[davecromberge]

Speeds up ThetaSketch::compact by gathering retained hashes without a per-slot branch and bucket-sorting them for ordered output; the compact bytes are unchanged.

• Hashes are uniform below theta, so a counting pass on their top bits followed by insertion sort replaces sort_unstable; a bucket over 32 entries or fewer than 256 hashes falls back to sort_unstable.
• Tables at most 1/16 full keep the simple loop, so tiny sketches are not slower.

Speedup vs main from the characterization theta_compact_serde_profile job (lg_k 12; geometric mean over each band, best of 3 interleaved passes):

TrueU	compact(true)	compact(false)
1–64	1.67×	3.1×
65–4,096	2.1×	5.8×
4,097–1,048,576	1.70×	3.9×

Ordered speedup split between the two parts (separate interleaved run including a gather-only build):

TrueU	Gather only	Gather + bucket sort
1–64	1.63×	1.59×
65–4,096	1.62×	2.01×
4,097–1,048,576	1.28×	1.67×

Unordered compaction gets its whole gain from the gather.

🤖 Generated with Claude Code