# Thread message recovery proof

Web: Chromium, 1280×800, dark. Native: iPad Pro 13-inch (M5), iOS 27.2, dark.
Each app checkout runs a real server on fresh isolated state. A seeded Replay (dev)
command-output turn supplies the same message history on both commits.

The injected fault models an undecodable replay. It does not reproduce the original
Interrupted/ClientAbort cause. In cursor mode, only subscriptions with afterSequence
fail with Die; cursorless snapshots work. In all mode, both replay and snapshot fail.

The web scenario instruments WebSocket responses in the browser. The native fixture
patch adds the same controlled fault at the server subscribeThread transport boundary.
It changes no client code. Apply the fixture patch to the current-main scenario harness.

Run the web scenario with scenarios:prove and the requested base/head commits.
Run the native scenario once per app checkout with T3_SCENARIO_APP_DIR pointing at it,
--target ios --action-timeout 120000 and separate --output directories. Set
T3_PROOF_FAULT_FILE to an owned temporary file, and include it plus the pool environment
variables in T3_SCENARIO_INHERIT_ENV. Both native runs use exactly the same scenario.
The pool's protocol-2 gateway requires agent-device 0.21.36; the harness used the copy
already installed for @e2e-dev/mobile instead of current main's 0.21.41 client.

Recordings are trimmed to start at the loaded-history capture, after pairing, and end
after the recovered-state capture. Pairing and connection details are excluded.

The full native three-stage scenario was run on iPhone 18 Pro. The shorter native-label
scenario was run on iPad Pro 13-inch (M5) to capture the complete subtitle without phone
header truncation; it fails replay and snapshot together, then recovers on foreground.
