fix: Broken voice states and connectivity drops

toju-app/src/app/infrastructure/realtime/README.md

@@ -117,6 +117,12 @@ The signaling layer's only job is getting two peers to exchange SDP offers/answe
 
 Each signaling URL gets its own `SignalingManager` (one WebSocket each). `SignalingTransportHandler` picks the right socket based on which server the message is for. `ServerSignalingCoordinator` tracks which peers belong to which servers and which signaling URLs, so we know when it is safe to tear down a peer connection after leaving a server.
 
+Room affinity is authoritative at this layer as well. The renderer repairs each room's saved `sourceId` / `sourceUrl` from server-directory responses and routes `join_server`, `view_server`, and room-scoped signaling traffic to that room's signaling URL first. If that route fails, alternate endpoints can be tried temporarily, but server-scoped raw messages are no longer broadcast to every connected signaling manager when the route is unknown.
+
+Cold-start routing now waits for the initial server-directory health probes so same-backend aliases can collapse to one canonical signaling endpoint before any saved rooms reconnect. When a room is reconnected on a chosen socket, its background rooms are re-joined on that same socket as well so stale per-signal memberships do not keep orphan managers alive, and reconnect replay only sends `view_server` for rooms that manager still has joined.
+
+This is still a non-federated model. Different signaling servers do not share peer registries or relay WebRTC offers for each other, so users in the same room must converge on the same signaling endpoint to discover one another reliably.
+
 ```mermaid
 sequenceDiagram
     participant UI as App