Why BM25 queries with more terms can be faster (and other scaling surprises)