test(rls): R2 — background-thread + connection-pool edge-case coverage

CHANGELOG.md

@@ -13,6 +13,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Flake in `tests/test_store.py::test_do_cleanup_logs_errors`.** Test was using the `caplog.at_level(...)` context manager plus `caplog.text` substring check. That pattern failed intermittently on CI (observed on Python 3.11 and 3.14) even though the code under test (`_do_cleanup`) is fully synchronous — the flake was in pytest's log-capture path, not in the production code. Rewritten to use `caplog.set_level(...)` at fixture scope and inspect `caplog.records` directly (logger name + level + message substring), which is sturdier across pytest/Python versions and produces a richer failure message when it does fire. Verified 10/10 consecutive local runs post-fix. Closes [#374](https://github.com/cmeans/mcp-awareness/issues/374).
 
 ### Security
+- **RLS background-thread + pool edge-case coverage** — new `tests/test_rls_background.py` (11 tests) covers the execution contexts the request-path `rls_store` fixture doesn't exercise: the `_do_cleanup` daemon thread, the `upsert_embedding` path used by `server._embedding_pool`, and Postgres's transaction-local `set_config` semantics plus the combined pool/Postgres contract. Four cleanup tests and two embedding tests verify the full owner-scoped call paths (RLS policy + explicit `WHERE owner_id = %s` SQL filter + request-path `_set_rls_context`) catch cross-tenant regressions — tests pass only when all three defenses are intact, acknowledging the layered design rather than claiming any single layer is independently load-bearing. Five pool-guarantee tests: two use a raw `psycopg.connect` to directly verify Postgres's `set_config(..., is_local)` semantic (the `true` variant does NOT survive COMMIT; the `false` variant does), isolating the Postgres contract from `psycopg_pool`'s `RESET ALL` check-in reset; the remaining three codify the *combined* pool+Postgres behavior (no RLS residue across checkouts, after ROLLBACK, or across concurrent threads). Test-only change; no production code modified. ~2.6 s added to the suite; full suite: 998 → 1000 passing. Module docstring captures the audit summary: cleanup's owner enumeration relies on the pool role having `BYPASSRLS`, the per-owner DELETE is doubly scoped; embedding upsert is doubly scoped (SQL `owner_id = %s` + `_set_rls_context`). Closes R2 of [#359](https://github.com/cmeans/mcp-awareness/issues/359); closes [#362](https://github.com/cmeans/mcp-awareness/issues/362).
 - **RLS migration-safety test** — new `tests/test_rls_migration_safety.py` walks the Alembic migration path (`N-1 → head`) with two-tenant data seeded at N-1, then asserts tenant isolation still holds after applying the head migration. Catches the class of bugs where a future migration regresses an RLS policy, weakens a `WITH CHECK`, renames `app.current_user`, or adds a new owner-scoped table without `ENABLE ROW LEVEL SECURITY` — failures no scanner catches at authoring time. Each test runs in its own database (`CREATE DATABASE` / `DROP DATABASE` inside the shared `pg_container`) for full isolation. ~2.5 s added to the suite. Closes R3 of [#359](https://github.com/cmeans/mcp-awareness/issues/359) (RLS harness coverage extension tracking); closes [#360](https://github.com/cmeans/mcp-awareness/issues/360). A follow-up companion test for the `_system`-schema carve-out migration specifically was prototyped but deferred — hit an unresolved FORCE-RLS/BYPASSRLS interaction on Postgres 17 during seed-path development; the core isolation invariant this PR ships is the scope R3 promised.
 - **Third-party GitHub Actions pinned to full commit SHAs** (instead of floating `@vN` major-version tags). A moving major-version tag on a third-party action is an unreviewed supply-chain channel — the upstream owner can move the tag to any commit at any time, including after a repo compromise. Pinning to the commit SHA freezes the code under review. Affected references in `.github/workflows/`:
   - `codecov/codecov-action@v6` → `@57e3a136…` (v6.0.0) in `ci.yml`