Fix onStop handling for container rollouts

[SagsMan]

1. Summary

This patch fixes container lifecycle handling when Cloudflare replaces a container during an image rollout.

1. Recognize both the existing runtime exit message and the rollout-specific message.
2. Preserve the numeric exit code so stopped-event recovery can replay lifecycle hooks.
3. Add regression coverage for message parsing, unrelated runtime errors, and one-time onStop() replay.

2. Root cause

The SDK only matched this message shape:

Runtime signalled the container to exit: 0

During a rollout, the runtime emits:

Runtime signalled the container to exit due to a new version rollout: 0

Because the rollout message contains additional text before the colon, the SDK did not recognize it as a runtime-signalled exit. The monitor then stored the container as plain stopped instead of stopped_with_code.

syncPendingStoppedEvents() does not replay onStop() for the plain stopped state, so cleanup and persistence hooks were skipped during recovery.

3. Fix

The runtime-exit matcher now uses the shared message prefix and extracts the terminal numeric exit code from either message format. This keeps the existing behavior intact while handling rollout replacements correctly.

4. Validation

1. pnpm run test:unit — 37 tests passed
2. pnpm run typecheck — passed
3. pnpm run lint — passed
4. pnpm run build — passed
5. pnpm run format:check — passed
6. git diff --check — passed

The regression tests verify that:

• the existing runtime exit message still produces stopped_with_code;
• the rollout message produces stopped_with_code with exit code 0;
• unrelated runtime errors are not classified as signalled exits; and
• onStop() is replayed exactly once.

5. Execution proof

The image below shows the fresh before-and-after unit-test evidence. The parent revision fails to recognize the rollout message and skips onStop(). The PR branch passes the targeted regression tests and the full unit suite.

Open the proof image directly

Fixes #253