Fix cross-test mock leak behind the flaky parallel test failures

[blaipr]

Merge PR #388 first — it makes full parallel Python test runs deterministic for reviewing everything else.

SUMMARY

Fixes #379 — and the root cause turned out to be none of the suspects in the issue. Six unit tests assigned a Mock directly onto the model manager, with no patch machinery:

task.model.objects.get = mock.Mock(return_value=job)   # task.model is the Job / AdHocCommand class

That permanently replaces Manager.get for the remainder of the pytest-xdist worker process. Whichever functional test files later shared that worker became the 'flaky' victims, and the leaked mock explains every observed failure family:

• api/test_survey_spec.py: Job.objects.get(...) returned a MagicMock → json.loads(job.extra_vars) failed with 'must be str … not MagicMock' (up to 13 parametrized failures per run).
• utils/test_update_model.py: update_model() returned the unit test's leftover in-memory Job built with Project(pk=1); the test saved it, inserting a row whose project_id=1 references nothing — which Django's whole-DB check_constraints() at teardown reported as the mysterious 'main_job … invalid foreign key' IntegrityError.
• commands/test_secret_key_regeneration.py: same frankenjob → garbage decrypted start_args.

The xdist scheduling lottery decided which files shared a worker with the leakers each run — hence the random victim set, why every victim passed in isolation, and why pairwise reproductions failed.

How it was found

A temporary pytest plugin scanned every model manager for Mock instance attributes after each test and logged transitions per worker. Both leak sources (unit/test_tasks.py, unit/tasks/test_jobs.py) were caught within two full parallel runs, attributed to the exact tests.

The fix

All six sites become mocker.patch.object(task.model.objects, 'get', return_value=...) — same behavior inside the test, automatically unwound by pytest-mock at teardown.

ISSUE TYPE

• Bug, Docs Fix or other nominal change

COMPONENT NAME

• API

ASCENDER VERSION

awx: 25.4.1.dev5+gcda0899.d20260610

ADDITIONAL INFORMATION

Validation: five consecutive full parallel runs (py.test --create-db -n auto --dist=loadfile awx/main/tests/unit awx/main/tests/functional awx/conf/tests awx/sso/tests) all returned 3476 passed, 0 failed, with the leak detector confirming no manager is ever left mocked. Before the fix, roughly two of every three such runs failed.

Note: the TransactionTestCase classes called out in the original issue text are innocent — their teardown flush verifiably leaves all tables empty. Test-only change; production code untouched. Independent of all open PRs.

[cigamit]

Thanks for tracking this down, was bugging me for a while but was so sporadic (only failed for me 1 in 10 times) that it was hard to pin down. Ran through the tests multiple times and it has not occurred again.