Synthetic Android pending-question comparison.

Base: ed6cb97f07df59c01467b919073550d310da3ef4
Head: da472053493279434989086f0add07dd09a5223b
Pixel 9 / Android 16 API 36 / 1080x2424 / light theme.

All base-25 files belong to a discarded early run whose backend disconnected; they are retained for transparency and are NOT proof.
The valid base run is base-retry-25 (missing controls on reopening 18) followed by base-recorded-25 (missing controls again on reopening 5). The recorded run starts in the previously blank thread and shows four recoveries, then another failure.
Head-first-5 runs the same five scripted reopenings. Additional head runs extend the sample.

Reproduction: prepare each checkout with vp i; use its scripts/scenarios/start-app.ts mobile command with distinct fresh homes, server ports, Metro ports, and the device-reachable host. Install the dev client using mobile:install. Both commits share native fingerprint 17de3e3141895b662dbd84f7b86fc852a561794f. Set agent-device settings animations off before fresh app launch, load that checkout through Metro, pair with its isolated backend. Create the synthetic project and Replay question while on the list, as seed.py demonstrates. Use reopen.py from a thread screen; it returns to list, opens the pending thread, waits three seconds, snapshots and screenshots. background.py creates ten new Replay questions from the list and opens each.

The driver stops when controls disappear or the client displays a reconnect banner. A script interrupted externally is counted only through its completed results.json rows.
No unit tests were used as evidence. All app data is synthetic.

Final counts: base controls missing on 2 of 23 completed scripted reduced-motion reopenings; head controls present on all 50 scripted reduced-motion reopenings and all ten background-question first openings. Original head question answered after about 20 minutes; app and CLI confirmed completion. Normal-motion control: five reopenings per commit, all visible. These finite samples do not establish an absolute failure probability.

Export friction: the initial base normal-motion video export was interrupted by a gateway ECONNRESET and retry reported no active recording. The base normal-motion control was recorded again; base-normal-retry-5 holds that repeated five-opening run.
