If you run incremental repair regularly (every night or so), the data in the unrepaired set of sstables will (almost) always be newer than the data in the repaired sstables - this means that we will not block any compactions among the repaired sstables from dropping tombstones based on the sstables in the unrepaired set (ie, a tombstone in the repaired set will always be older than the data in the unrepaired set, and therefor cannot cover any data there). And since data will move quickly from unrepaired to repaired, this is not a problem in the unrepaired set either.