Evidence recapture 4 for [#31721](https://git.cafe/pingdotgg/t3code/pull/31721).

AI-assisted; verified with the tests and tools listed above.

Revisions: [c52df24d](https://git.cafe/pingdotgg/t3code/commit/c52df24d285a072ad9c3f7ab5b3656c0be1cb0f5) → [3f0217a1](https://git.cafe/pingdotgg/t3code/commit/3f0217a16f13bff6800e7fff9b7b1d4159254218).

Create before4/after4 worktrees at those commits and run `vp install` in each. Parameterize the absolute scratch root, checkout paths and actual Vite ports in the scripts. Run `prepare-fixtures.py`, make `fixtures/op` executable, then start each `start-servers.sh <side>` with `setsid`; retain process-group PIDs. Every dev runner uses `T3CODE_DEV_SEED=0` and `T3CODE_FEATURE_SECRETS=1`; no production userdata, copied database or migration fixture is used. Runtime credentials and startup logs are private and omitted from this package.

Use T3 preview tools. Call preview_status then preview_open with separate default/incognito profiles for after/before, because localhost cookies collide across ports. Authenticate each tab using that server's own pairing URL; never upload it. Set store keys `beforeTab`, `afterTab`, `tab`, `proofRoot`, `beforePort`, `afterPort`. Evaluate `capture-helpers.js(tools,store,load,text)` to obtain h. Before each side set `tab` to its owned tab and instantiate h again. The real run used 6099 before, 8182 after.

The app creates one starter thread even without dev seeding. Remove it on each fresh environment through the authenticated app command before capture: import appAtomRegistry from `/src/rpc/atomRegistry.ts`, primaryEnvironmentIdAtom from `/src/state/primaryEnvironment.ts`, threadEnvironment from `/src/state/threads.ts`; read environmentId, read threadEnvironment.snapshotAtom(environmentId), then await threadEnvironment.delete.run(registry,{environmentId,input:{threadId}}) for each starter thread. No agent turn is sent. Both captures then have no thread age/status chrome.

Run `scenario-recapture.js(h,side,row,viewport)` for F02/F03/F06/F20/F21/F25 at DL and F09 at DL/NL. F02 covers closed plus popup. F03 saves one account and reopens Edit. F06 shows blank-email. F09 includes top/middle/bottom and last selection: scroll the popup's `[data-slot=scroll-area-viewport]`, not its content list. Observed after scrollTop: DL 0/160/200; NL 0/160/280. Before has no picker and uses equivalent manual account entry. Keep responsive sidebar state matched: in this run the final F09 DL pair navigated at NL then resized to DL on both sides; all other DL pairs share the same startup viewport history. Do not compare differently collapsed sidebars.

F20 `delayed` is 9000ms. Record before opening Add, then type `custom-team-pending` one character every 650ms plus tool latency, without clearing. The selected after recording shows the first character at about 5s, no trigger at 10.5s with `custom-t`, and trigger plus popup at 11s with `custom-te`; typing finishes about 19s. Focus remains Account across all observations. The baseline performs the same manual typing and has no account-list query. F21 uses the real timeout profile (60000ms op process, 45000ms driver bound) and continuous recording beyond completion; a read-only query subscription observes Initial → Success([]). The baseline has no account-list query. Its recording keeps the manual value focused for the comparable wait; a failed attempt to observe a nonexistent baseline query was corrected by continuing the recording, without starting any query.

Every still awaits fonts and 500ms resource quiet, disables animations/transitions/carets identically, and matches another original screenshot at least 500ms later in RGB. Date is frozen after app boot to 2026-11-06T00:00:00Z. No data was created after that freeze except synthetic provider settings; thread-age chrome was removed through the app. Captures remain full surface. `diff.py` produces saturated-magenta masks: every nonzero maximum RGB-channel delta for Diff0, and >16/255 for DiffT. `validate-recapture.py` validates all 36 selected source-frame pairs and measures pixels outside the exact dialog/popup rectangles. For all requested matched-background rows (F02/F03/F06/F20/F21/F25), both outside counts are zero. F09 DL has 8352 nonzero pixels outside the exact popup rectangle from its shadow, all ≤16/255; after padding the dialog/popup rectangles by 20 raster pixels there are zero residual pixels. No background chrome differs.

The first F09 after attempt scrolled the content element and produced unchanged positions. Those files were kept locally under rejected-scroll-target and are excluded from uploads. The final scrolling pair uses the viewport element and was visually reviewed. No application source was edited. Run diff.py, validate-recapture.py, video-sheets.py and targeted-video-review.py; review all captures/diffs and recording contact sheets before upload. The targeted reviewer decodes F20 through the response boundary. F21 was inspected every two seconds through both full recordings. This is not a claim to have watched every encoded video frame.

Upload selected captures, diffs, sheets, recordings and public audit files with `ghdrop upload -m`; include this script/fixture package and corrected matrix. Stop only retained process groups, close both owned tabs, and remove only before4/after4 worktrees. Nothing is committed.
