Platform saves previous state after reconnect

I’ve recently observed an interesting test failure, caused by the following sequence of events:

1) Round `X` reaches consensus, and we know we want to write it to disk eventually
2) Before we gather enough signatures on round `X`, we fall behind
3) When we eventually reconnect, the current round is way ahead of `X`
4) The node decides (correctly) that it will never be able to gather enough signatures on round `X`, and writes the state from round X to disk anyways.

This is not ideal. Will update this ticket with the proposed changes, once those details have been hashed out.