QUALITY: Test-suite coverage gaps and schema drift

[echel0nn]

Labels: quality, tests
Severity: S2. The suite gives false confidence on the exact paths this sweep found broken.

Finding (path:line)	Issue
tests/platform/conftest.py:36 (12+ fixtures)	Tables built via SQLModel.metadata.create_all instead of running Alembic -> the schema under test can diverge from the migrated production schema (missing constraints/indexes/column types); tests pass on code that would fail on a real DB
tests/test_knowledge_hybrid_retrieve.py:13	Whole file pytest.mark.skip(reason='SQLite-era tests') -> 4+ knowledge-retrieval tests never run (false coverage)
tests/modules/vulnerability/workflow/test_two_phase_dispatch.py:188,204	Two dispatch tests permanently skipped pending a fixture never built -> two-phase dispatch has zero automated coverage
tests/platform/llm/test_client.py:310,336	Retry tests mock asyncio.sleep but never assert retry count or backoff intervals -> retry logic could lose its backoff silently without failing

Highest-value missing tests:

1. PersistContract.upsert/upsert_many (conflict path, no-natural-key path, None-id exclusion) -- zero tests.
2. ObligationSet/adjudicate() platform contracts -- zero unit tests (only implicit via nday).
3. RunState/PlatformResponse/RouteDecision JSON round-trip / model_dump(mode='json').
4. ReasoningCaseState.observables non-serializable-value rejection before the task queue.
5. Team isolation at the register_team_scope_listener() ORM hook in isolation (API tests pass even if the hook is disabled).
6. Workflow crash -> /re-enqueue recovery (cursor wipe + re-submit), the bridge between stuck and recovered.

Acceptance: fixtures run Alembic (or a documented parity check); no permanently-skipped test files (delete or fix); retry tests assert count + backoff; the six missing tests added, prioritizing the team-scope hook and serialization guards.

[echel0nn]

Design decision (2026-07-19). This test backbone is front-loaded. Tests/benchmark FIRST is the chosen extraction order (#25), so the frozen eval benchmark + record-replay tests over current vr/malware behavior are the critical-path first build. Given multi-tenant + evidence-grade, the highest-value items are fixtures running Alembic (not create_all) and the team-scope-hook test in isolation.

[echel0nn]

Implementation-ready design (2026-07-19). Wave 1c (gates all). Full design on disk: .run/designs/DESIGN_foundation_eval.md (8-heading template: grounded problem, design with signatures+DDL+Alembic, reworked concerns, test plan, rollout).

Key moves:

• Alembic-based fixtures replace create_all (kills schema drift); restore skipped knowledge_hybrid + two_phase_dispatch; retry tests assert backoff 1/2/4 + Retry-After.
• Six missing tests: persist upsert_many, obligations adjudicate, RunState serialization, observables JSON guard, team_scope listener, crash-recovery.