bench: investigate synthetic benchmark deltas after 0.5.0 release

[Fieldnote-Echo]

Summary

After the 0.5.0 crate release, investigate the synthetic bench_rank deltas against the committed README baseline. This is curious and worth understanding, but it is not a 0.5.0 release blocker.

A 10-run aggregate on the current 0.5.0-pre branch shows several large speedups that are plausible given recent kernel and two-stage optimization work, plus a few slower rows that should be explained before refreshing public synthetic claims.

Evidence

Command:

cargo run --release --example bench_rank

Method used for this note:

• baseline: committed README / benchmarks/rank_modes_results.txt synthetic table
• current: 10 consecutive runs on the current release-prep branch
• machine: same local x86_64 AVX-512 desktop class as the existing synthetic fixture
• quality columns are deterministic for this seeded synthetic corpus; latency and throughput are wall-clock measurements

Headline current 10-run p50 aggregate versus README baseline:

Mode	Current p50 mean +/- sd ms	Current CV	README p50 ms	Delta vs README	Current R@10
Rank asym	3.6400 +/- 0.0748	2.1%	3.7118	-1.9%	0.8330
RankQuant b=4 asym	0.3258 +/- 0.0071	2.2%	0.3132	+4.0%	0.8095
RankQuant b=2 asym	0.2640 +/- 0.0082	3.1%	0.2382	+10.8%	0.5785
RankQuant b=2 FastScan	0.1086 +/- 0.0024	2.2%	0.0901	+20.6%	0.5845
RankQuant b=2 byte-LUT	0.6212 +/- 0.0168	2.7%	0.7542	-17.6%	0.5785
RankQuant b=4 byte-LUT	1.2246 +/- 0.0314	2.6%	1.6437	-25.5%	0.8095
Bitmap n_top=64	0.0634 +/- 0.0022	3.5%	0.0812	-21.9%	0.2495
SignBitmap probe	0.0498 +/- 0.0015	3.0%	0.0912	-45.4%	0.2745
TwoStage b=2 M=100	0.0538 +/- 0.0012	2.3%	0.0977	-44.9%	0.5795
TwoStage b=2 M=500	0.0675 +/- 0.0024	3.5%	0.1089	-38.0%	0.5785
TwoStage b=2 M=1000	0.0804 +/- 0.0029	3.6%	0.1225	-34.4%	0.5785
TwoStage b=2 M=5000	0.1884 +/- 0.0073	3.8%	0.2398	-21.4%	0.5785
SignTwoStage b=2 M=500	0.0676 +/- 0.0022	3.2%	0.1057	-36.0%	0.5785

Initial read

The large speedups in bitmap/sign/two-stage and byte-LUT rows are likely explained by recent kernel and caller-owned/two-stage optimization work. The b=2 asym and FastScan slowdowns are larger than the observed run-to-run CV, so they deserve a focused follow-up rather than being dismissed as jitter.

Suggested investigation after 0.5.0

• Make the built-in synthetic benchmark optionally run and emit N-sample aggregates, probably with a flag such as --samples 10.
• Preserve single-run output for quick local smoke runs.
• Compare current main against the README baseline and, where useful, against nearby historical commits around the kernel/two-stage changes.
• Attribute deltas to specific changes where possible: byte-LUT scoring, bitmap/sign probe kernels, two-stage candidate generation, exact rerank, FastScan layout/scoring, and RNG/corpus/ground-truth changes.
• Refresh benchmarks/rank_modes_results.txt and README synthetic rows only after the investigation separates expected kernel wins from benchmark harness or corpus drift.

Non-goals

• Do not block the 0.5.0 release on this.
• Do not update public performance claims from one-off local timings.
• Do not change BEIR/real-corpus claims as part of this issue; use the BEIR harness for those.