ci(android): un-flake the emulator integration-test job

[ddfreiling]

Problem

The API 24 emulator job fails about one run in five, and it is not a boot failure — the emulator boots fine. reactivecircus/android-emulator-runner presses the unlock key the instant sys.boot_completed reads 1, with no hook in between:

await waitForDevice(port, emulatorBootTimeout);
await adb(port, `shell input keyevent 82`);

A restored quickboot snapshot sets that property before its services are back, so the keypress is killed (exit 137) and the step dies before our script: runs — which is also why the failed runs upload no diagnostics. Upstream issue #489 is open; the fix (#495) is unmerged, so bumping the action does not help.

Signature, identical across every affected run:

INFO | Successfully loaded snapshot 'default_boot' using 3563 ms
Emulator booted.
[command] .../adb -s emulator-5554 shell input keyevent 82
##[error]The process '.../adb' failed with exit code 137

Survey of the last 60 CI runs:

Job	Failures	Concluded	Of those, this bug
API 24	9	42	6
API 36	1	48	0

Fix

A sleep cannot be inserted where it is needed — the action exposes only pre-emulator-launch-script (before launch) and script (after the keypress). So, two guards instead:

• Snapshot a settled system. The AVD-snapshot step now idles (script: sleep 60) before teardown, so the cached snapshot captures a system that has finished starting rather than one mid-startup. Only runs on a cache miss.
• Retry, narrowly. Run integration tests is continue-on-error, and retries once only when android-test-api<N>.log is absent. A missing log proves the suite never started; a real test failure always leaves its log behind and is never re-run.

Both attempts call a new shared composite action, .github/actions/run-android-integration-tests, so the two paths cannot drift. The AVD cache key now hashes that file too.

Verification

• actionlint clean; both YAML files parse.
• The flutter test one-liner moved byte-identical (diffed), apart from matrix. → inputs..
• Editing these files changes the AVD cache key, so the settled snapshots regenerate on the first run of this branch.

Not verified: this has not run in CI yet. The retry logic is sound on paper; the settle sleep is well-reasoned but unproven — the failure is intermittent, so a single green run will not confirm it either. Worth watching the API 24 job over the next couple of weeks.

No CHANGELOG.md entry: CI-only, nothing a consumer upgrading the package would notice.

🤖 Generated with Claude Code