Design a messaging protocol between GPU workers

GPU workers in one cluster send small control messages straight to named peers. Design the messaging layer: sessions and sequence numbers per direction so each message lands in the receiver's queue once and in order despite retries and lost acknowledgements; what a receipt does and doesn't promise; what a sender learns after a timeout (an unknown outcome, not a failure); detecting that a peer restarted so stale messages from the old session are dropped; when a session can resume after a reconnect; and limits on queued bytes and messages, with rejection when they are full.