Orchestrator deadlock in `executing` with stale `local://` assignment causes repeated `create_goal BLOCKED` sleep loop

[bpnace]

Summary

The agent gets stuck with an active goal in executing when a task remains assigned to a dead local worker (local://...).
The parent loop then repeatedly tries create_goal, receives BLOCKED, and goes to backoff sleep, while progress does not move.

Observed behavior

• Active goal remains active (example: 01KJXDGJ2212GXQDGVGE4H9WD8)
• orchestrator.state.phase = executing
• task_graph for active goal:
  • assigned: 1 (assigned to dead local://...)
  • blocked: 9
• children table still has the same local:// worker as running
• Logs repeat patterns like:
  • Recovering stale task from dead worker
  • then parent does create_goal -> BLOCKED
  • then sleeps with exponential backoff

Expected behavior

1. Dead local:// worker assignment is recovered once and does not re-enter the same stale loop.
2. Stale local worker rows are marked dead/unhealthy and excluded from reassignment.
3. Parent loop should not attempt create_goal while active goal is in progress.
4. Goal execution should resume (reassign to live worker or self-assignment fallback), not repeatedly sleep.

Reproduction (high-level)

1. Start with active goal in executing.
2. Ensure one task is assigned to a local:// worker that no longer exists.
3. Restart runtime.
4. Observe repeated stale-recovery + create_goal BLOCKED + backoff sleep loop.

Notes

This appears separate from planning/classifying fallback logic; here tasks exist but assignment targets stale local worker state.