fix(recall): enforce expiry/supersession in top-K recall SQL (TAP-4586)

[wtthornton]

What

Push the expiry/supersession predicate into the KNN + FTS recall SQL so expired and superseded rows no longer consume the top-K budget. Previously they were dropped only after the top-K cut by the Python is_temporally_valid filter, so stale rows could crowd out live ones under budget pressure.

Live-row predicate (added to _postgres_private_sql.py):

AND superseded_by IS NULL
AND (valid_until IS NULL OR valid_until = '' OR valid_until::timestamptz > now())

valid_until is TEXT NOT NULL DEFAULT '' (migration 001), so a live row is the empty string — the issue's literal valid_until IS NULL OR valid_until > now() would have dropped every live row. Adjusted to the canonical form that mirrors MemoryEntry.is_temporally_valid.

How

• Predicate uses now()/IS NULL only — no bound params, so existing positional param tuples are unchanged and tenant/RLS predicates (project_id/agent_id) are untouched.
• Thread include_expired: bool = False through build_search_sql + new build_knn_search_sql, the sync + async backends, MemoryStore/QueryMixin, and the PrivateBackend protocol so the include_historical recall path still surfaces stale rows for the Python filter to decide.
• Suppress the predicate when as_of is set — point-in-time queries resolve validity via the bi-temporal valid_at/invalid_at window and legitimately want the superseded version. (This fixed a TestTemporalFiltering::test_search_as_of_returns_old_version regression the naive predicate introduced.)
• Python is_temporally_valid gate kept as defense-in-depth (unchanged).

Test

tests/integration/test_recall_temporal_sql.py (@pytest.mark.requires_postgres, ran — not skipped, against an isolated throwaway pgvector:pg17 container):

• Expired + superseded rows placed with a closer vector distance / higher FTS rank than the live row → asserted they do NOT occupy top-K while the live row does.
• Live-row (valid_until='') and future-valid_until rows still returned.
• include_expired=True escape hatch surfaces stale rows (KNN + FTS).
• as_of point-in-time still returns the superseded version.
• Cross-project rows never leak (tenant scoping intact) for both KNN and FTS.

Results: new test 10 passed; with the temporal unit tests 13 passed; existing recall integration 25 passed (no regression); affected-module unit sweep 512 passed / 26 skipped. ruff clean; mypy --strict src/tapps_brain/ → no issues in 152 files.

Closes TAP-4586.

🤖 Generated with Claude Code