Android: snapshot helper exits without parseable output after a screen transition (still on 0.21.33) #4005

Summary

On an Android emulator in CI, snapshot intermittently fails because the snapshot helper exits without returning parseable output. The client gets COMMAND_FAILED with details.androidSnapshotHelperFailureReason: "Android snapshot helper failed before returning parseable output". It reproduces on 0.21.28 and 0.21.33, so the vanished-child fix in #3901 (0.21.31) did not remove it for us.

Environment

• Emulator: system-images;android-36;google_atd;x86_64 (API 36, Google ATD), in CI
• Driven through @e2e-dev/mobile 0.12.4 (persistent snapshot-helper session, then the one-shot am instrument fallback)
• App: React Native (Expo)

Frequency: an A/B over the same suite gave 1 failing run of 5 on 0.21.33 and 1 of 3 on 0.21.28.

Repro shape

In the 0.21.33 failure, the step was an assertion polling snapshots right after a screen transition (a peer mention opened from the notifications list). Failure is intermittent, and the request logs also show it on repeated snapshots.

Evidence excerpts

Run A, 0.21.28:

command=snapshot; helper version=0.21.28; versionCode=21028
android_snapshot_helper_capture durationMs=629: failed before returning parseable output
request_failed code=COMMAND_FAILED: Android snapshot helper failed: ...

Run B, 0.21.28:

android_snapshot_helper_session_stop: forceKilled=true
android_snapshot_helper_session_fallback: "Android snapshot helper session returned truncated XML"
android_snapshot_helper_capture durationMs=947: failed before returning parseable output
request_failed code=COMMAND_FAILED

Run C, 0.21.33 (two requests):

helper version=0.21.33; versionCode=21033
persistent-session answered=false; instrumentation fallback failed
capture durations: 1022 ms and 602 ms
reason: Android snapshot helper failed before returning parseable output
request_failed code=COMMAND_FAILED

Logcat:

• In run C, the Android automation logcat shows Force stopping com.callstack.agentdevice.snapshothelper ...: start instr, then the process start and UiAutomation: Initialized. No FATAL EXCEPTION, AndroidRuntime or NullPointerException line follows.
• The adb logcat -d -b crash dump taken after the suite is empty (run C; runs A/B predate that dump).
• So the artifacts show the helper exiting, but not an in-process exception.

Asks

1. Is there a helper-side log, or a flag such as --debug, that would record why the instrumentation exited (exit code, instrumentation result bundle, a INSTRUMENTATION_ABORTED / shortMsg)? Our host only records "failed before returning parseable output".
2. Could the one-shot fallback be retried once by agent-device itself when the persistent session has just been force-killed? In our logs, both paths fail inside the same ~1 s window right after a transition.

Related: #3901 (vanished-child NPE, 0.21.31), #3929 (persistent helper frozen on ColorOS; a different symptom).