900; NL/ND=390×844; DPR1. Light/dark, loaded fonts, capture-only reduced-motion CSS, no masks. Native mobile and browser portability are excluded, not proven.

Source diff reviewed: timeline inclusive selection-span pins; both react.js/react.mjs gain selection-aware next-frame ordering/restoration, immediate alwaysRender recalculation and off-buffer unpin removal; lock patch hash updates. Existing native entry points are unchanged. No transport/provider/backend behavior claim.

## All affected rows

| Row | Final-head status | Remaining obligation |
|---|---|---|
|01|Core demonstrated in WL/WD/NL/ND|Unchanged still comparison remains unmatched; no exact Reply80 claim|
|02|Partial; older WL/ND backward/reversal diagnostics|Final-head R2, all S4, direction/contraction checks|
|03|Unverified|Tall rich markdown and nested block endpoints|
|04|Partial final four-variant release/collapse/far-end counts|removeAllRanges, new small span, return; matched near-control proof; +9/+1 excess explicit|
|05|Partial older keyboard attempts|Final cross-row keyboard reversal and select-all|
|06|Partial older inside-to-header case|Outside-to-list, composer/input selection, final head|
|07|Partial older short Cite/pointer diagnostics|Final Tab, invalid multi-message, over-limit, correct composer/clear/pin release|
|08|Unverified|Existing offscreen citation, fold/page targeting|
|09|Unverified|Thread A/B/A, draft/loading/held lifecycle and pending callback|
|10|Unverified|Turn/attempt fold, endpoint removal, prepend|
|11|Unverified|Streaming/finalization/follow/working states|
|12|Unverified|Nested tools inner/outer selection, collapse/restore/append|
|13|Unverified|Work log wrappers, disclosures, subagent paths|
|14|Unverified|Remaining row-content/status/attachment branches|
|15|Unverified|Embeds, fullscreen pins and surrounding host selection|
|16|Partial four-variant rapid scroll|Minimap/find jumps, controlled resize; unmatched end positions|
|17|Unverified|Model picker|
|18|Unverified|Branch picker|
|19|Unverified|Font picker|
|20|Unverified|Command palette|
|21|Unverified|Synthetic thread PR panel|

Rows22–24 are matrix scope exclusions/shared checkpoints, not extra passes. There are five named generic pickers/panels in rows17–21 plus the nested work list in row12 among the non-timeline runtime callers.

## Visual inspection coverage

Opened 183 local final r3/base stills and raw/amplified base→r3 diff images through 23 labeled contact sheets (inspection-manifest.json identifies each). This includes final selected/unselected/repeat/cleanup/control and matched-reshoot stills, and baseline supplemental diagnostics. Final accepted stills show correct viewport/theme, coherent selection highlights and no obvious ghosted/mid-transition text. Black repeat differences agree with stability metrics. Boot/devtools/incomplete-history diagnostic screenshots are not accepted flow proof.

Opened eight sampled frames spanning each of 25 recordings: four base core, four final core, three playable base cleanup/scroll, four final cleanup/scroll, four final unselected controls, two base backward/keyboard/header and four base Cite diagnostics. Baseline WL cleanup clip is excluded as unplayable. These are contact-sheet/keyframe inspections, not frame-by-frame playback; no audio listening claimed. Narrated versions carry visible step captions; SRT inspection confirms held pause/resume/release/copy narration. Spoken/raw duplicate encodes and superseded head recordings were not separately opened; therefore the literal request to inspect every delivered historical/duplicate artifact is **not fully satisfied**. They add no final-head coverage. No fabricated all-artifact pass.

## Independent rebuild

Rebuilt ND in the same real app at final head using native xdotool pointer down on test12, 90 wheel clicks with pointer motion, an 800ms held pause after tick45, release, Ctrl+C and xclip. Instrumentation only reads selection/mounted rows; positioning/theme setup is programmatic. No synthesized selection or repair.

Anchor selection-user-12 offset0 remained at pause/end; final focus selection-assistant-63 offset0; selection text5988 chars; clipboard5780 bytes, fixture prefix true, final markers Reply62/test63. Two settled captures600ms apart are identical.

| Comparison | Diff0 | Pixels exceeding channel tolerance8 | Percent |
|---|---:|---:|---:|
|Rebuilt selected vs delivered r3-ND selected|39830|8637|2.623952%|
|Rebuilt unselected vs r3-matched-ND|41356|12939|3.930915%|
|Rebuilt selected repeat|0|0|0%|

No masks; raw=masked. Selected delta bounds x12..389/y525..827; inspected difference is composer compact/expanded geometry, scroll-to-end button position and lower hover controls, with selected text above unchanged. Rebuild is behaviorally consistent but **not pixel-identical** to delivered evidence. The first attempts were rejected because DevTools/blank preview obscured the app; final capture has neither in its390×844 crop. No rebuild pixel acceptance is claimed above tolerance.

Delivered unselected comparison metrics: WL Diff0 90701 / >8 32097 (2.47662%); WD76467 /14380 (1.10957%); NL2188 /1604 (0.48730%); ND2226 /1648 (0.50067%). Reshoots remain above tolerance. Final r3 repeat pairs are identical; stability alone does not establish before/after correctness.

## What else changed / FRICTION

No checkout edit, push, forge comment, merge, server launch or primary-app stop. No public forge writes. Review scratch/evidence only under /tmp. The built-in preview probe created a blank floating preview; closed through normal preview lifecycle before accepted rebuild. An initial clipboard paste landed in the synthetic composer; cleared without sending. DevTools required clicking the app before its shortcut; failed/obscured attempts rejected. Final capture-only CSS was added for matching conditions and remains in the app DOM; polling cleared. An attempted later CSS cleanup could not run after polling removal; no claim it was removed. Primary was notified rebuild finished and UI may stop. Bridges this reviewer started (1055662 and1061793) were stopped by captured PIDs. Primary app left running.

Core repro supported; full matrix and literal exhaustive duplicate/historical-artifact inspection remain insufficient. This audit does not authorize merge.

VERDICT: INSUFFICIENT 02–21