Return 503 when the database drops a request's connection #905

Summary

When Postgres ends a session, for example after the API's 60-second idle-in-transaction limit or during a restart or failover, the next query fails and the API returns a 500 and sends a Slack alert. The client can't tell that retrying will work. This issue returns 503 with Retry-After for a dropped connection instead, and keeps the Slack alert.

Steps to reproduce

1. Open a session with SET idle_in_transaction_session_timeout = '500ms'.
2. Run a query, wait 1.5 seconds, run another.

Observed

sqlalchemy.exc.OperationalError: server closed the connection unexpectedly, with no SQLSTATE (err.orig.pgcode is None) and err.connection_invalidated set. CatchAllErrorMiddleware returns 500.

Fix

• In lib/middleware/errors.py, treat a DBAPIError with connection_invalidated as a 503 with Retry-After: 5.
• Unlike the database timeouts, keep sending the Slack alert: a dropped connection can also mean an outage.

Acceptance criteria

• A route whose connection is closed by the server returns 503 with Retry-After.
• The Slack alert still fires for it.
• Other database errors still return 500.