Routes NOTIFY wake-ups from ergon_job_notifier to the workers that care.

A thin wrapper over OTP's pg, using the queue name as the group. Every ergon_worker joins its queue's group on init, and several workers draining the same queue coexist under one group, which is exactly pg's semantics. When a job lands, wake/1 fans a wake message out to them; they then race to drain, and FOR UPDATE SKIP LOCKED in checkout sorts out the contention exactly as it does on a periodic poll. A duplicate or spurious wake is therefore always harmless.

pg lives in kernel, so this costs no dependency.

The wake is deliberately node-local

wake/1 uses pg:get_local_members/2, not get_members/2, even though pg synchronises group membership across the whole cluster and the cluster-wide call would "work".

Every node runs its own notifier and receives the same NOTIFY, so waking remote workers means each worker is woken once per node in the cluster. That is harmless but pointless, and it quietly turns a local message send into cluster-wide chatter proportional to the square of the node count. Keeping it local also preserves the semantics a single-node deployment already has, so adding a second node changes throughput and nothing else.

Cluster-wide membership is still useful for observability, which is why pg is the right primitive rather than a private ETS table: pg:get_members/2 answers "who is draining this queue anywhere?" without any extra machinery.

This is the fast path only

If the notifier is disabled, or a wake is lost, workers still drain on their periodic fallback poll. The registry never being reached costs latency, never correctness.