Add plain postgres ready and build connectivity waits into schema commands

[davegaeddert]

What this does

plain postgres ready — a serving gate for deployment entrypoints. A readiness engine (plain/postgres/readiness.py) classifies whether the database is ready for this code to serve, and the CLI maps it to an operator contract:

Exit	Meaning
0	Ready to serve
1	Not ready, retryable — migrations pending, schema missing objects, DB unreachable
2	Configuration error a human must fix — bad credentials, bad URL, missing database

--wait [--timeout N] standardizes the initContainer/entrypoint poll loop (config errors still exit 2 immediately); --json includes the exit code alongside the truthful classification.

Built-in connectivity waits for schema commands — sync, migrations apply, and converge now wait for the database themselves (up to POSTGRES_WAIT_TIMEOUT, default 60s), retrying connection failures while failing immediately on configuration errors via the same classifier. plain postgres wait is removed; plain-dev drops its wait step and relies on sync's built-in wait.

The resulting grammar: writers wait for connectivity, servers wait for readiness, one classifier decides retry-vs-refuse.

Design decisions

• The verb belongs to the operator. The engine returns a truthful classification; exit-code policy (including DEBUG downgrading schema gaps to warnings) lives in the CLI. JSON status and the Python API (check_database_ready()) never lie.
• Data-only migrations warn, not gate — a pending all-RunPython backfill shouldn't hold autoscaled capacity hostage while existing pods serve fine.
• Existence-only schema check — one read-only catalog query for model tables/columns (works against replicas). Drift is convergence's job; extra/unknown objects are invisible, which is what makes code rollbacks against a newer database pass by design (directional satisfaction — a contraction is the rollback horizon).
• Unknown connection failures default to retryable. psycopg3 exposes no sqlstate on connection-time failures, so classification uses a small marker set for clearly-permanent errors; a DB restarting mid-deploy must never read as a config problem.
• Direct probe connections, never the pool — pool timeouts are classification-opaque. The engine installs its connection as the active connection so migration-recorder ORM reads ride it too.

Also: float-annotated settings now accept int values (PEP 484 numeric tower) — PLAIN_POSTGRES_WAIT_TIMEOUT=30 and X = 30 in settings.py work naturally instead of failing type validation.

Testing

• 766 plain-postgres + 433 plain + 26 plain-dev tests pass; type checks clean across all three packages.
• New public tests pin the classification contract and CLI exit codes; internal tests pin the connection-failure taxonomy and the wait loop.
• Live-verified against real postgres: wait-through-outage recovery, bad password/scheme/option → exit 2, DEBUG warn-through, JSON reconciliation, --timeout bounds.