fix(db): StaticPool for in-memory SQLite + module-cache reset between tests

[gizmax]

Summary

Architectural prerequisite for issue #218 (test isolation). Forces `poolclass=StaticPool` for in-memory SQLite URLs in the engine factory + adds a function-scoped autouse fixture that clears module-global dicts in routes.py and the rate-limiter window store between tests.

Why

The test suite uses `sqlite+aiosqlite://` (no path) for an in-memory database. Without StaticPool, every `async_session()` call may open a fresh aiosqlite connection, and each fresh in-memory connection sees its OWN empty database. A row written by session A is invisible to session B, which surfaces in CI as intermittent 'no such table: runs' / 'no rows returned' failures - the symptom diagnosed at length in `.autoresearch/test-isolation-results.tsv`.

This also explains why previous cleanup-fixture attempts in conftest.py made things strictly worse (65 failures -> 400+ errors): the fixture's row-wipe ran on a different in-memory DB than the test code subsequently read.

What changes

src/sandcastle/models/db.py

• New `_is_in_memory_sqlite(url)` helper recognising 3 URL shapes
• `_build_engine_kwargs()` adds `poolclass=StaticPool` when URL is in-memory
• File-backed SQLite and PostgreSQL paths unchanged

tests/conftest.py

• Function-scoped autouse fixture clears `routes._hub_cache`, `_batch_store`, `_stats_cache`, rate-limiter window store before every test
• Shared `_run_async` helper for session-level fixtures

What this does NOT do

Does not close #218 by itself. Local pytest delta with this change:

• Passed: 15,009 (unchanged)
• Failed: 70 (unchanged)

Zero new failures, but the 65-test pollution baseline remains. This is the architectural prerequisite.

Follow-ups in #218

1. Per-test SAVEPOINT isolation via session.begin_nested()
2. Move module-global caches behind FastAPI Depends()
3. Move execution_limiter similarly

After those three land, the baseline collapses to zero.

[gizmax]

CI confirmation

Job	Result
Dashboard build + tests	✓ PASS
Python tests	matches main baseline

Python: 15,009 passed / 70 failed / 1 error - identical to main. Zero regression, zero improvement, as predicted in the PR description.

The architectural change is correct (StaticPool is the right pool class for in-memory SQLite) but it does not reduce the 65-test pollution baseline on its own. The remaining work in #218 needs to land alongside this for the count to drop.

Merge options

1. Merge now - foundation work, clearly correct, makes the per-test SAVEPOINT refactor in Test isolation: 65 tests fail in full suite, pass in isolation #218 actually possible.
2. Hold - bundle with the SAVEPOINT work into one PR so the failure-count delta is visible at merge time.

Recommend option 1 - the file diff is tiny (97 lines), the change is obviously right, and follow-on PRs land cleaner once this foundation is in.