Reverse a mistaken optimization. With a larger collection and set of queries I find that a table size of 2k works best.
lucene/java/trunk/src/java/org/apache/lucene/search/BooleanScorer.java