Deeper fixes #1: FK indexes, session-factory invalidation, HTTP pooling

[mousebrains]

Stacked on top of #3 (cleanup-and-fixes). Set the base back to main once #3 merges.

Summary

Second batch from the deep-review pass. Three small targeted fixes — each one a real correctness or scaling concern surfaced by the original review.

1. FK indexes (migration 0013)

Four FK columns weren't covered by an existing index:

• gauge_source.source_id — composite PK only indexes (gauge_id, source_id), so reverse-direction lookups full-scanned.
• reach_guidebook.guidebook_id — same composite-PK reason.
• reach_class.reach_id — non-PK FK, no covering index.
• latest_gauge_observation.source_id — non-PK cache FK.

Tables are small today (224 / 1093 / 401 / 461 rows), so latency is fine; the indexes remove the future scaling cliff and match the ix_reach_state_state_id pattern already used for the same reason. Migration applied cleanly on a copy of the prod DB; foreign_key_check + integrity_check both ok; init-db on a fresh DB produces matching indexes.

2. Session factory invalidation (db/engine.py)

get_engine(url=...) disposed the old engine and created a new one, but the cached _session_factory was only rebuilt when get_session_factory() itself was called with a url. A caller going through get_engine() directly left the factory bound to the now-disposed engine — sessions created from it would either fail or write to a stale connection pool kept alive by GC.

Fix: invalidate _session_factory at the same time the old engine is disposed. Regression test exercises the exact path.

3. HTTP connection pooling (utils/http_client.py)

A typical levels fetch run hits ~50 URLs across a handful of hosts. Calling requests.get() opened a fresh TCP+TLS connection per call. Hoisted a module-level requests.Session (lazy via _get_session()) so the second call to a host reuses the pool. reset_session() for test isolation. fetch() now routes through _get_session().get(...).

Test fixtures simplified: per-method @patch("...requests.get") collapsed into one autouse fixture. New TestSessionPooling covers singleton-ness, User-Agent priming, reset behaviour, and shared-session usage.

Test plan

• make test — 579 passed, 1 deselected (was 574)
• ruff check and ruff format --check clean
• mypy src/ clean
• Migration 0013 applies on cp ../DB/kayak.db /tmp/kayak-mig13.db; all four indexes present; foreign_key_check + integrity_check both ok
• init-db on a fresh DB produces all four indexes
• CI on this PR (3.13 + 3.14 + gitleaks + lint + typecheck)

🤖 Generated with Claude Code