If we drop an event (or otherwise end up with a gap) when processing a response to a /backfill request we'll subsequently try always try and /backfill at the gap whenever a client paginates past this point. This is a bit sucky and slows things down a lot.

Potential solutions include:

• try to backfill a few times and then mark that backwards extremity such that we don't try again
• for v2 rooms and above we could process just the prev_events, for example if we have events C -> B -> A, where B is a dodgy event, then we can still note that B fills a gap between C and A and so there is no longer a gap there. (This is safe due to event ID as hashes)