test(ios): stop cold-host timing from failing the occupancy test

[thymikee]

Summary

Fixes testAbandonedTreeCaptureSkipsQuerySweepAndHonorsWarmupExemption, which runs first in "Run targeted iOS runner XCTest regressions" and failed 8 times from 09-15 to 09-23 on cold or loaded hosts, once on main. Test-only; 1 file, +9/−1.

• Line 106 (4 failures): a cold viewport read overran its 1 s cap before the tree capture the test targets, so the recorded reason was "preparing tree snapshot". The test now takes that read uncapped before the timed section. The production cap is unchanged.
• Line 111 (2 failures): the blocking stub's own 20 s timeout released the main thread before the test read the abandonment count (Optional(0)). It is now a named 75 s leak guard, well past the plan deadline, so only the test's own release ends the block.

Dropped after review:

• A class-level warm-up. RunnerTests.setUp also runs before the production runner's testCommand.
• The bare-type relaxation. refactor(ios-runner): restrict the tap-witness synthesized type route to the bare submit key #2845 moved bare type off the synthesized route whose dropped keystrokes caused that flake, so the relaxation was obsolete. It was also broken: failure responses carry no data.

Not fixed here: 4 cold app.launch() timeouts.

Validation

• ba936608c6, rebased on main: pnpm check:affected --run passed.
• Unit-test build on a fresh iPhone 17 Pro / iOS 26.2 simulator: the occupancy test passed 5 of 5 iterations.
• Mutation: pinning the stub timeout to the 8 s slice budget reproduces Optional(0).

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.79 MB	4.79 MB	0 B
Package (unpacked)	4.79 MB	4.79 MB	0 B
Package (download)	1.43 MB	1.43 MB	+8 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	22.4 ms	22.5 ms	+0.1 ms
CLI --help	62.1 ms	61.1 ms	-1.0 ms

[thymikee]

This does not fix the bare-type half of the flake, at commit 41249a0. executeTypeCommand builds a failure response as Response(ok:false, error:ErrorPayload(...)) with data == nil, and only sets textEntryRoute on the ok:true DataPayload (https://github.com/callstack/agent-device/blob/41249a0/apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/RunnerTests+CommandExecution.swift). The test still asserts XCTAssertEqual(typeResponse.data?.textEntryRoute, "synthesized-first-responder") unconditionally (line 408), so when the runner returns TEXT_INPUT_COMMIT_NOT_OBSERVED, data is nil, the comparison is nil-vs-expected, and the test still goes red on a cold host, now with a route-mismatch message instead of the real commit-not-observed cause. The new prefix and code checks never gate the outcome, so a host that drops keystrokes still fails this test.

Can the route assertion move inside the if typeResponse.ok branch, and can the not-ok branch assert the route from evidence the failure path actually carries — either the AGENT_DEVICE_RUNNER_TEXT_ENTRY_ROUTE log line, or textEntryRoute attached to the failure response's data (the failure path's textResult already carries it through dispatchFailureResult)? A unit case that drives the commit-not-observed response through this assertion would prove the fix.

The rest of the diff (occupancy stub timeout) reads sound, so this is the one blocking change before merge. All ten checks on this commit are still queued or in progress, none have failed yet, and the diff only touches iOS runner XCTest source, so those checks are unlikely to speak to this issue either way.

[thymikee]

Addressed at ba93660. The bare-type change is gone: #2845 moved bare type to xctest-element, so the synthesized-route keystroke drop no longer applies to this test. Main's version is kept as is. Also applied a thermo-nuclear quality pass: the occupancy leak guard is now a named constant, and the narration comments are cut. The PR is now 1 file, +9/−1.

[thymikee]

The unconditional textEntryRoute assertion on a nil-data failure response from the earlier review (41249a0) is gone at ba93660; this PR now only touches RunnerTests+SnapshotCapturePlanOccupancyTests.swift and is test-only, so it does not carry that risk.

18/18 checks are green on ba93660, and the diff touches only that occupancy test file, which is exactly what the previously failing 'Run targeted iOS runner XCTest regressions' job exercises for this test name, so the green run is direct evidence the fix runs. A single green run does not rule out the cold-host flake recurring, so a repeat run or two on that job would help confirm the fix holds.

I did not run the runner XCTest suite myself; this reads on the diff plus your stated validation (5/5 pass on iPhone 17 Pro/iOS 26.2, and a mutation test pinning the stub to the 8s slice budget reproduces Optional(0)) and the 18/18 green checks. I also didn't re-run pnpm check:affected --run independently.

This looks ready to merge from the code side.