perf(vector/search): ahash for the pool doc_id->position maps probed per candidate in Flat/IVF searchers

[mosuka]

Background

#853 (PR #854) removed every per-candidate hash probe from the HNSW graph traversal by moving it to segment-local ordinals. The Phase 0b prototype gate for that work measured the probe cost directly: swapping the pool doc_id → position maps from std HashMap (SipHash) to ahash::AHashMap improved the HNSW store-level deletion benches by −30/−32% before any structural change (gate record).

The Flat and IVF searchers were out of #853's scope and still pay that SipHash probe per scanned candidate:

• flat/searcher.rs — the int8 hot path resolves pool.field_position_index(field_name) once, then probes the inner HashMap<u64, u32> per candidate inside the scan closure.
• ivf/searcher.rs — parallel_scan over the probed clusters calls pool.get_record(*doc_id, field_name) per candidate, which probes the same map.
• Map definitions (inner maps are std HashMap<u64, u32>): quantized_storage.rs, pq_storage.rs, pq_fastscan_storage.rs, rerank_storage.rs (field_index).

ahash is already a workspace dependency, and post-#853 the HNSW hot path no longer touches these maps at all — so the swap is low-risk and its benefit is confined to the Flat/IVF scan paths and cold lookups.

What to do

Swap the inner HashMap<u64, u32> (and the reader-side HashMap<u64, usize> prefetch map if convenient) to ahash::AHashMap in the four pool types, following the type through the few construction/return sites.

Acceptance criteria

• A/B on the store-level Flat/IVF benches (Flat Search, Flat Search int8 kernel, IVF Search, Flat/IVF Filtered Allowset), with the usual A/A null-pair discipline
• No regression on the HNSW gate cases (Deletion Search, HNSW Graph Search sparse_ids) — expected neutral since perf(vector/hnsw): segment-local u32 ordinal graph + LVS1 v2 format (sub-issue of #686) #853 removed those probes
• fmt / clippy (incl. 1.97 parity) / tests clean

References

• perf(vector/hnsw): segment-local u32 ordinal graph + LVS1 v2 format (sub-issue of #686) #853 Phase 0b gate measurement (the −30% prototype): the swap recipe and the boundary-conversion caveats are described there
• perf(vector): migrate to InternalId = u32 throughout vector index + search #686 closing status comment (this issue is the "small follow-up candidate" recorded there)

Task list

• A-side baseline (main): 2 rounds, A/A bands recorded
• Implemented as specified (then reverted per the NO-GO verdict)
• fmt / clippy (incl. 1.97 parity) / tests / wasm check — all clean pre-revert
• B-side benches → NO-GO: neutral everywhere (see closing comment) → reverted
• [~] PR — not opened (nothing to merge; closed as not planned)

[mosuka]

Measured: NO-GO — closing with the negative result

Implemented as specified (inner HashMap<u64,u32> of the four pools' field_index, the HNSW searcher ctx types, and the reader prefetch map → ahash::AHashMap; fmt/clippy 1.95+1.97/tests/wasm all clean) and A/B-measured against main (A = 2 runs, B = 1 run, worst-pairing):

case	A (min..max)	B	Δ vs worst	A/A band
Flat Search int8 kernel/cosine_dot	2069.5..2085.5µs	2042.8µs	−1.3%	0.8%
Flat Search int8 kernel/euclidean_sqdiff	2140.5..2381.9µs	2116.9µs	−1.1%	11.3%
Flat Search int8 kernel/manhattan_absdiff	2014.8..2027.7µs	2014.0µs	−0.0%	0.6%
IVF Search Large K/512	93.2..94.4µs	91.2µs	−2.2%	1.2%
Deletion Search del0/del10 (HNSW-neutrality guard)	—	—	+0.5% / −3.6%	✓ as predicted
Flat Filtered / Multi-field	—	—	±2% mixed, one +16.5% outlier at n=1000 (codegen-noise class)

Why the premise didn't transfer: the per-candidate idx.get(&doc_id) probe does exist in the flat scan (flat/searcher.rs), but that scan is a parallelised, memory-bandwidth-bound sequential sweep (parallel_scan + 128-byte int8 records streamed per candidate) — the ~20ns SipHash probe is ≲5% of per-candidate CPU and hides entirely behind bandwidth. The −30% Phase 0b measurement in #853 came from the serial, latency-bound HNSW pointer-chase, where the probe sat on the critical path; that path no longer uses these maps at all (ordinals), which the deletion-bench neutrality above confirms.

Takeaway recorded in the maintainers' notes: a probe-cost lower bound measured on a latency-bound serial walk does not transfer to a bandwidth-bound parallel scan, even over the identical data structure — the access regime, not the structure, decides (the same regime split as #679's contiguity lesson, in the other direction).

Change reverted, nothing merged.