A pipeline that is “usually green” is not the same as a release gate. In more than one Android audit we have found cold-start crashes gated behind a flaky emulator job that engineers learned to re-run until it passed.
What to inspect
- Jobs marked optional or
continue-on-errorthat still cover startup - Emulator images that differ from the minimum OS you claim to support
- Initialization order in
Applicationsubclasses and ContentProviders - Third-party SDKs that assume network on first launch
A practical habit
Treat re-runs as a smell. If a startup job fails twice in a week, log it in the release readiness memo even when the eventual colour is green. Auditors will ask for that history; having it ready shortens the review.
When we flag a cold-start issue as a blocker, we include the stack signature and the exact job name that should have caught it. That combination is what moves a debate from opinion to scheduling.