Proof audit pass 2 for [#24568](https://git.cafe/pingdotgg/t3code/pull/24568).

Findings: no blocking proof defects or contradicted claims in the supplied replacement Evidence.md. E5 documents an actual cold-hover regression; sufficient proof does not mean that regression is fixed or that the PR is approved for merge. Synthetic Azure and the inert provider are labelled accurately. Live Azure remains outside scope.

Reviewed AFTER [e13265440a](https://github.com/pingdotgg/t3code/commit/e13265440a44b51116f9c3427b50c3d0e2c3ca1e) against BEFORE / merge-base [3cc31ed27c](https://github.com/pingdotgg/t3code/commit/3cc31ed27c8956b36cee35a50509fb7c6869592b). Local checkout, forge metadata, table columns, runtime identities and all six production-source hashes agree. Read the complete five-file diff and relevant hover/reactor callers.

| Row | Status | Evidence checked |
| --- | --- | --- |
| E1 | SUFFICIENT | Replacement failure-only pair shows exactly one AFTER summary and active watches. Original ordered runtime logs show scheduled pending/failure/duplicate sweeps; committed event and inert-provider receipt carry Tests (#12), head A. Independently reproduced AFTER exactly. |
| E2 | SUFFICIENT | Paired still shows failure plus required pass. Runtime recording shows required-pending silence, required pass with optional pending, optional completion, and duplicate suppression. Paging is skip 0 then 100, with 100 non-build rows on the first page. |
| E3 | SUFFICIENT | Replacement pair has six AFTER summaries, active watches and identical composers. Ordered logs prove rerun and A-to-C transitions; committed C-pass text matches the actual provider receipt. |
| E4 | SUFFICIENT | Pair shows retained failure. Real completed malformed-page sweep has typed error/warning; persisted watch fields and message IDs remain equal through recovery. Separate cold endpoint recording proves preview success and strict detail failure. |
| E5 | SUFFICIENT, regression captured | BEFORE metadata card below link; AFTER URL tooltip above link. Both full recordings, all 23 encoded frames and the narrated sequence inspected. Cold PR91 subprocess records show BEFORE show/iterations/changes versus AFTER show/denied policy. |
| H1 | SUFFICIENT at reused harness scope | Decoder, filtering, names, status, pagination and error evidence retained as labelled harness results. All 58 bundle hashes verify. Raw runtime metadata confirms Node v26.12.1 on both revisions. |
| H2 | SUFFICIENT at reused harness scope | Labelled empty/all-filtered, logical +16min, nullable and iteration-fallback evidence agrees with raw logs. No actual 16-minute runtime or new device claim. |

Visually inspected all 38 delivered PNGs: selected before/after captures, both diffs per row, settlement frames, two video sheets, and the superseded E3-final images in the old archive. Also inspected both videos through every encoded frame. The old E3-final pair includes the former composer mismatch and inactive watches; Evidence.md explicitly replaces it with E3-active-C-pass, whose diffs contain neither problem. The old pair was not counted as current proof.

All five selected pairs are 1280x900 dark, as authorized. No stale overlays, duplicate elements, loading placeholders, accidental clipping, missing rendering, or unrelated changed regions were found in selected settled images. Intentional author-label ellipsis in the existing hover card is unchanged rendering. What-else-changed descriptions agree with the images and source.

Recomputed every selected diff. All ten images match their declared formulas exactly: Diff0 is absolute per-channel delta amplified 32x, DiffT16 is per-channel delta >16/255. Selected pair AE counts are E1 2171, E2 2992, E3 6243, E4 2184 and E5 53362. Every delivered settlement frame is pixel-equal to its selected frame.

Independent reproduction used the supplied setup, az fixture, RPC driver, browser preparation and native capture procedure. Only scratch-root paths and the unused baseline side were changed in the copied setup. Fresh home, T3CODE_DEV_SEED=0, explicit --home-dir, --no-seed; no migrate-dev-db or developer database. Started the real dev server from the exact AFTER checkout, seeded running policies, observed a native scheduled sweep, changed policy12 to rejected and awaited the next native sweep and terminal inert-provider run. The persisted event carries the expected failure text. The browser was prepared using the supplied Date/status-pulse/sidebar/details controls; no inserted UI content.

Matching rendering environment: fs-29; Linux 6.1.102 x86_64; Node 26.12.1; Vite Plus 1.3.0; T3 preview Chromium 156; 1280x900 CSS/PNG; DPR2; dark; root16px; visual scale1; no device emulation; fonts loaded; sidebar collapsed; no scroll or popup; browser Date 2026-10-27T16:00:00.000Z; server clock live; status-pulse paused at time0. Resource idle exceeded 173 seconds before capture.

Rebuilt E1-failure-after.png: settlement AE0, delivered AE0, threshold16 AE0, max channel delta0. The PNG files are byte-identical, SHA256 b59b539e00fa769dc930883b63064a6d0b3e4d6e5c2e73a62cd6863e5f419f9d. Capture and comparison metadata are retained under run-20261027T151452Z/captures/E1-failure-20261027T151955Z/.

![Independently rebuilt E1 failure AFTER, pixel-identical to delivered evidence](run-20261027T151452Z/captures/E1-failure-20261027T151955Z/E1-failure-rebuilt-after.png)

Cleanup verified: fixture watch unregistered, owned preview tab closed, observer exited, captured server process group and provider PIDs 2844196/2844906/2844859 absent. No preview tabs remain. Worktree porcelain status empty and HEAD unchanged. 29 GB free. No edits to source, pushes, forge comments or PR changes. Scratch evidence and raw logs retained, not deleted. Completed approximately eight minutes into the 60-minute budget; no coverage was truncated.

FRICTION:
infra | One browser-driver.js download exhausted retries on DNS resolution; a later direct retry succeeded | Reliable DNS resolution.
model | Initial orchestration/browser calls rejected invalid escaping, about:blank, or trailing whitespace; corrected before capture | Raw shell strings, omit the URL for a blank tab, and trim expressions.

VERDICT: SUFFICIENT