Wouldn't the simple fix be to delay backend (= one connection) closing until all pending replication it initiated is finished? That still leaves actual crashes

Wouldn't the simple fix be to delay backend (= one connection) closing until all pending replication it initiated is finished? That still leaves actual crashes, which would need to use the shared memory to store the list of pending replications before the recovery of transactions is finished.