Dennis, would it be possible to isolate some of the cases with more extreme results and run them repeatedly under the same timing framework, as a test of how trustworthy the _framework_ is? From decades of bitter experience, most benchmarking efforts end up chasing ghosts ;-)

For example, this result:

length=3442, value=ASXABCDHAB...  | 289 us  | 2.36 ms: 8.19x slower (+719%) 

Is that real, or an illusion?

Since the alphabet has only 26 letters, it's all but certain that a needle that long has more than one instance of every letter. So the status quo's "Bloom filter" will have every relevant bit set, rendering its _most_ effective speedup trick useless. That makes it hard (but not impossible) to imagine how it ends up being so much faster than a method with more powerful analysis to exploit.