iOS runner smoke: alert-observation XCTests are red or flaky on main #2546

Purpose

Smoke Tests in the iOS workflow is not deterministic on main. Two alert-observation XCTests fail on the runner host without any Apple source changing, so every PR that reaches that step has to be re-run before a reviewer can trust it.

Observed on main

• main@973b74cc14 (push run 34741577232, job 103682270067, step Run targeted iOS runner XCTest regressions), RunnerTests+AlertObservationTests.swift:72:

  XCTAssertEqual failed: ("false") is not equal to ("true")
  - Optional(… code: "ALERT_DEADLINE_EXCEEDED", message: "alert dismiss exhausted its deadline")
  

  Failing test: testAlertDismissDoesNotActivateAReplacementWithTheSameTitle.

• PR docs: fix phantom specifiers, the duplicate ADR 0019, and the Node floor #2533 head a16821b661 (run 34742468875 attempt 1, docs-only diff), same file at :75 and :80:

  XCTAssertEqual failed: ("First actions: 0; replacement actions: 0") is not equal to ("First actions: 1; replacement actions: 0")
  

  Failing test: testAlertCannotProveAnIdenticalReplacementAndDoesNotActivateIt. Attempt 2 of the identical run and commit passed at 06:46 UTC, so the second case is a flake rather than a regression.

Both symptoms share one shape: a presented alert reports no actions, and the assertion times out through ALERT_DEADLINE_EXCEEDED.

Separate symptom on the same step family, worth tracking apart from the two above: main@8e8eeb2ced (run 34712516058) failed Run fixture-backed iOS simulator E2E smoke with an assertion raised at test/integration/ios-simulator-e2e/live-automation-scenario.ts:33 (observeFixtureHome).

Required behavior

• The alert-observation assertions in apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/UnitTests/RunnerTests+AlertObservationTests.swift pass or fail on their own terms, not on simulator alert presentation timing.
• A presented system alert that the test can see must expose its button list to the probe, or the probe must report the empty-button case as a distinct typed outcome instead of consuming the whole deadline.
• Whatever the resolution, an unchanged main must not require a manual re-run to go green.

Observable completion conditions

• A diagnosis recorded here naming the mechanism (alert not yet realized when the action list is read, deadline sized below host latency, or a genuine runner-side ordering bug).
• If it is timing: the probe or deadline change lands and the iOS workflow's Smoke Tests job is green across at least five consecutive main pushes, with no manual re-runs.
• If it is a real defect: a failing-then-passing XCTest that pins it, plus the same five-run streak.
• The observeFixtureHome case is either fixed or split into its own issue with the above evidence linked.

Dependencies

• Reproduce on a hosted macOS runner or a local simulator; the failure does not appear in unit or check:affected runs, so this is device-facing and docs/agents/device-verification.md applies.
• Related history for the alert horizon: ADR-tracked alert deadline work behind testAlertResolutionCannotBypassRequestedDeadline and testAlertDeadlineBeforeActivationLeavesTheOriginalUntouched, which pass in the same runs where the two above fail.

Discovered while babysitting #2533, whose diff touched no Apple source.