With a very common 'must' token, the 'should score' wasn't taken into account
enough, and on small enough `limit` the best result wasn't returned.