I wasn't sure whether to label this a bug or an improvement.
For collections with stateFormat=2, we now fail requests (because of stale state) which we didn't previously. The client having stale cached cluster state is not a sufficient reason to fail and retry the entire request because in most of such cases, the node receiving the request is still perfectly capable of returning a valid response (either from local replicas or remote ones).
We should find a better way to notify clients that they have stale state. Perhaps we can modify the response and add a "routing" section instead of outright exception.