So it's true that the client can't unilaterally terminate COPY BOTH mode; it can only send CopyDone. But an error on the server side should do so.
Hm, you're right. Even though the server sends COPY_BOTH message before the plugin startup, an attempt by the client to actually read the data messages results in ErrorResponse handled on the client, then the client can re-submit the corrected START_REPLICATION command and continue without the need to reconnect. This cannot be actually tested in psql, but I can verify the behavior with my python scripts.
Then I don't have a preference for the early error reporting in this case. If the current behavior potentially allows for more flexible error reporting, I'm for keeping it.
It looks like a decision needs to be made here whether to apply this patch, send it back to the author, or reject it so I'm marking it "Ready for Committer".
Robert, since you were participating in this conversation can you have a look?