fix: improve parallel orchestrator agent tracking clarity and cleanup

- Add comment on running_coding_agents explaining why feature_id keying
  is safe (start_feature checks for duplicates before spawning), since
  the sister dict running_testing_agents required PID keying to avoid
  overwrites from concurrent same-feature testing
- Clear running_testing_agents dict in stop_all() after killing
  processes so get_status() doesn't report stale agent counts while
  _on_agent_complete callbacks are still in flight

Follow-up to PR #130 (runaway testing agent spawn fix).

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

parallel_orchestrator.py

@@ -169,6 +169,7 @@ class ParallelOrchestrator:
         # Thread-safe state
         self._lock = threading.Lock()
         # Coding agents: feature_id -> process
+        # Safe to key by feature_id because start_feature() checks for duplicates before spawning
         self.running_coding_agents: dict[int, subprocess.Popen] = {}
         # Testing agents: pid -> (feature_id, process)
         # Keyed by PID (not feature_id) because multiple agents can test the same feature
@@ -906,6 +907,11 @@ class ParallelOrchestrator:
                 status=result.status, children_found=result.children_found,
                 children_terminated=result.children_terminated, children_killed=result.children_killed)
 
+        # Clear dict so get_status() doesn't report stale agents while
+        # _on_agent_complete callbacks are still in flight.
+        with self._lock:
+            self.running_testing_agents.clear()
+
     async def run_loop(self):
         """Main orchestration loop."""
         self.is_running = True