# Proof audit pass 2

No blocking defect found in the required visual evidence. Independently read the complete pass1 matrix, final index, rebuild.md, phases.json, driver snippets, replacement uploads, revision/fixture/diff scripts, and the three-file source diff. HEAD and clean source verified as 9573ae6258f690ee8acd0306f33eb9b9d2fcfe84; base resolves to e3592efe7f5dbd3a869e46e608d8f6815e1bb7f6. No source modifications, commits, pushes, forge writes, server lifecycle actions, SQL or fake DOM/VCS results.

|Row|1440 light|1440 dark|390 light|390 dark|Assessment|
|---|---|---|---|---|---|
|C card + hover|Pass|Pass|Pass|Pass|Stale icon/branch removed; popup branch omitted|
|L slim hover|Pass|Pass|Pass|Pass|Actual popup visible; branch omitted at HEAD|
|S search hover|Pass|Pass|Pass|Pass|Actual popup visible; branch omitted at HEAD|
|G legacy|Pass|Pass|Pass|Pass|Icon/worktree tooltip removed; ordinary title tooltip allowed|
|R lifecycle/reacquire|N/A|Pass|N/A|N/A|Known false → true → false, released/reacquired C|
|V valid checkout|N/A|Pass|N/A|N/A|Recovered C branch and G icon/tooltip remain|
|U initial unknown|Unproven|Unproven|Unproven|Unproven|Control, no claimed visual proof|
|M root mismatch|Unproven|Unproven|Unproven|Unproven|Control, no claimed visual proof|
|X scratch fallback|Conditional|Conditional|Conditional|Conditional|Fixture does not exercise it|

Inspected every final still and both amplified diffs in four full-frame comparison sheets. Independently recomputed all eight pair counts/bounds and all 16 diff pixels: exact match to supplied formula (max RGB channel delta >0 / >20, amplification 8x). All 16 final/stability pairs RGB-identical. 390-dark G bbox is [30,283,520,377]. No unrelated chat/composer/scroll/theme differences in final still pairs.

Decoded every recording independently with ffmpeg, extracting every documented phase plus adjacent hover frames where approximate wall-clock labels preceded the encoded popup. Actual L/S popups appear in all four configurations at both revisions. Replacement HEAD390dark recording has 20m/18m labels and visible L popup. Recovery HEAD encoded frames approximately 20.5s show C restored, 21.5–24.5s show G restored (including tooltip at24.5), later removed/reacquired states clear metadata. Transitional drawer/cross-view ghosted frames were rejected as endpoints; stable later frames and delivered clean stills supply proof. Settled-age text can differ (e.g. baseline dark L4m versus HEAD now), so recordings are interaction evidence, not full-frame pixel comparisons.

Downloaded all 40 authoritative public visual URLs from final index via curl and compared SHA256 to inspected local files:40/40 identical, including replacement recording. Python urllib initially received403; curl succeeded.

Independent rebuild: fresh own native preview tab, existing8407/16447 server, exact retained fixture/route, HEAD source, Linux Chromium154, DPR2, CSS1440x900, actual Light preference, default sidebar width, composerCollapseOnScroll false, scroll0, document fonts ready, frozen Date2026-10-25T22:00:00Z. Waited for actual20m/18m labels; hovered C. PNG1280x800 equals delivered head-1440-light-C-removed.png exactly: pixels>0=0, pixels>20=0, maxdelta0, bbox null. No environment gap. Capture and quantitative result retained in this directory.

Restored Follow system and composer collapse true; default sidebar and active Hello unchanged. Restored page Date and closed own tab. Fixture never mutated during recapture; exact worktree remains absent/unregistered and t3/reply-hello absent. Unrelated fixture worktree untouched. Servers untouched; evidence retained.

Not covered: U,M,X; temporary-null same-key retention and different-key isolation; native mobile, Electron shell, other providers (no changed provider/server/mobile paths). R pass covers lifecycle/reacquire only, not these additional retention controls. Prior59 tests/check:changed reported by lead, not rerun or claimed as independent logic proof here. Core affected set meets matrix minimum; explicit control limitations remain.

FRICTION timing|Wall-clock phase offsets can precede encoded tooltip or follow a short recovered phase|Inspect adjacent encoded frames and choose observed stable states.
FRICTION transport|urllib403 on public artifact download|Use curl and hash downloaded bytes against inspected files.
FRICTION clock|Frozen Date does not instantly refresh existing relative labels|Wait for observed20m/18m before recapture.
VERDICT: SUFFICIENT