After repeated close / open cycles for a database, we can get into the situation where the metadata checkpoint LSN is earlier than the first LSN in the log.

A partial fix was backed out in e6f15b1b3e (on the LevelDB branch). I think the real issue is that at the end of recovery, if there were modifications, we should do a real checkpoint rather than just logging a checkpoint. When recovery was first implemented, we were relying on all files being closed (and flushed) when the recovery session was closed. That is no longer true.

@sueloverso, can you please take a look at this one?