Expose the Auth DB connection pool in config.toml (local Auth keeps no idle connections) #6981

Existing issues

• I have searched the existing issues.

Affected area

Auth

Problem to solve

The local Auth container keeps no idle Postgres connections. With GOTRUE_DB_MAX_IDLE_POOL_SIZE and GOTRUE_DB_CONN_PERCENTAGE unset, GoTrue's fixed pool strategy (internal/storage/dial.go) calls SetMaxIdleConns(0). Every Auth request then opens and closes its own database connection, and each closed socket sits in TIME_WAIT for 60s inside the Auth container.

Under a concurrent integration-test suite against supabase start, this exhausts the container's ephemeral ports (32768–60999). The symptom is random Auth failures, for example in user creation.

Repro, CLI 2.117.0, Auth image gotrue:v2.197.0: 20 password-grant requests sent from inside supabase_auth_<project> raised its TIME_WAIT sockets from 6 to 26, one per request.

The CLI gives no way to change this:

• The [auth] config has no pool key.
• The Auth env builder sets only GOTRUE_DB_DRIVER and GOTRUE_DB_DATABASE_URL, with no extra-env passthrough.
• supabase stop prunes containers, so an Auth container recreated by hand with extra env does not survive a restart.

Hosted Auth does get a pool sized as a percentage or an absolute number (supabase/supabase#50932), so local differs from production here.

Proposed solution

Either of these:

1. Expose the Auth DB pool in config.toml, for example:

   [auth]
   db_max_pool_size = 10
   db_max_idle_pool_size = 10
   db_conn_max_idle_time = "60s"

   These map to GOTRUE_DB_MAX_POOL_SIZE, GOTRUE_DB_MAX_IDLE_POOL_SIZE and GOTRUE_DB_CONN_MAX_IDLE_TIME.

2. Default the local Auth container to a non-zero idle pool, matching hosted.

Alternatives considered

• Recreating the Auth container with extra env after supabase start. The next supabase stop / start undoes it, and it means copying about 100 CLI-built env vars, labels and network settings by hand.
• Capping test concurrency. This works around the port exhaustion but does not remove it.

Additional context

Related: #6922, where the HTTP gateway opened a new upstream connection per request and exhausted ephemeral ports.