MOB-114: pin the two behaviours that arrived for free

[GenericJam]

Final step of epic MOB-108. Follows #95, #96, #97, #99, #101.

What this step turned out to be

MOB-114 was scoped as three pieces of migration work. Two were already done — they fell out of MOB-112 rather than needing separate work. Each live screen schedules its own state sync and dumps in its own terminate/2, and hot reload has been a broadcast over every live screen since screens became processes.

Neither had a test. That's the real gap: behaviour that arrived as a side effect of another change, asserted nowhere. Reviews of MOB-112 flagged hot_reload as having zero coverage, and nothing had ever round-tripped state across a multi-stack app — this issue's stated acceptance criterion.

Both are pinned now, verified as negative controls. Before MOB-112 the second wasn't merely untested but impossible: only the active screen held a socket, so a screen you'd navigated away from couldn't be dumped at all.

The third piece was attempted and reverted

That's the more useful result, so it's recorded in the ADR rather than quietly dropped.

Mob.Test's tap/2, select/3, send_message/2 were changed to resolve the screen pid and send directly, reasoning that a native tap reaches a screen via Mob.Listener without touching the router. Review showed that was wrong three ways:

• It made the thing it was avoiding worse. get_screen_pid/1 is a GenServer.call into the router — the serialisation point the change cited as its motivation. One async send became a synchronous round-trip plus a second send. Measured at 5002 ms for one send_message while the router was blocked, against a documented fire-and-forget contract.
• Resolve-then-send isn't atomic. The router reads state.current.pid and delivers in one step. Resolving separately opens a full RPC round trip in which the screen can restart and the event be delivered to a corpse.
• The premise was factually wrong. Alert actions and webview messages go to :mob_screen on both platforms and are documented send_message/2 payloads. Router-handled shapes stopped working entirely — send_message(node, {:mob, :back}) no longer popped.

Addressing :mob_screen is correct, not just incumbent: the router resolves and delivers atomically in one non-blocking RPC, and it's where native itself sends the messages this function simulates. lib/ is unchanged by this step.

Two barriers that looked like barriers

Worth recording — each produced a passing test that proved nothing.

Mob.Sender.sync/1 doesn't order a hot reload: a background screen's tree is dropped rather than committed, so the sender catching up says nothing about whether that screen processed its cast. Nor does :sys.get_state(router), since hot_reload/1 is a cast per screen. The screens are drained — and because render/1 is user code that may take arbitrarily long, the assertion is a bounded wait rather than a snapshot comparison. With a 30 ms sleep in render/1 the snapshot version failed every run; the bounded version passes.

Separately, screens dump in their own terminate/2, which runs after the router exits, so GenServer.stop(router) returning doesn't mean the writes landed. The state tests monitor the screens and wait for their exits.

Verification

• 1243 tests, 3 new. 12/12 clean full-suite runs
• Negative controls: reverting the hot-reload broadcast, or removing the dump from terminate/2, fails exactly the tests written for them
• Timing probe: a 30 ms render/1 passes 3/3 with the bounded wait
• mix format, credo --strict clean

Epic complete

MOB-108's six steps are done. Deferred findings are tracked as MOB-115, MOB-116 (resolved by MOB-112), MOB-117, MOB-121; unrelated bugs found along the way as MOB-118 (deploy durability), MOB-119 (suite flakes), MOB-120 (Android dark-mode text).