fix: surface UI content warnings in skill preflight

[onevcat]

Supersedes #139 by @SunsetWan. This branch starts from the #139 head, so the original commits stay unchanged. Merge with a merge commit (not squash or rebase). Then #139 also shows as merged.

From #139

When ui returns ok: true with a remote_content_recovery advisory, preflight now prints a content warning instead of "All checks passed". The exit code stays 0. The warning asks the agent to compare the outline with the screen. On an iOS simulator, it also points at ApplicationAccessibilityEnabled, which an app reads at launch.

Added on top

• Merge the latest main to resolve the CHANGELOG.md conflict.
• Empty outline: a successful read with no elements now gets the same warning. Before this change, preflight reported a clean pass for it. This is the worse case because the remote-content retry found nothing, so no advisory is set. Physical iOS reads are excluded: their entries list is always empty, and the outline text is the payload.
• Skill docs: the guidance moves from the Preflight section of SKILL.md into the Pitfalls symptom index, next to the other [i] advisories, because the advisory can appear on any ui read. The pitfall is renamed to Missing app controls in the outline and covers both cases.
• Tests: use real advisory kinds (orientation_calibration_fallback, full_screen_tap_target) and add empty-outline cases for iOS, Android and physical iOS.

Validation

• swift test --filter PreflightScriptTests passes (6 tests, 20 cases). The new empty-outline cases fail against the fix: surface remote content recovery in skill preflight #139 script.
• make test: all 310 XCTest tests pass. In both Swift Testing runs, the only failure was cancellableSleep wakes early when the flag is cancelled (7.1 s against a 5 s bound under full-suite load). It passes in isolation, and fix: surface remote content recovery in skill preflight #139 reported the same flaky failure. This PR changes no Swift sources.
• Not run on a live simulator in this session.

🤖 Generated with Claude Code

[onevcat]

Review round 1

• R1-P1 (accepted as hardening, fixed in a2f1a71). The empty-outline exclusion keyed only on data.kind == "physical". The current CLI cannot produce an empty-outline false pass: the top-level ui payload sets kind only for physical iOS. executeAndroid() never sets it, and the field docs at IOSSimDescribeUICommand.swift:85-90 say "absent otherwise". Still, the exclusion now names its real reason (platform == "ios" and kind == "physical"). A new regression case (Android, kind: physical, empty entries) failed before the fix and passes after it.

swift test --filter PreflightScriptTests: 6 tests / 21 cases pass.

[onevcat]

Review round 2

• R1-P1: The reviewer confirmed this is resolved.
• R2-P2 (CI red on a2f1a71, not caused by this PR). The only failure was cancellableSleep wakes early when the flag is cancelled (9.6 s against a 5 s bound). This PR does not change ProcessControlTests.swift or ProcessControl.swift, and the test passes in isolation. A re-run of the failed job (run 36812590531, attempt 2) passed, and all checks are now green. We will fix this timing-sensitive test separately. The 100 ms pre-cancel Task.sleep is inside the measured window, so starvation of the cooperative thread pool under full-suite load counts against the bound.

[onevcat]

Review round 3: clean, with no new findings. R1-P1 and R2-P2 are confirmed resolved. make build passes, and all checks on a2f1a71 are green. The PR is ready to merge with a merge commit.