fix(tests): repair the order- and load-dependent failures parallelism exposed

[dieterolson]

Two failures that #4548 surfaced by running the suite under xdist. Neither is caused by the diffs that hit them — main is red on one of them right now.

1. E-stop state leak (real isolation bug)

main.control_context is a module-level singleton shared by the entire P1AM backend suite. test_estop_trigger cleared the latch as the last statement of its body:

response = client.post("/api/estop")
assert response.status_code == 200
assert "E-stop triggered" in response.json()["message"]
mock_estop.assert_called_once()
control_context.clear_estop()   # <- skipped if any assert above fails

Any assertion above it failing leaves E-stop engaged for whatever xdist schedules next on that worker. It surfaced as test_pid_tuning_tag_guards failing with assert "PID loop 1 CV tag 'TAG_255'" in 'E-stop active; output writes are inhibited.' — an unrelated assertion, in an unrelated file, only on some runs.

Reproduced deterministically before fixing rather than inferred: latch E-stop, then run the victim module → 1 failed; with the guard → 2 passed.

Fix is two-layer: the clear moves into a finally, and a new package conftest.py clears the latch on both boundaries of every test. Teardown alone would still leave each module's first test dependent on whatever ran before it — which is the coupling being removed. This is the same class as the import-time P1AM_DEV_NO_AUTH bug (#4061), whose fix was likewise a per-test fixture.

2. Wall-clock flake

test_100k_rows_remain_bounded_and_fast asserted elapsed < 0.5. Main tripped it at 0.5146 s under contention. The assertion guards against an accidental super-linear pass — which at 100k rows costs minutes, not tenths of a second — so any ceiling in this range catches the real regression equally well, while 0.5 s reports a scheduling accident as a code regression. Ceiling moves to 5.0 s with the reasoning inline; displayed_count == 2_000 remains the deterministic contract.

Verification

• Full P1AM backend suite: 1252 passed, 6 skipped, both serial and under -n 4.
• test_launch_monitor_linked_scatter.py: 9 passed.
• ruff check + format clean on all three files.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.