# Proof audit pass 2

Findings first: the delivered evidence does not satisfy the specified complete paired visual matrix. This is an evidence verdict, not a new source-code objection.

1. Before V1 is only 1440 light. Before 1440 dark, 390 light, and 390 dark end-to-end recordings are absent. Supplemental failure screenshots cannot replace interactions; narrow failure details are below the viewport. After V1 supplemental recordings re-submit an existing alias/deduplicate and do not show reconnect. Fresh registration and reconnect are supplied only by R1 at 1440 light.
2. V2 requires blank selection and a second typed-override selection in every cell. Before provides typed keyboard in three cells and blank pointer in 390 dark. After provides username-only/port-only/both/blank respectively. These show representative behavior, not the required dual-case coverage. Different ports and typed input combinations make the paired result states noncomparable for pixel equivalence. The blank Before 390-dark selection also uses the short discovered alias rather than the stipulated long alias.
3. Strict settlement fails nine of 24 After a/b pairs. Independently recomputed 15 passes. All four V1 result pairs differ as success toast disappears; caret differences affect comparable 1440-dark, V2 fields 1440-light/390-dark, V1 fields 1440-dark/390-light. Before countdown and narrow caret differences are also real failures. No threshold erases these failures.
4. Published two-diff sets are correct but restricted to query/suggestion crops. They exclude typed fields, changed labels and almost all screen context. Background route counts, provider state, viewport scroll position, and translucent dialog context differ. They cannot certify full dialog, label, or full-frame regression coverage. R1 backend 15477 differs from baseline/visual backend 15474.
5. Exact source checkout confirmed, but loaded-build SHA attribution remains the supplied provenance, particularly the Before bundle. Git HEAD and asset hashes alone do not independently bind a running renderer to a source SHA. R1 recording has no included record-start UTC marker; its transcript is timestamped but is not a fully synchronized action-by-action narration of its entire 82-second video.

## Row status

| Row | Disposition | Actual evidence and limits |
|---|---|---|
| R1 typed, two port changes | Affirmative runtime proof | Before retry video connects then fails at stale 24569 while terminal alias resolves 24570. After primary video visibly registers, switches off/on twice, and returns Connected; transcript records effective 24570→24571→24572, same alias/alias/ubuntu/null, stable connectionId/environment and profileCount1. Effective descriptor success is not raw argv interception or direct credential association inspection. |
| R2 discovered blank | Affirmative within disclosed IPC scope | Video visibly selects discovered alias with blank overrides, Adding keeps blanks, then route added. Transcript records alias/alias/null/null, stable IDs and count2 at 24570 then24571. Preferred active route stays primary: off/on UI alone cannot establish discovered-route reconnect. Direct ensure/descriptor supplies that evidence. The discovery recording ends before changed-port reconnect. |
| R3 username-only | Affirmative | 1440-light selection visibly preserves ubuntu with empty port; conflicting nonexistent configured user makes success meaningful. Runtime transcript saves ubuntu/null and returns ubuntu/24571 again after defaults change and explicit IPC disconnect. |
| R3 port-only | Affirmative | 1440-dark selection preserves24571 with blank user; configured24579 then24578 are different. Saved null/24571; runtime descriptor succeeds ubuntu/24571. Configured username remains ubuntu, so no independent valid-account User-change claim. |
| R3 both | Affirmative | 390-light shows both ubuntu/24571 before selection and still populated during Adding; saved tuple and reconnect descriptor honor both despite conflicting configured user/port changing again. |
| V1 1440 light | Partial paired coverage | Before registration/reconnect failure; After R1 registration/two reconnects plus supplemental dedup video. Different backend/port/state, not a controlled pixel pair. |
| V1 1440 dark | Incomplete | Before interaction missing; After dedup video present, no reconnect; result pair unsettled. |
| V1 390 light | Incomplete | Before interaction missing; After dedup video present, no reconnect; fields caret and result toast unsettled. |
| V1 390 dark | Incomplete | Before interaction missing; After dedup video present, no reconnect; result toast unsettled. Original mislabeled Before failure is genuinely1440 wide and correctly excluded. |
| V2 1440 light | Partial | Before both overrides24570, After username-only: keyboard path visible; missing dual-case cell coverage. |
| V2 1440 dark | Partial | Before both overrides24570, After port-only24571: keyboard path visible; missing dual-case cell coverage. |
| V2 390 light | Partial | Both sides typed overrides but ports24570/24571 and route counts differ; missing blank case. Long suggestion shortcut/address clips at right edge; result secondary alias ellipsizes. |
| V2 390 dark | Partial | Both blank pointer cases visible; Before autofills ubuntu/24570 during Adding, After retains blanks. Before short alias / After long alias; missing typed case. |
| Zero/threshold diffs | Correct, narrow scope | Four cells each have before/after crops, amplified8× zero-delta image and binary max-channel >16/255 mask. Independently recomputed counts below. |
| Provenance | Partly independently verified | Checkout HEAD78d05b9e0c0315303ba1985742e284a6d734f762, base d53873b6a4799e87afd63c8d288ef0089f5a3748; clean checkout. Declared CSS1440/390×900, DPR1 light/dark matches media. Running-bundle source attribution is supplied, not independently rebuilt in this audit. |
| Focused tests | Lead-reported only | 25/25 and374 focused/scoped checks were supplied in brief. No rerun and no independent test-result certification in this proof-only pass. |

## Independent pixel calculation

| Cell | changed >0 | changed >16 | pixels |
|---|---:|---:|---:|
|1440 light|97|22|56004|
|1440 dark|2514|20|56004|
|390 light|841|0|28044|
|390 dark|5456|0|28044|

All match delivered pixel-analysis.json. Black narrow threshold masks mean no pixels over16, not zero difference. Low-contrast structure in zero-diff images is consistent with differing translucent background context. Wide visible caret lines account for the prominent above-threshold marks.

## Visual inspection performed

Opened every PNG in both evidence directories:99 total, including all a/b captures, all eight diff images, before/after crops, supplied timelines and the explicitly invalid/mislabeled images. Contact sheets were generated in /tmp/issue7568-pass2/static-00.jpg through static-16.jpg. Inspected every WebM by decoding frames every3seconds and opening all generated timestamped contact sheets. This covers all7 Before recordings (including unsuccessful V1 and mixed-theme exploratory V2) and all10 After recordings (R1,R2,four V1,four V2). Opened key screenshots at native size to read baseline stale-port error, narrow long alias fields and result labels. Contact sheet timestamps mark the centers of each3-second sample interval; this is sampled visual inspection, not a claim to inspect every encoded frame or prove absence of subsecond artifacts.

No sampled source frame is blank or shows an unexplained ghosted duplicate. Blur behind open dialogs is the actual modal background. Black unused tiles in supplied timelines are contact-sheet padding, not blank source frames. Before stale resolved labels are expected bug reproduction. Initial Before V1 spends a long interval in an uncompleted form and is correctly not successful registration proof. The exploratory Before v2-1440-dark recording visibly changes width/theme and is correctly excluded as a single-cell pass. Narrow captions/subtitles truncate; Before error details go below the fold; long picker suggestion metadata/shortcut clips at right. These limit readability claims, not newly diagnosed regressions.

Recordings are silent. Capture scripts/logs provide useful UTC action narration for V1/V2, including keyboard vs pointer and failure recovery; they do not turn the R2 discovery video into a recording of its later reconnect. R3 runtime changes are transcript/IPC proof, not recorded UI reconnects.

## Rebuild one capture

Used existing CDP19470 with agent-browser session issue7568-final, after browser release. No app navigation or mutation was necessary: final isolated app already occupied the delivered V1 390-dark result state. Fresh screenshot /tmp/issue7568-pass2/rebuilt-v1-result-390-dark.png is390×900 and RGB pixel-identical to /tmp/issue7568-after-evidence/v1-result-390-dark-b.png (difference bbox None). Snapshot independently exposes SSH ubuntu@issue-7568-repro · Connected. This confirms reproducibility of this affected result capture; it does not replay registration/reconnect or repair missing matrix coverage.

All122 entries in inventory.json match their local SHA256. Only audit artifacts were written under /tmp. No checkout/source/catalog/SSH config edits, app starts/stops, forge actions or tests. Parent runtime left unchanged.

FRICTION: CLI was not directly on PATH; used its documented npx invocation after reading core/Electron skill content. Missing baseline cells, representative V2 variants, unstable caret/toast/countdown captures, limited crop comparisons and supplied loaded-build provenance remain explicit. No gaps waived, no new code-review round requested.

VERDICT: INSUFFICIENT — missing three Before V1 recording cells; missing required V2 blank+typed interactions in every paired cell; unsettled required capture pairs; no comparable label/dialog-wide diff evidence; loaded-build SHA provenance and R1 recording narration not independently closed.