A libcluster strategy: nodes find each other through a Postgres channel.

Every :heartbeat ms, a node sends its name with pg_notify; every node LISTENs on the channel and connects to the names it hears.

Both Postgres connections reconnect by themselves, so a database outage only pauses discovery — it never crashes the node. (libcluster_postgres 0.2 crash-looped here and took the whole app down with it.)