Refactor queue core and fix orphaned task cleanup

[himattm]

Summary

• Extract shared queue infrastructure into queue_core.py for code reuse between MCP server and CLI
• Fix critical bug where cancelled sub-agents leave orphaned tasks that block the queue forever
• Add PID reuse detection to distinguish dead processes from PIDs reused by unrelated processes (Chrome, etc.)
• Add instance tracking (server_id) to detect orphaned tasks from crashed/restarted processes

Changes

New: queue_core.py

Shared infrastructure used by both task_queue.py (MCP server) and tq.py (CLI):

• QueuePaths dataclass for consistent path management
• Database functions: get_db(), init_db(), ensure_db()
• Process management: is_process_alive(), is_task_queue_process(), kill_process_tree()
• Queue cleanup: cleanup_queue() with orphan detection
• Logging utilities

Bug Fix: Orphaned Task Cleanup

Problem: When sub-agents are cancelled, their queued tasks remained in the database forever, blocking subsequent tasks.

Solution:

1. is_task_queue_process(pid) - Checks if a PID is actually running task_queue vs being reused by Chrome/Safari/etc.
2. SERVER_INSTANCE_ID / CLI_INSTANCE_ID - Unique ID per process instance to detect stale tasks even if PID is reused
3. _active_task_ids tracking - Server tracks which tasks are actively being processed
4. asyncio.CancelledError handling - Properly cleans up when MCP clients disconnect

Other Improvements

• Add server_id column to queue table (with migration for existing DBs)
• Fix wait time calculation to use POLL_INTERVAL_WAITING instead of hardcoded value
• Add comprehensive tests for orphan cleanup scenarios

Test plan

• All 61 existing tests pass
• Manual testing: Start 2 sub-agents with 60s tasks → Cancel them → Queue is empty
• demo/test_pid_reuse.py verifies PID reuse detection works correctly
• CLI still works correctly with instance tracking

🤖 Generated with Claude Code