Native iOS proof: comment after refresh

Both sides ran the real, unmodified T3 server and mobile JavaScript from their respective commits in separate checkouts and fresh T3 homes. The same cached native development client fingerprint was used on one iPhone 18 Pro simulator (iOS 27.2), portrait, 402 x 874 logical pixels, light theme. No agent turn was needed for this Issues flow.

GitHub transport was synthetic. github-fixture.py serves GraphQL and comment POST responses over a Unix socket named github.sock beside the script. It supplies 105 synthetic comments, holds an older-page response for 240 seconds, answers issue detail promptly, and records accepted comments. Configure an isolated gh config.yml with http_unix_socket pointing to that socket, and give the app GH_CONFIG_DIR plus a synthetic GH_TOKEN. No live GitHub issue was modified.

seed.mjs takes checkout and fresh T3 home arguments, creates the synthetic repository/project through the scenario API helper, and supplies its synthetic GitHub remote. Each app was started with scripts/scenarios/start-app.ts mobile, a fresh home, reachable server and Metro ports, T3CODE_DEV_SEED=0, and the isolated GitHub configuration. Install the client through mobile:install, load Metro, and pair with the test server.

Open PRs, switch its header menu to Issues, then select the synthetic issue. Its Activity and Load older comments control fit on the initial screen. rapid-flow.sh takes base or head and drives the same device session: Load older comments, a slow downward pull to refresh, capture after refresh, focus the composer, type Comment after refresh, and tap Post comment. It saves fixture state immediately after refresh and posting. Keep the whole interaction below the server's 30-second GitHub request timeout.

The base recording shows the draft remaining after the enabled Post tap, without an error. The head recording shows the composer clearing and Activity increasing from 105 to 106. runtime-evidence.json confirms the older response was pending at both taps; the base accepted zero comments and head accepted the draft. Final PNGs are frames from these recordings. The proof recordings end shortly after the result, before the older request times out.

The scenario runner was attempted first, but three native navigation/locator corrections and slow accessibility captures prevented it from reaching the assertion. These recordings came from direct device control, not from a passing scenario or test.
