# Announcement lifetime evidence

Surface for every base/head video: Linux web app served by its own checkout's `T3CODE_DEV_SEED=0 vp run dev`, Chromium 157, 1280 x 900 CSS viewport, device scale factor 2, dark appearance. Raw videos are 2560 x 1800; short clips are scaled to 1280 x 900. Fresh independent databases; synthetic project and threads only. No live database was read or copied, no paid agent turn was started, and no application source was changed.

## Reproduce

1. Install dependencies with `vp i` separately in each checkout; start each with `T3CODE_DEV_SEED=0 vp run dev`. Pair each browser tab with its corresponding server privately. Never include pairing URLs in evidence.
2. The fresh app registers the checkout's server directory as a default project. From each checkout run `node apps/server/src/bin.ts project rename --base-dir .t3 "$PWD/apps/server" "Announcement demo"`.
3. Evaluate `seed-expression.js` in each paired app. This uses the existing client command and real server RPC to create stopped, empty threads named Expiry example and Undo example. The Replay selection is a synthetic fixture; it never starts a turn. The app's initial New thread draft is also present.
4. Click the Undo example sidebar row in both tabs, open the thread details panel, ensure Open in Files is present, and reload. Set both tabs to the viewport/theme above.
5. Run `capture-driver.js` with T3 Browser tools and each tabId. Supply the text of `browser-instrumentation.js` and `capture-schedule.js` as its instrument/schedule arguments. The function returns the recording path, actual DOM samples, and accessibility excerpts. Save each result under a new timestamped run directory. The capture tools require expressions without leading/trailing whitespace. Large observations are retrieved in batches to fit their response size limit.
6. Copy each raw recording to base-raw.mp4 or head-raw.mp4. Put both in one directory and run `python3 make-clips.py <directory>` to recreate all six short clips. No annotation or result panel is composited into any video.
7. Evaluate `restore-native-region-css.js` and inspect the browser accessibility tree once more. This restores the original nodes to clipped 1 x 1 presentation; base still exposes the used Undo instruction, head does not. See native-accessibility.json.

## Exact timeline (seconds on the recorded diagnostic clock)

- 1, 11, 21: click the real Open in Files button. These are three identical general announcements, each separated by ten seconds, with no undo notice active.
- 31: click Settle on Expiry example. Wait through the five-second Undo window and observe until 39 seconds.
- 40: click Settle on Undo example. At 42, click the real Ctrl+Z to undo button, while it is active. Observe the restored sidebar row and announcement text through 47 seconds.

The driver invokes existing controls' click handlers. It never imports or calls announce(), cancelAnnouncement(), a notice store, or a replacement implementation. It never advances or replaces application timers.

## Inspection instrumentation

The two visible boxes are the ORIGINAL `body > [aria-live="polite"][aria-atomic="true"]` nodes, moved and enlarged with diagnostic CSS. They are not copies of their text. A and B mean stable original DOM order, not the module's changing internal tuple order. Empty means the node's real textContent is empty. The diagnostic headings and elapsed clock are aria-hidden. The composer and machine-context row are masked with CSS before recording to keep operational metadata private. All other data is synthetic.

Observations sample the real nodes and the real Undo button every 100 ms. Undo presence is determined from its actual text, layout, visibility, and absence of inert/aria-hidden ancestors; it is never assigned from a sample number or time. Accessibility excerpts are measured separately with the browser tool. The final native accessibility inspection checks that the styling did not manufacture the base's accessibility-tree exposure.

## Observed

- Base: the general text remains at 9.4, 19.4, and 29.4 seconds. Each subsequent action replaces it; it does not clear in between.
- Head: the general text first appears near 1.3 seconds and clears near 8.3; the next two clear near 18.3 and 28.3. This demonstrates the seven-second timer without cancellation from an Undo window.
- Both: identical general announcements land in original nodes A, B, A. Head continues alternating after both nodes have become empty between announcements.
- Base: after Undo expires near 36.1, its control is absent but its text remains. After Undo is clicked at 42 and the thread is restored, the text also remains.
- Head: the instruction and control both disappear near 36.1 on expiry and near 42.1 on Undo use. All times are measured at 100 ms sampling resolution.

Inspection images in the bundle are frames/contact sheets extracted from videos for quality review, not separate still captures. base-attempt-raw.mp4 is a superseded exploratory run with an incorrect Undo-button selector; no matched proof relies on that attempt. The final base/head recordings and observations use the corrected shared driver.

Not exercised: NVDA/JAWS/VoiceOver speech or virtual cursor; desktop/mobile clients; the other individual callers of announce (model picker, boot-error/copy); mixed-source announcement batches. This evidence directly proves web DOM/accessibility-text lifetime and node alternation.

AI-assisted; verified with the tests and tools listed above. No tests are used as evidence here.
