chore(vector/bench): resolve the mmap/OnDemand HNSW 50k bench class verdict after #853

[mosuka]

Background

During #853's A/B (PR #854), the mmap/OnDemand f32 HNSW benches at n=50000 could not produce a verdict on this bench host:

• HNSW Graph Search/top10/50000 measured +17% vs main, while HNSW Graph Search Rerank/top10/50000 — which runs the same traversal plus a rerank tail — measured −0.2% in one branch run and +20.4% in another.
• The +20.4% run executed a strictly lighter binary than the −0.2% run (it skips an entire prefetch pass on this path), which is direct evidence the case cannot resolve ±20% on this host.
• The untouched HNSW Fallback Search benches (code not modified by perf(vector/hnsw): segment-local u32 ordinal graph + LVS1 v2 format (sub-issue of #686) #853) swung −6.2%..+9.6% across the same sessions.

Full tables: the final A/B comment on #853. The store-level eager-int8 benches (the actual production VectorStore path) resolved cleanly at −56..−86%, so this is a measurement-quality question, not a known regression.

What to do

Observe whether a consistent post-#853 delta exists on the OnDemand/mmap path at 50k, using measurement the local host cannot provide:

• pinned methodology (./scripts/bench-stable.sh --bench vector_search_bench -- "HNSW Graph Search/top10/50000"), and/or a second host / CI runner
• interleaved A/B/A/B with medians, A/A null pair first

If a consistent regression is confirmed, file a targeted fix issue — candidate mechanisms to check: the graph.doc_id(ord) array read added to the f32 fallback distance path, and graph memory layout after v1→ordinal load translation.

Acceptance criteria

• Pinned measurement done (bench-stable.sh, worktree A/B, shared corpus, interleaved; 5 A runs + 3 B runs)
• Verdict: within noise — no consistent regression (+1.2–3.8% on tight pairs; the flagged +17–22% was host interference, proven by untouched-code spikes). Closed, no fix issue needed

References

• perf(vector/hnsw): segment-local u32 ordinal graph + LVS1 v2 format (sub-issue of #686) #853 final A/B comment (evidence tables), perf(vector): migrate to InternalId = u32 throughout vector index + search #686 closing status comment

[mosuka]

Verdict: no consistent regression — within noise. Closing.

Measured with the instrument the original flag called for: scripts/bench-stable.sh (CPU-pinned + nice), A = pre-#854 (f3b0409a) in a git worktree, B = post-#854 main, identical corpus files (shared bench cache), interleaved runs. mmap relevance confirmed first: FileStorageConfig defaults to use_mmap = true on Linux (platform::default_use_mmap()), so the OnDemand path is production-relevant and this verdict mattered.

HNSW Graph Search/top10/50000 (pinned, medians)

run	A (pre-#854)	B (post-#854)
samples	446.4 / 500.9 / 501.9 / 502.8 / 504.9 µs	507.9 / 521.1 µs (+ one 637.0 µs spike)
tight-cluster verdict	med 501.9 µs	+1.2% / +3.8% vs A median

Cross-checks

• HNSW Graph Search Rerank/top10/50000 — the only case where BOTH sides measured tight in the same window (A/A 0.6%, B/B 0.6%): +1.3%.
• HNSW Fallback Search/top10/50000 — code untouched by perf(vector/hnsw): segment-local u32 ordinal graph + LVS1 v2 format (#853) #854 — spiked +11.9% between two B-side runs in the same session; the plain case's 637 µs spike (B/B 22%) occurred in a different run's window. Interference hits untouched code equally → host-side (intermittent background load), not the change.
• The A side itself spans 446–505 µs (13%) across pinned sessions.

Conclusion

The +17–22% flagged during #853's A/B was host interference, not a code regression. The best available measurements put the post-#854 mmap/OnDemand 50k path at +1.2–3.8% — at the edge of even pinned resolution on this host, dwarfed by the −56..−86% wins on the eager int8 path, and with a known bounded candidate (the one extra graph.doc_id(ord) array read in the f32 fallback) should a better instrument ever show it to be real. No fix issue warranted.