This is the same as this and this.
The dgraph team tried to reproduce it on dgraph cloud and stated they could not.
That peer is corrupt now - to remove him and add a new peer: run /removeNode on the zero specifying the messed up peer, wipe out his state and bring him back with new state. Instructions are here (they are k8s specific but basically do as above).
This corruption issue has hit us a lot and is part of our bi-weekly issues with running dgraph ourselves. Removing nodes from a group also, in our case, caused them to stay in the raft membership and messed up our cluster which required forking dgraph to fix, but the dgraph team has never even responded to that.