fix: surface silent task failures instead of acking them as completed

[kpscheffel]

Type of Change

• Fix - bug fix or security fix to source code

Description

What. Stops the agent-runner from acking scheduled tasks as completed when the SDK call actually failed silently — and adds a host-side scan that catches anything that bypasses the runner-side fix.

Why. When the Claude SDK terminates with an error subtype (error_max_turns, error_during_execution, …) or the stream closes without a terminal result event, the v2 poll-loop unconditionally called markCompleted and produced no user-visible signal. Combined with the host's syncProcessingAcks collapsing both completed and failed processing_ack rows into messages_in.status='completed', scheduled tasks could fail every day and look healthy — operators had no way to tell. Reproduced in real data: three consecutive morning-briefing runs marked complete with zero messages_out.

How it works.

1. container/agent-runner/src/providers/types.ts — result ProviderEvent carries the SDK's subtype. Three buckets: undefined = synthetic mid-turn (e.g. compact_boundary, dispatch text, do not ack); error_* = terminal failure; otherwise = terminal success.
2. container/agent-runner/src/providers/claude.ts — yield site reads (message as { subtype?: string }).subtype and passes it through.
3. container/agent-runner/src/poll-loop.ts — processQuery tracks sawTerminalResult and lastErrorMessage. Error subtype → markFailed(initialBatchIds) and write a ⚠️ Task did not complete: <subtype> chat to the originating channel. After the stream ends without a terminal result → same. The outer catch path now calls markFailed instead of always markCompleted. markFailed becomes array-aware to mirror markCompleted. MockProvider updated to yield subtype: 'success'. New stopSignal: AbortSignal on PollLoopConfig so tests can halt the loop deterministically (production callers omit it).
4. src/db/session-db.ts — syncProcessingAcks propagates failed as failed instead of collapsing to completed, with a guard so an already-completed row isn't downgraded by a stale ack.
5. src/host-sweep.ts — new detectAndNotifySilentTasks step. Scans kind='task' rows acked completed/failed in the last 15 min with zero messages_out in the same window, DMs the session's originating messaging group via getDeliveryAdapter().deliver(...), dedups per-process. Caps the dedup Set at 1000 entries.

How it was tested.

• 5 new container tests in container/agent-runner/src/poll-loop-failure.test.ts covering: error subtype path, stream-closed-without-terminal path, in-stream error captured in DM, empty success result legitimately ack'd, synthetic mid-turn followed by terminal success.
• 3 new host tests in src/db/session-db.test.ts covering syncProcessingAcks propagating failed as failed, propagating completed as completed, and refusing to downgrade an already-completed row.
• Verified on a real install — the day before this fix, three consecutive morning-briefing tasks (task-1777349525294-capq6s, task-1777435932858-rqa22y, task-1777522278899-8mqkfc) each acked completed with 0 messages_out. With this patch deployed, future failures will markFailed and DM ⚠️ Task did not complete: … to the channel; the host-side scan will catch any path the runner-side missed.