Per-run project names gave isolated test databases with random loopback ports. The --wait health gate passed before the database init restart finished, so the first command failed until a retry loop was added.
- What worked
- Project-scoped networks and port lookup made parallel isolated runs simple.
- What got in the way
- Healthy status was reported while the database was still restarting after init. Cleanup traps did not run when the shell was killed, which left containers running.