Restart reviewer terminal when a newer PR head supersedes an in-flight review

[neversettle17-101]

Migrated from aoagents/ReverbCode/issues/406

Source: aoagents/ReverbCode#406

---

Problem

AO currently supersedes stale running review_run rows when a worker PR gets a newer head SHA, but the shared reviewer terminal is only notified with the new task if it is still alive.

That protects the daemon from accepting stale results, because a later ao review submit --run <old-run> is rejected once the old run is marked failed. However, the reviewer process itself may still be actively reasoning about the old commit. Injecting a new prompt into the same terminal can mix old and new review context.

Example:

1. Trigger review for PR feat: implement web dashboard with attention-zone UI and API routes #1 at sha-old.
2. AO creates run-old and starts reviewer terminal review-<worker>.
3. Worker pushes sha-new while the reviewer is still working.
4. User triggers reviews again.
5. AO marks run-old failed and creates run-new for sha-new.
6. AO currently reuses the live reviewer terminal and sends a new prompt.

The storage state is correct, but the live reviewer execution is not cleanly cancelled.

Desired behavior

When a new trigger supersedes one or more running review runs for an older SHA, AO should stop the current reviewer execution and start a fresh reviewer terminal for the current review queue.

High-level behavior:

• Same current-head running run: reuse/skip as today.
• Older running run for the same PR: mark stale run failed, destroy reviewer runtime, spawn fresh reviewer for the current queue.
• If old reviewer later submits the old run, submit remains rejected because the old run is no longer running.

Probable fix

• Extend the review launcher interface with Destroy(ctx, handleID).
• Extend the private reviewer runtime interface to include existing ports.Runtime.Destroy.
• In Trigger(sessionID), track whether SupersedeStaleRunningReviewRuns(...) superseded any old running runs.
• If stale running work was superseded and the shared reviewer handle is alive:
  • destroy the reviewer runtime handle,
  • spawn a fresh reviewer instead of notifying the old terminal.
• Keep the session-level reviewer handle stable (review-<workerID>); zellij can destroy/recreate the same runtime id.
• If destroy/spawn fails, mark newly created runs failed using the existing trigger failure path.

Tests

• Stale running old SHA causes destroy + spawn, not notify.
• Current-head running same SHA does not destroy or spawn.
• Multi-PR trigger with one stale running run restarts reviewer once and launches the fresh current queue.
• Destroy failure fails newly created runs and returns an error.
• Submit for the old stale run remains rejected because the run is failed.

Notes

This issue is separate from #395. #395 makes review runs PR-scoped and uses one shared reviewer terminal per worker session. This issue tightens the behavior when that shared reviewer is working on a commit that has already been superseded.