test(checks): fix cross-test pollution in test_checks.py (#1741)

[johnrtipton]

Summary

Fixes #1741 — two tests in python/tests/test_checks.py failed deterministically under cross-test ordering but passed in isolation:

• TestC003DaphneOrdering::test_c003_daphne_missing_info
• TestSuppressChecks::test_no_suppress_by_default

Both failed with a polluted C003 assertion (0 == 1 — C003 did not fire despite all firing conditions being met).

Proven polluter + mutated global state

Polluter: python/djust/tests/test_dev_server_watchdog_missing.py — specifically the block_watchdog fixture, exercised by test_check_hot_view_replacement_survives_without_watchdog, which evicts and re-imports djust.checks (and djust.dev_server) to test the no-watchdog import path (#994).

Mutated state (proven, not guessed): a re-import via importlib.import_module("djust.checks") rebinds both sys.modules["djust.checks"] and the parent-package attribute djust.checks. The fixture restored only sys.modules, leaving the two pointing at different module objects.

Diagnostic probe in the failing order showed:

globals_is_module = False      # check_configuration.__globals__ is NOT vars(checks-imported-by-test)
sysmod_is_checks  = False      # sys.modules["djust.checks"] != (from djust import checks)
pkg_attr_is_localimport = True # djust.checks (pkg attr) == (from djust import checks)  -> module A
cc_from_sysmod    = True       # sys.modules["djust.checks"].check_configuration is check_configuration -> module B
globals_has_asgi_result = True # the function under test sees the REAL (unpatched) _has_asgi_server

So the failing test did:

• from djust import checks → resolves the package attribute → module A (which the fixture restored)
• from djust.checks import check_configuration → resolves sys.modules → module B (the stale re-imported copy)

monkeypatch.setattr(checks, "_has_asgi_server", lambda: False) patched A, but check_configuration's globals belonged to B, so the patch silently no-op'd → the real _has_asgi_server() returned True → C003 was suppressed → len(c003) == 0.

Reproduced failing order

.venv/bin/python -m pytest python/djust/tests/ python/tests/test_checks.py -p no:randomly -q
# -> 2 failed (the two cited tests), 3294 passed

(Also surfaces intermittently in CI shards / xdist, per #1736/#1739 Stage-11 reviewers.)

Isolation fix

In block_watchdog (function-scoped fixture): snapshot and restore the parent-package attribute (djust.checks, djust.dev_server) alongside sys.modules, so the two stay consistent after teardown and from djust import checks / from djust.checks import X resolve the same object. Test-only change; no runtime behavior change.

Module-cache / registry audit (recurring class)

• djust check registry (@register("djust") into Django's global registry): the actual tech-debt: fix 2 pre-existing test-pollution failures in test_checks.py + audit module-level caches for test-reset #1741 cause was the module re-import desyncing the registered-check module copies — fixed by the package-attr restore above. ✓
• _route_map_cache (routing.py:27, has _reset_route_map_cache()): already wired into autouse reset fixtures in both tests that mutate it — test_route_map_derivation.py (autouse _reset_state) and test_client_config_tag.py (autouse _reset_route_map). No gap. ✓
• checks.py module-globals: all compiled regexes / frozensets (immutable; not mutated by tests). No reset needed. ✓
• No second polluter found.

Verification

Gate-off (#254): reverted the fix (git stash) → the exact 2 tests fail again; restored → pass. The fix is confirmed to be what resolves the failure.

3-clean-runs gate (pollution-class, mandatory):

• test_checks.py 3× consecutive → 204 passed each. ✓
• Cross-dir order python/djust/tests/ python/tests/ 3× consecutive → 5471 passed, 9 skipped each. ✓
• xdist -n auto cross-dir → 5471 passed, 9 skipped. ✓
• Exact previously-failing order python/djust/tests/ python/tests/test_checks.py → 3296 passed. ✓

Closes #1741

🤖 Generated with Claude Code

[johnrtipton]

Stage 11 Adversarial Review — PR #1743 (fix #1741 cross-test pollution)

Verdict: 🟢 APPROVE. Test-only fix, root cause correctly diagnosed, independently reproduced pre-fix / confirmed post-fix, 9/9 clean pollution-gate runs, restore is complete and order-correct, no second/latent polluter found.

---

Mandatory pre-review checks

• 🟢 Stale-base (Full HTML update path mangles page rendering #250): git rev-list --count HEAD..origin/main = 0. Not stale. git log origin/main..HEAD shows exactly the 2 PR commits.
• 🟢 Two-commit shape: c7ebcb51 = fix only (1 file: test_dev_server_watchdog_missing.py, +33/−2, no CHANGELOG); 6ef261af = CHANGELOG only (+1). Clean boundary per Action feat: add event replay button to Events tab (#177) #181.
• 🟢 Scope: PR touches exactly 2 files — the test fixture + CHANGELOG. No prod module touched.

Load-bearing empirical reproduction (ran independently)

• 🟢 Pre-fix FAILS (cited order, origin/main test file): swapped in the origin/main version of the fixture file, ran pytest python/djust/tests/ python/tests/test_checks.py -p no:randomly -q →
  2 failed, 3294 passed — the exact two cited tests:
  TestC003DaphneOrdering::test_c003_daphne_missing_info + TestSuppressChecks::test_no_suppress_by_default.
• 🟢 Post-fix PASSES (same order, PR HEAD): 3296 passed, 3 skipped, 0 failed. The 2 previously-failing now pass.

3-clean-runs pollution gate (re-ran myself, did not trust the report)

• 🟢 python/tests/test_checks.py ×3: 204 passed / 204 passed / 204 passed.
• 🟢 cross-dir python/djust/tests/ python/tests/ (serial, pytest-randomly → 3 different seeds) ×3: 5471 passed, 9 skipped each.
• 🟢 cross-dir -n auto (xdist, mimics CI shard) ×3: 5471 passed, 9 skipped each.
• 9 clean runs total, zero failures under every ordering. No second polluter trips.

Second-/latent-polluter probe

• Grepped the whole test suite for importlib.reload / import_module / del sys.modules / sys.modules[...] writes.
• Only two other tests use the dangerous evict+reimport pattern: tests/unit/test_hot_view_replacement.py and tests/integration/test_hvr_flow.py. Both evict/reimport temporary top-level modules (hvr_fixture_*), which are not attributes of the djust package, so they cannot create the package-attr/sys.modules split. They clean up with sys.modules.pop (correct for top-level modules). Not same-class.
• tests/unit/test_ws_compression_config.py uses importlib.reload(djust.config) — reload mutates the same module object in place (no new object, no split). Safe.
• test_theming_register_get_roundtrip_invariant.py uses import_module on already-loaded modules (cache hit, no eviction). Safe.
• No latent same-class polluter found. 🟢

Restore correctness (#1741 invariant)

• 🟢 Snapshot is taken in fixture setup, before yield (lines 79-82, before blocker/yield) → captures the original objects, not a post-reimport copy.
• 🟢 Teardown restores BOTH sys.modules (lines 93-97) AND the parent-package attrs djust.checks + djust.dev_server (lines 101-106) to those original objects, with correct _MISSING/delattr handling for the attr-didn't-exist case.
• 🟢 Because both paths restore the same original object, the post-teardown invariant sys.modules["djust.checks"] is djust.checks holds — the exact invariant the original fixture violated. Empirically corroborated: gating the fix off (origin/main file) reproduces the failure; restoring the fix clears it.

Module-cache audit completeness (#1741 "stop the class recurring")

• 🟢 _route_map_cache claim verified: both python/djust/tests/test_client_config_tag.py:38 and python/djust/tests/test_route_map_derivation.py:33 have autouse _reset_route_map_cache fixtures (reset before+after each test). The 9 clean runs empirically confirm no residual route-map pollution.

CHANGELOG

• 🟢 Accurate, under ### Fixed, cites tech-debt: fix 2 pre-existing test-pollution failures in test_checks.py + audit module-level caches for test-reset #1741, root cause + "Test-only change; no runtime behavior change."

No 🔴, no 🟡. Recommend merge.

[johnrtipton]

Retrospective — PR #1743 (pipeline-drain v1.0.2-3)

Task: v1.0.2-3 — #1741 (Action Tracker #289): fix 2 deterministic test_checks.py pollution failures.

Quality: 5/5

Pollution proven by causation (not guessed), narrowest fix, 3-clean-runs gate honored, reviewer independently reproduced both directions.

What went well

• Root cause proven via a diagnostic probe, not bisect-and-hope. The polluter was test_dev_server_watchdog_missing.py's block_watchdog fixture: re-importing djust.checks rebinds BOTH sys.modules["djust.checks"] and the parent-package attribute djust.checks, but the fixture restored only sys.modules → the two pointed at different module objects. A later monkeypatch.setattr(checks, "_has_asgi_server", ...) patched the package-attr copy while check_configuration's globals lived in the sys.modules copy → patch silently no-op'd → C003 not suppressed → the 2 tests fail. The probe showed globals_is_module / sysmod_is_checks flip with/without the polluter — causation, not correlation.
• Narrowest robust fix: snapshot+restore the package attr alongside sys.modules in the one offending fixture; test-only, no runtime change. Re-establishes the sys.modules[x] is djust.<attr> invariant the original fixture violated.
• 3-clean-runs gate (pollution-class) honored and re-verified by the reviewer: 9 clean runs (file ×3, cross-dir serial ×3 across seeds, xdist -n auto ×3). No second polluter.

What didn't

• Nothing. 0🔴/0🟡.

Process finding

The module re-import / sys.modules desync is a distinct pollution sub-class from the usual "cache not reset": re-importing a SUBMODULE rebinds two references (sys.modules entry + parent-package attribute) and restoring only one leaves them split, so attribute patches land on the wrong object. Any fixture that evicts+reimports a djust.* submodule must restore BOTH. The reviewer's second-polluter probe confirmed the other evict+reimport tests operate on temp top-level modules (not djust.* package attrs), so they're immune — but the pattern is worth watching.

Verified

• Reviewer independently reproduced pre-fix-fails (2 failed on main, cited order) / post-fix-passes (3296 passed); gate-off non-tautological. _route_map_cache autouse resets confirmed present; checks.py globals immutable. All CI green (both Python shards). Two-commit shape intact.

RETRO_COMPLETE