fix(tests): conftest re-executed already-loaded modules, breaking isinstance across the suite (#12839)

[mrveiss]

Closes #12839

Thinking Path

Two failing tests, and I assumed one cause. There were two independent defects — fixing the obvious one left both tests still red, which is what forced the second investigation.

Defect 1 — the KB mock lied about its own interface. ClaimClassifier._search_kb picks its branch with hasattr:

if hasattr(self.kb, "search_async"):   results = await self.kb.search_async(...)
elif hasattr(self.kb, "search"):       results = self.kb.search(...)

The fixture was a bare MagicMock(), which auto-creates every attribute, so hasattr(kb, "search_async") was True and the classifier awaited a plain MagicMock — object MagicMock can't be used in 'await' expression. The classifier caught that, logged it, and returned no results, so the test was quietly asserting against a degraded error path instead of the one it meant to exercise. spec=["search"] makes the probe answer honestly.

Defect 2 — the same module executed twice. Fixing the mock did not fix isinstance, so the duplicate-identity theory needed proof rather than assumption. A stack probe in knowledge_grounding_models.py during a real pytest run showed both executions:

### EXEC knowledge_grounding_models
  services/claim_classifier.py:31  from .knowledge_grounding_models import (

### EXEC knowledge_grounding_models
  conftest.py:1141  pytest_configure -> _real_load_light_services()
  conftest.py:1118  _real_load_and_bind(f"services.{name}", svc_path)

_real_load_and_bind re-executes the file and swaps a fresh module into sys.modules, while services.claim_classifier keeps holding the classes from the first execution. Hence isinstance(x, Claim) returning False for an object whose repr says it is a Claim, with only one sys.modules entry to inspect.

Re-execution exists solely to replace a MagicMock stub, so an already-real module loaded from that same path is now left alone. The parent bind is still (re)applied on that skip path — patch("pkg.mod.NAME") resolves via getattr(pkg, "mod"), and dropping it would resurrect the inert-patch bug from #11532/#11618. That case has its own test.

What Changed

• conftest.py — _real_load_and_bind returns early when sys.modules[name].__file__ already matches the target path, after re-applying the parent bind.
• tests/services/test_claim_classifier.py — KB fixture uses MagicMock(spec=["search"]).
• tests/test_conftest_real_load_identity.py (new) — pins the loader invariant.

Verification

conftest.py is loaded by every backend test, so this was measured against a pristine origin/Dev_new_gui worktree using an identical command:

baseline: 14 failed, 5304 passed, 130 skipped
branch:   12 failed, 5310 passed, 130 skipped

failures only on this branch (regressions):  NONE
failures fixed:  test_claim_classifier.py::TestIndividualClassification::test_classify_claim_basic
                 test_claim_classifier.py::TestBatchClassification::test_batch_classify_single_claim

The 12 remaining failures are byte-identical on both trees (gitlab/gitea connector tests requiring a live Redis, etc.) and are untouched here. The +6 passes are the 2 fixed tests plus 4 new ones covering: an already-real module is not re-executed, an instance stays isinstance-valid across a repeat load, a stub is still replaced by the real module, and the parent bind is applied on the skip path.

A note on the first verification attempt, since it nearly produced a false conclusion: running tests/ llc/tests/ together reported "Interrupted: 49 errors during collection", which looked like this change breaking the suite. The identical command on the pristine baseline produced exactly the same 49 errors — pre-existing, triggered by combining those two paths in one invocation, and it meant nothing had actually executed. The numbers above come from a subset that collects cleanly on both trees.

Model Used

Claude Opus 5

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!