fix: stream replay suite progress

[thymikee]

Summary

Stream bounded replay-suite progress for agent-device test --maestro: suite start, per-file start/pass/fail/skip, attempts/retries, session names, and artifact paths now appear while the suite is still running. Ctrl-C/client disconnect now cancels parent replay work and best-effort aborts active iOS runner sessions so simulator-driving work does not continue detached.

Also fixes two iOS Maestro runtime issues found while validating the React Navigation suite: custom-scheme simulator deep links no longer pre-launch stale app state before simctl openurl, and horizontal percentage swipes avoid the iOS interactive-back edge. Scope: 17 files. Closes #794.

Validation

Focused tests passed: progress/client/session-test/cancellation/Maestro runtime/iOS open tests, 9 files and 230 tests. pnpm format, pnpm check:quick, pnpm build, and node --test test/integration/smoke-*.test.ts passed.

Live iOS simulator validation on iPhone 17 Pro C25DBB5B-9254-4293-A8D5-2785C78DE03A: full React Navigation Maestro suite passed 38 passed, 0 failed in 532.0s with streaming progress and artifact/session paths. Ctrl-C verification exited immediately with code 130, and a follow-up one-file Maestro run on the same simulator passed in 14.9s, showing no active simulator-driving work was left blocking the device.

Local pnpm check:unit wrapper could not run because pnpm 11.1.2 signature verification/fetch is blocked in this environment; the direct underlying unit command also has sandbox/environment failures unrelated to this change (localhost listen EPERM, Swift typecheck, Android helper/upload/materialization suites).

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.3 MB	1.3 MB	+2.8 kB
JS gzip	410.0 kB	411.0 kB	+969 B
npm tarball	543.1 kB	544.0 kB	+953 B
npm unpacked	1.8 MB	1.8 MB	+2.8 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.8 ms	29.5 ms	+1.6 ms
CLI --help	52.8 ms	55.0 ms	+2.2 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/session.js	+1.5 kB	+467 B
dist/src/9542.js	+750 B	+290 B
dist/src/2415.js	+645 B	+226 B
dist/src/apps.js	+29 B	+15 B

[thymikee]

Android validation update:

• Booted AVD Pixel_9_Pro_XL_API_37 as emulator-5554 and installed React Navigation example with ./gradlew installDebug -PreactNativeDevServerPort=8081.
• Configured Metro reverse for tcp:8081 and tcp:8082.
• Smoke Android Maestro subset passed with streaming progress: bottom-tabs.yml + stack-basic.yml, 2 passed in 64.6s.
• Full Android Maestro suite streamed bounded progress throughout and completed 36 passed, 2 failed in 1185.1s.
• The two failures reproduce in isolation, so they appear Android-specific/pre-existing rather than caused by this PR's progress/cancellation changes:
  • drawer-master-detail.yml: launch gate fails after retries with uiautomator dump did not return XML while waiting for Pages.
  • screen-layout.yml: Suspend remains visible after tapping it; final assertNotVisible Suspend fails.

Artifacts:

• Full run: /private/tmp/agent-device-794-android-full/1b9ab21f1d75300e
• Isolated failure rerun: /private/tmp/agent-device-794-android-failures/0a0aee53a77a6fe9

[thymikee]

CI update: all checks passed except Fallow Code Quality. Please address before merge readiness.\n\nFallow findings:\n- src/daemon/handlers/session-test.ts: runReplayTestSuite at line 31 is now 94 LOC, cyclomatic 24, cognitive 9.\n- src/cli-test-progress.ts: formatReplayTestCaseProgressEvent at line 23 is cyclomatic 10, cognitive 13, CRAP 31.6.\n\nKeep the live iOS/Android validation evidence in the PR. Fix should be targeted: extract focused helpers where it improves locality, keep progress output bounded/agent-friendly, preserve JSON behavior, and rerun pnpm check:fallow --base origin/main plus the focused tests already used. Also make sure the cancellation path remains covered and does not regress the Ctrl-C live behavior.

[thymikee]

Addressed the review blockers in dd3f6698b.

What changed:

• Fallow findings fixed by splitting runReplayTestSuite setup into focused helpers and reducing formatReplayTestCaseProgressEvent branching.
• HTTP disconnect cancellation now arms only after auth succeeds.
• The iOS runner abort loop now follows request cancellation state, not res.writableFinished, so streaming disconnects stop aborting as soon as handler cleanup clears the request.
• Disconnect-triggered iOS runner abort is scoped to explicit iOS requests or replay/test requests, so Android-only RPC disconnects do not trigger global iOS abort.
• Unknown future progress event types are ignored again instead of failing older clients.
• Replay test retry loop now stops when the parent request is canceled.
• Removed the no-op announceReplayTestRun path and renamed the misleading HTTP NDJSON progress test.

Regression timing/root cause:

• The streaming-progress silence/cancel behavior regressed when replay suites moved to long-running daemon-side Maestro execution without per-test progress and without complete parent-request cancellation propagation. This PR fixes that path.
• The Android helper slowdown observed while validating was separate: persistent Android snapshot-helper sessions were introduced in d087a62d1 / PR fix: improve Maestro Android reliability and snapshot speed #612 on 2026-05-29 with a session socket budget of timeoutMs + 2s. Busy RN accessibility captures could exceed that, causing a premature session timeout and slow one-shot fallback. That is handled separately in PR fix: relax Android snapshot helper session timeout #796.

Validation:

• pnpm exec vitest run src/__tests__/cli-test-progress.test.ts src/daemon/handlers/__tests__/session-test-suite.test.ts src/utils/__tests__/daemon-client.test.ts passed: 43 tests.
• pnpm exec vitest run test/integration/provider-scenarios/daemon-http-server.test.ts passed: 7 tests.
• pnpm check:quick passed.
• pnpm check:fallow --base origin/main passed.
• node --test test/integration/smoke-*.test.ts passed: 8 tests.

Note: direct ./node_modules/.bin/vitest run --project unit is not a clean signal in this sandbox; unrelated tests fail on loopback bind EPERM / Swift typecheck / daemon entrypoint environment. The impacted focused tests above pass.