fix(agent-runner): exit on persistent inbound.db corruption errors

[kartast]

Problem

On Docker Desktop macOS, the agent-runner container can get stuck in an endless database disk image is malformed log loop on the follow-up poll, blocking message delivery indefinitely. Observed in production — container ran for 25+ minutes emitting:

```
[poll-loop] Follow-up poll error: database disk image is malformed
[poll-loop] Follow-up poll error: database disk image is malformed
... (2/sec, indefinitely)
```

Host-side `PRAGMA integrity_check` on both `inbound.db` and `outbound.db` returns `ok` — the file isn't actually corrupt. This is the cross-mount page-cache coherency issue already documented in `container/agent-runner/src/db/connection.ts:11-18`: Docker Desktop's virtiofs / gRPC-FUSE layer can latch a torn snapshot mid-host-write, and every fresh `openInboundDb()` in the same process sees the same broken view. Reopening the handle does not recover; only a fresh container mount does.

The follow-up poll's catch block (poll-loop.ts:358-367) logs the error and keeps polling — which is correct for transient errors but catastrophic for this one. The user's only recourse is `docker restart `.

Fix

After `CORRUPTION_STREAK_EXIT` (10) consecutive corruption errors (~5 s at `ACTIVE_POLL_INTERVAL_MS = 500ms`), exit the process with code 75. The host's existing container-respawn-on-wake path then brings up a fresh container with a fresh mount, which clears the poisoned page cache.

• Streak counter is scoped to `processQuery`'s `pollHandle` so it resets on any successful poll or non-corruption error.
• Threshold of 10 is intentionally well above any plausible single torn read during a host write burst but well below the 30-min stale-heartbeat ceiling, so users see fast recovery instead of long stalls.
• Exit is deferred 100 ms so the explanatory log line flushes through Docker's log driver before the process dies.

Test plan

• `bun test src/poll-loop` — 3 new `isCorruptionError` tests pass (existing failure in `formatter > should format multiple chat messages as XML block` is pre-existing and unrelated).
• `bunx tsc --noEmit` — clean.
• Reproduced locally by triggering the symptom (host-side message write race), confirmed container now exits + respawns instead of looping forever.

Notes

• Only addresses the symptom (stuck-forever loop), not the underlying Docker Desktop coherency race. The race itself is hard to eliminate without switching off Docker Desktop's gRPC-FUSE backend. The symptom fix is sufficient because the race is rare enough that occasional 5-second self-heals are acceptable, whereas indefinite stalls are not.
• Exit code 75 (`EX_TEMPFAIL`) is conventional for "transient failure, please retry".

[gavrielc]

@kartast Thank you for the contribution!