fix(kanban): terminate live workers on manual exits from running

[Dusk1e]

Summary

• Manual running -> done/blocked/scheduled/archived transitions now best-effort terminate the live claimed worker before clearing run metadata.
• Added a shared helper in hermes_cli/kanban_db.py so CLI/core paths and dashboard direct-status moves use the same worker termination logic.
• Preserved the self-completing worker path with a guard that skips killing the current process when the worker is closing its own run.
• Added regression coverage for complete_task, archive_task, and dashboard drag-drop off running.

Validation

• PASS: python -m pytest tests/hermes_cli/test_kanban_db.py tests/plugins/test_kanban_dashboard_plugin.py -q --timeout-method=thread -k "complete_task_terminates_foreign_running_worker or complete_task_skips_terminating_self_worker or archive_task_terminates_foreign_running_worker or patch_status_ready_terminates_running_worker or patch_status_archive_closes_running_run or patch_status_done_with_summary_and_metadata"
  Result: 6 passed, 257 deselected
• PASS: manual smoke repro with a real sleeping subprocess
  Result: complete_ok=True, status=done, worker_pid=None, run_outcome=completed, proc_exited=True, returncode=15

Additional broader check

• PARTIAL: python -m pytest tests/hermes_cli/test_kanban_db.py tests/plugins/test_kanban_dashboard_plugin.py --timeout-method=thread -vv --maxfail=3
  Result before stop: 125 passed, 3 failed
• The 3 failures were unrelated to this change and came from existing _resolve_hermes_argv expectations on this Windows/Python shim environment:
  • test_resolve_hermes_argv_prefers_path_shim
  • test_resolve_hermes_argv_hermes_bin_bare_name_uses_path
  • test_resolve_hermes_argv_falls_back_to_module_form_when_no_path_shim

Notes

• I had to run pytest with --timeout-method=thread because this Windows environment does not support the default SIGALRM path used by pytest-timeout.
• There is also a non-blocking pytest temp cleanup PermissionError at process exit on this machine; it does not affect the kanban fix itself.

[teknium1]

Thanks for addressing a real worker-lifecycle gap. Current complete_task clears worker_pid while moving a task to done (hermes_cli/kanban_db.py:4053-4069), and the same gap remains in archive, schedule, and dashboard direct-status transitions (hermes_cli/kanban_db.py:5431-5449, 5812-5854; plugins/kanban/dashboard/plugin_api.py:1023-1039).

Problems

• The diff adds termination behavior to block_task and schedule_task, but its new tests cover only complete, archive, and dashboard drag-drop. Current block_task has separate dependency, loop-routing, and ordinary branches that clear worker_pid (hermes_cli/kanban_db.py:4825-4967), so these paths need explicit coverage too.

Suggested changes

• Add foreign-worker and self-worker regression coverage for manual running -> blocked and running -> scheduled transitions when salvaging this into the current implementation.
• The base has substantially moved; port the helper deliberately into current lifecycle code rather than applying the old hunks mechanically.

This is an automated hermes-sweeper review.