[davecromberge]

Replaces SipHash with a seeded folded multiply for the hash set that ThetaANotB and TupleANotB build from B when either input is unordered.

• The keys are already uniformly distributed 64-bit hashes, so SipHash's cost buys nothing here. A per-call random seed keeps crafted sketches from forcing collisions; an unseeded multiply let a 65,536-entry B take 212 ms instead of 0.25 ms.
• Ordered inputs use the existing merge path and are unchanged.

Speedup vs main from the characterization *_set_operations_speed_profile jobs (update-sketch inputs, ordered result; geometric mean over each band, best of 3 interleaved passes). Union, intersection and Jaccard in the same runs stay at 1.00×.

InU	Theta A-not-B	Tuple A-not-B
1–64	1.18×	1.16×
65–4,096	1.30×	1.33×
4,097–65,536	1.42×	1.81×
65,537–1,048,576	1.45×	1.83×

🤖 Generated with Claude Code