* fix: deliver classic tool approvals across server instances

Classic (non-durable) agents kept a pending tool approval in the memory of the single instance running the agent. In multi-instance production the approval response often hit a different instance with no record of it, so it was dropped and the agent timed out after 10 minutes.

When an approval lands on an instance that isn't holding it, persist the decision to a new tool_approval_decisions table; the instance running the agent polls and consumes it. Same-instance approvals still resolve in memory. Includes the table, migration, cleanup, and a cross-instance test.

* fix: harden cross-instance tool approval per review feedback

- Guard the approval poll with an in-flight flag so DB latency can't stack one query per second per approval and exhaust the connection pool.
- Use Promise.allSettled for the approval batch so one failed DB write reports per-toolCallId instead of 500-ing the whole batch.
- Add a settled flag set synchronously on resolve, checked before the destructive DB consume, and re-record the decision if it was consumed after the approval already settled (closes the timeout/poll TOCTOU).
- Pass an empty conversationId instead of 'unknown' so the poll guard disables rather than querying an id that never matches.
- Convert waitForApproval to a single options argument (7 positional params).
- Log full error objects in the poll catch to preserve stack traces.
- Run the two cron cleanups independently so one failure doesn't skip the other.
- Add a cross-instance denial test covering reason propagation.

* fix: distinct log for approval re-record failure

The re-record after a post-settle consume shared the poll's outer catch, so a re-record (write) failure was logged as a poll (read) failure. Wrap it in its own try/catch with a distinct error log.

* chore: add conversationId to cross-instance delivery log

Makes the delivery log self-contained for triage correlation by conversationId.

GitOrigin-RevId: 1f0279996b4055a67695b1fbc5e875f8e1ed6a9d