fix(ios): stop charging every capture for a slow Simulator app discovery

[thymikee]

Summary

Removes repeated concurrent app discovery on the iOS Simulator AX-bridge route and bounds how long each capture waits for it. Resolver and route regression tests pass; loaded-host smoke validation is what this PR's own CI run will show, and it was still pending when this was written.

The failure it targets

Since #2279 landed (09-05 21:06), the iOS Smoke lane has failed four times with one signature, the first wait text "Agent Device Tester" after a cold open --relaunch reporting wait_capture_stalled with readableCaptures: 0:

• main 80997b6b: https://github.com/callstack/agent-device/actions/runs/33992502696
• main f30328d0: https://github.com/callstack/agent-device/actions/runs/34014253127
• PR fix(ios): confirm alerts without repeating activation #2326: https://github.com/callstack/agent-device/actions/runs/34015120372
• branch fix/2314-memoize-platform-runtime-la: https://github.com/callstack/agent-device/actions/runs/34016148244

The request log of the third run (artifact ios-artifacts, sessions/ios-e2e-smoke-14ax/requests/45ad58af8fd6970d.ndjson) accounts for the 10s budget:

phase	cost
runner findText on the just-launched app	5.8s
AX-bridge target discovery: xcrun simctl spawn … launchctl list, xcrun timed out after 3000ms	3.4s
fallback XCTest snapshot	cancelled at the deadline

The failed-step screenshot and snapshot the harness took right after (#2315) show the home screen rendered with a healthy tree. That is consistent with the timeline, not proof the app was readable throughout the wait.

createSimulatorSnapshotTargetResolver (snapshot-target.ts) ran that spawn with a fixed 3s timeout on every cache miss, on the capture's critical path, and kept nothing on timeout, so the next capture ran its own probe again while xcrun was slow. xcrun takes about 1s on an idle Mac with simulators booted; on these CI hosts it exceeded 3s in both captures of that run, and the same hosts time out simctl terminate at 2s in other artifacts.

Change

• Discovery is single-flight per target and detached from the capture that starts it. A capture waits at most 1.5s for the in-flight discovery, then takes the XCTest fallback with reason simulator-target-discovery-pending. Captures that keep arriving while the probe runs each wait up to that slice; the probe is shared, not repeated.
• A discovery has one 15s deadline shared by both simctl probes and the ps identity read.
• A cancelled caller no longer aborts the probe. A failed discovery is forgotten, so the next capture starts a new one.
• The resolver error names its reason, so the ios_snapshot_route_fallback diagnostic says why the fallback ran.

A responsive host still reaches the bridge on the first capture (local probe ≈ 1s < 1.5s).

What is verified

• Resolver tests (deferred spawn mock, fake timers): slow discovery yields to the fallback and finishes in the background; concurrent captures join one discovery; a cancelled caller leaves it running; one deadline is shared by the probes and the identity read. The first three fail on the old code (hang, second spawn, second spawn).
• Route test with the production resolver over a simctl that answers only when released: the first capture falls back within the slice with the target-resolution-failed warning, the released discovery then serves the bridge without a second spawn.
• Existing resolver and route tests pass (20 total); package typecheck, oxlint, oxfmt, fallow clean.
• Live fixture E2E on a private iOS 26.2 simulator with the rebuilt dist: cold launch, wait for Agent Device Tester 269ms, 23 captures served by the bridge, smoke:automation-input green. This proves the healthy route, not the loaded-host case. (A later keyboard dismiss failure in that run is my local Simulator.app hardware-keyboard environment.)

Not verified here

• A loaded-host wait completing through the fallback: no local way to make xcrun slow; the lane's own runs are the evidence.
• The production poll path end to end under slow discovery; the route test covers one capture, not the wait loop.

Relation to open work

#2329 changes snapshot-route.ts (launch grace on typed bridge failures) and makes wait observe runner-free; it does not touch the probe, and its design needs the probe to be cheap on the critical path. #2325 covers open startup budgets. This PR does not change the wait contract.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.49 MB	4.49 MB	+455 B
Package (unpacked)	4.49 MB	4.49 MB	+455 B
Package (download)	1.33 MB	1.33 MB	+181 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.7 ms	27.8 ms	+0.1 ms
CLI --help	79.6 ms	81.0 ms	+1.4 ms