max_concurrent_users is leaking when leaving channels

[mandar1045]

Issue

When a client unsubscribes from a channel, RealtimeChannel.terminate/2 doesn't call UsersCounter.remove. This means the user stays in the census until the idle socket sweeper (Tracker) eventually kills the connection, which can take up to 10 minutes.

I found this while debugging an SPA where users frequently switch between channels. The connected user count kept increasing because users weren't being removed when they left a room, eventually causing "Too many connected users" errors.

Consequences

1. Unnecessary Postgres connections: Ghost users can keep tenant Postgres pools alive for up to 10 minutes, potentially wasting connections.
2. Customer friction: Stale users can consume max_concurrent_users slots, causing customers to hit their limits sooner and trigger rate limiting.

To Reproduce

1. Open a WebSocket without joining a channel — uses 0 user slots.
2. Join a channel — UsersCounter.add/2 is called, using 1 slot.
3. Call channel.unsubscribe() — RealtimeChannel.terminate/2 runs.
4. Bug: UsersCounter.remove is never called. The user has no active channels but still counts as 1 user for up to 10 minutes.

Proposed Fix

Remove the user from UsersCounter when their active channel count drops to 0 in terminate/2:

if Tracker.untrack(transport_pid) <= 0 do Realtime.UsersCounter.remove(transport_pid, tenant_id) end

[edgurgel]

Great issue report 🏆