emrg: test: harden stop_all hermeticity guard to all five killers (rant 2026-08-25T10:42:47)

[argszero]

Summary

Closes the ⛔ red-line test hole reported in rant 2026-08-25T10:42:47: on a Windows pytest run, tests/test_stop_all.py::test_stop_all_process_residual_still_aborts mocked is_win → True but only isolated stop_daemon — so stop_gui, stop_tui, stop_bundled_git and stop_lock_owners executed for real, killing the live EMRG GUI/TUI (and their process scans) on the host.

Changes

1. tests/conftest.py — the autouse red-line guard _guard_stop_all_hermeticity (renamed from _guard_stop_daemon_hermeticity) now replaces all five stop killers (stop_daemon, stop_gui, stop_tui, stop_bundled_git, stop_lock_owners) with raisers, so any test that reaches a real stop path fails loudly with AssertionError instead of killing processes. Existing isolation pattern (monkeypatch.setattr(_stop_all, "stop_x", lambda: None)) still overrides it as before.

2. Escape hatch — a new @pytest.mark.allow_real_stop("stop_lock_owners") marker opts a test out per function when it deliberately calls a stop function with its internal kill machinery already isolated. Applied to the two verified-hermetic direct-call sites:

   • test_posix_noop (is_win=False → early noop; _lock_owner_ps mocked)
   • test_stop_lock_owners_logs_diag (find_install_module_holders → [] + _lock_owner_ps mocked)
     Marker registered in pyproject.toml ([tool.pytest.ini_options]).

3. tests/test_stop_all.py::test_stop_all_process_residual_still_aborts — the flawed test now isolates all five killers (mirroring test_stop_all_retries_lock_kill), so the Windows branch can never perform a real kill.

4. Canary — TestStopAllHermeticityGuard is now parametrized over all five functions, verifying each guard raiser is installed (function-identity check only, never calls a stop path).

5. Agent.md — documented pytest count 1052 → 1056 (net +4: 5 parametrized canaries − 1 old canary), satisfying the emrg: sync test counts to 484 + guard test against doc drift (recurs #426/#430/#510) #511 doc-count guard.

Verification

• uv run pytest tests/ -v → 1055 passed, 1 skipped (1056 collected, doc-synced)
• uv run python -c "from emrg.client.app import run_client" → OK
• uv run python -m emrg --help → OK

No production code touched — emrg/_stop_all.py is unchanged (the Windows stop path is legitimate emrg stop behavior; the hole was purely test-side).

[argszero]

✅ LGTM — cycle

Fresh review of the stop_all hermeticity fix (rant 2026-08-25T10:42:47):

• tests/conftest.py: autouse guard now replaces ALL FIVE stop killers (stop_daemon/stop_gui/stop_tui/stop_bundled_git/stop_lock_owners) with raisers — matches the Windows 5-step _step_plan; any un-isolated real stop path fails loudly with AssertionError. The allow_real_stop marker opts out per function only for verified-hermetic direct calls.
• tests/test_stop_all.py:770: the flawed Windows-branch test now isolates all 5 killers (mirrors test_stop_all_retries_lock_kill), closing the hole where stop_gui/stop_tui/stop_bundled_git/stop_lock_owners ran for real.
• Escape hatch applied only to the 2 safe direct-call sites (test_posix_noop with is_win=False + _lock_owner_ps mocked; test_stop_lock_owners_logs_diag with both killers mocked).
• Canary TestStopAllHermeticityGuard parametrized over all 5 functions (identity check only, never calls a stop path).
• Marker registered in [tool.pytest.ini_options]; Agent.md count 1052→1056 syncs the #511 doc guard.

CI double-green (run 32803059504: test + test-windows both pass). Local verification: 1055 passed / 1 skipped, import + CLI OK. Production code untouched — the fix is purely test-side. No issues found.

[argszero]

✅ LGTM — cycle (2/3)

Head unchanged (06e79a0), CI double-green (run 32803059504: test + test-windows both pass), MERGEABLE. Re-verified the diff from cycle 105902: conftest guard replaces all five stop killers with raisers (matches Windows 5-step _step_plan), allow_real_stop escape hatch applied only to the two verified-hermetic direct-call sites, test_stop_all.py:770 now isolates all five (mirrors retries_lock_kill), 5-way parametrized canary, marker registered in pyproject, Agent.md count 1052→1056 syncs the #511 doc guard. No issues found.

[argszero]

✅ LGTM — cycle (3/3)

Head unchanged since the last review (6cebd41, includes the master-merge conflict resolution — Agent.md pytest count merged to 1058), CI double-green (run 32804511105: test + test-windows both pass), MERGEABLE/CLEAN. Full diff reviewed across cycles 105902/111222: conftest guard on all five stop killers, allow_real_stop escape hatch, test_stop_all.py:770 full isolation, 5-way canary, doc-count synced. No issues found — merging.