fix: exclude dead workers from reassignment to break orchestrator deadlock

[Matthew-Selvam]

Summary

Fixes #266, fixes #259. The orchestrator's stale-task recovery (handleExecutingPhase) resets a dead worker's task back to pending when isWorkerAlive() says the worker is gone — but never updates the worker's own status. Meanwhile findBusyAgentForReassign() only excludes idle agents when picking a busy worker to hand a freed task to, not dead ones. So the very next matching pass hands the recovered task straight back to the same dead worker. It dies again, gets "recovered" again, and reassigned again — an infinite recover-reassign-die loop that never completes the goal. That's what leaves orchestrator.state permanently stuck in executing and blocks all new goal creation (both #266 and #259 describe this from different angles — a low-RAM VM where workers OOM, and a stale local:// worker after a restart).

Fix

• When stale-task recovery detects a dead worker, mark its children row 'dead' and call agentTracker.updateStatus(address, "dead"), so it's excluded from future reassignment.
• Defense-in-depth: findBusyAgentForReassign() now also checks the existing isWorkerAlive() hook directly, in case a task ever ends up assigned to a dead worker through some other path than the recovery block above.

Testing

• tsc --noEmit passes.
• Added two regression tests:
  • findBusyAgentForReassign (via matchTaskToAgent) falls through to self-assignment rather than handing a task back to a worker isWorkerAlive reports dead.
  • The stale-task recovery path in handleExecutingPhase (via tick()) marks the dead worker's children row and tracker status, and — since the worker is now excluded — the same tick reassigns the freed task off it instead of looping back to it.
• Full orchestration test suite: 286/286 passing (orchestrator.test.ts, messaging.test.ts, task-graph.test.ts, orchestrator-harness.test.ts, local-worker-harness.test.ts, attention.test.ts, workspace.test.ts, local-worker-security.test.ts).

[Matthew-Selvam]

This is ready for maintainer review whenever you get a chance — happy to make any changes requested.