Fix latest_output runtime bug and run DB-backed tests in CI

[seanwevans]

Summary

CI ran pytest without a PostgreSQL service, so every database-backed test
silently skipped (10 of 35). That blind spot let two real defects ship to
main. This PR makes CI actually exercise the database, then fixes everything
the now-running tests surface.

Root cause

The test job had no Postgres service and never set TEST_DATABASE_URL, so the
DB fixtures hit pytest.skip(...). Locally, running the suite against a real
Postgres 16 produced 4 failures.

Changes

Production bug

• sql/latest_output.sql: the function failed at runtime with
  column reference "id" is ambiguous — the RETURNS TABLE output columns
  (id, command, …) collide with the selected commands columns inside the
  RETURN QUERY. The SPEC's core read endpoint was completely non-functional.
  Fixed by qualifying every column with a table alias (c.id, c.command, …).

Test correctness

• tests/test_cleanup_agent.py: test_cleanup_once_resets_multiple_envs
  inserted two users with the same username (violating the UNIQUE constraint)
  and compared uuid = text in its verification query. Now uses distinct
  usernames and casts the array parameter (ANY(%s::uuid[])).
• tests/test_functions.py: test_command_indexes_query_plans never ran
  ANALYZE, so the planner chose an index from empty-table statistics
  (commands_status_completed_at_idx instead of the expected
  commands_status_submitted_at_idx). Added ANALYZE commands before
  EXPLAIN; the planner then picks the correct index.

CI

• .github/workflows/ci.yml: added a postgres:16 service (with a
  pg_isready healthcheck) and TEST_DATABASE_URL so the database tests run on
  every push and pull request.

Verification

• Against a live Postgres 16: 35 passed (was 4 failed / 31 passed).
• Without a database (mirrors the old CI): 25 passed, 10 skipped — the suite
  still degrades gracefully.

🤖 Generated with Claude Code

---

Generated by Claude Code