agents-runtime: background compaction stays running after normal idle exit

Versions

Reproduced in a local TypeScript integration with agents-runtime 0.6.3 based on 91758919586d7c0c70a7b8190a39b6e06fc2ea14. This build has downstream compaction configuration changes, but the background processWake lifecycle is unchanged. I also inspected main at bb397424db0e1c153dc356713fd3dfd40315470c; the same lifecycle is present there. I have not run the integration against that main SHA.

Bug description

Background compaction can finish successfully after the normal idle timeout ends its wake, but its checkpoint stays running and the summary is never applied. The original conversation remains available.

The non-blocking behavior in #4605 is intentional. The problem is the unfinished checkpoint after the background work has ended, not that a user turn returns before its summary finishes.

To reproduce without live inference, use a fixture turn whose reported usage exceeds the background threshold and hold the public summarizeComplete hook behind a barrier:

const gate = Promise.withResolvers<void>()
const returned = Promise.withResolvers<void>()
const summarizeComplete: NonNullable<AgentConfig['summarizeComplete']> = async () => {
  await gate.promise
  returned.resolve()
  return { content: [{ type: 'text', text: 'Generated summary.' }], stopReason: 'stop' }
}

1. Configure a 100 ms runtime idle timeout and a longer summary deadline.
2. Send one user message. Let the turn finish and background summarization start.
3. Keep the gate closed until the runtime reports that the wake has ended.
4. Resolve the gate and await returned.promise.
5. Read the persisted stream again. The latest status for that compaction generation is still running; no complete or failed checkpoint is saved.

The local regression fails with BACKGROUND_CHECKPOINT_MUST_FINISH, after confirming summarySettled: true and checkpointStatuses: ["running"]. This uses controlled completion, not a sleep or a stopped runtime process.

In packages/agents-runtime/src/process-wake.ts, the promise callback only updates the wake-local pendingBackgroundCompaction slot and interrupts an active idle wait. After an idle timeout with a pending slot, the loop exits and cleanup closes the writer and sends done. There is no continuation or cancellation for that slot. A subsequent wake creates a new slot.

The 100 ms idle window makes the test quick. The defaults also allow this ordering: 20 seconds idle versus a 120 second summary deadline. This does not mean every background summary fails.

Expected behavior

On a normal idle exit, finish the generation through an owned continuation, or cancel it and record a terminal status. Either policy is reasonable. A summary that has finished should not leave an orphaned running checkpoint or perform work whose result no active wake can use.