Reject stale Postgres readiness

[AntoineToussaint]

Summary

• probe the Nix Postgres process through its private Unix socket during bootstrap instead of trusting the public TCP port
• observe the launched process exit so bind failures surface immediately
• clean up partially initialized native runtimes on every Init failure
• prove lib/pq resolves the bootstrap DSN to the owned socket

Regression

A stale OrbStack Postgres proxy was listening on the proposed port. Before this change the Nix process logged a fatal bind failure but the agent pinged the stale database, ran migrations against it, and reported success. With the locally built agent, the same live collision now fails runtime-init immediately with postgres exited before becoming ready: exit status 1.

Verification

• codefly test source --target ./... --timeout 10m (173 passed)
• focused socket/credential tests (3 passed; final DSN assertion rerun passed)
• codefly agent build --dir ./service-postgres --native-only --skip-audit
• live Mind local-profile collision acceptance using the built Postgres 0.0.117 agent