test(mcp): stop MCP dispatch tests from leaking real spawn watchdogs

[basil-k-aji-dev]

Closes #43.

Problem

mobile_run_task arms _start_spawn_watchdog after Popen returns. Four tests in test_mcp_tools.py mock Popen but leave the watchdog live, so each starts a real daemon thread monitoring a pid that came from a MagicMock.

Those threads outlive temp_trace_env's directory and its monkeypatches. Past the 60-second deadline they go on to terminate a process tree, cancel a reservation, write into the restored trace directory, and dispatch failure notifications — for tasks that never existed. The issue reports a real desktop "Artemis Task Failed" notification naming fake pid 54321.

Reproduction

I reproduced it with the guard the issue describes — a temporary conftest.py that intercepts threading.Thread.start for spawn-watchdog-* names, records the attempt without starting a thread, and asserts none occurred:

$ uv run pytest -q -p no:randomly tests/unit/mcp/test_mcp_tools.py
ERROR test_mobile_run_task_reserves_and_passes_global_queue_ticket
ERROR test_mobile_run_task_with_device_serial
ERROR test_mobile_run_task_forwards_pro_tuning_to_background_runner
ERROR test_mobile_run_task_omits_pro_tuning_flags_when_unset
AssertionError: Test leaked 1 real spawn watchdog thread(s)
25 passed, 4 errors in 0.50s

Same four tests, same counts as the issue.

The run also surfaced something not in the original report — the leaked watchdog reached the notifier and made an outbound HTTP request from a unit test:

urllib.error.URLError: <urlopen error [Errno 111] Connection refused>

so the blast radius includes network egress, not just local process and filesystem effects.

Change

A module-local autouse fixture records watchdog launches instead of starting them, and returns the recorded calls so tests can assert on them.

Rather than only suppressing the call, the queue-ticket and device-serial dispatch tests now assert its exact arguments:

assert stub_spawn_watchdog == [(result["trace_id"], 43210, "queue-ticket-1", "conv-1")]

so arming the watchdog goes from untested-but-harmful to actually covered.

The fixture is scoped to this module, not the suite. Production watchdog behaviour — startup, terminal tasks, deadlines, cleanup, notifications — stays covered by tests/unit/mcp/test_spawn_watchdog.py, so no coverage is lost.

Test-only; no production behaviour changes, as the issue notes none are needed.

Verification

Command	Result
tests/unit/mcp/test_mcp_tools.py under the leak guard	25 passed (main: 25 passed, 4 errors)
tests/unit/mcp/test_mcp_tools.py normally	25 passed, no URLError
tests/unit/mcp/test_spawn_watchdog.py	4 passed
uv run pytest -q (full deterministic suite)	2096 passed, same 89 failing node ids as main — 0 new
uv run ruff check tests/	All checks passed

Both match the issue's "after" log (25 passed, 4 passed). The guard conftest.py was temporary and is not part of this PR.

make typecheck could not be run in my environment — the pyright wheel downloads a Node runtime on first use and my sandbox blocks it. The touched file is not in pyright-core.json.

[basil-k-aji-dev]

Closing as a duplicate. #42 by @wellorbetter already covers issue #43 and was opened first — that one should get the review, not this.

My mistake: I didn't check the open PR queue for an existing claim before sending this. Sorry for the noise.