Cohere 1M measurement and fix for k=100 concurrent-search scaling

[matengo]

Summary

Measures Cohere 1M in the configuration Zvec publishes (int8 / float, M = 15, efSearch = 180, recall@100, 12 concurrent clients), and fixes a scaling bottleneck that the measurement exposed.

Bug: result materialisation serialised concurrent searches

At k = 100 with 12 query threads, QPS was not monotonic in efSearch: efSearch 100 gave 1,883 QPS, efSearch 180 gave 3,129. Single-threaded and k = 10 sweeps were monotonic, so the per-hit work under concurrency was the suspect.

GetMetadata and ReadGuidFromDisk read through MemoryMappedViewAccessor, whose every ReadX/ReadArray call takes an interlocked reference on the shared SafeBuffer. At k = 100 that is a few hundred atomic operations per query on one cache line; twelve threads serialised on it. Both now read through the raw mapping pointer the distance computations already use.

Cohere 1M float, efSearch 100, 12 threads: 1,883 → 6,441 QPS. Single-threaded throughput unchanged (so the SIFT-1M numbers in the README did not move).

Harness

• --passes N: repeat the query set N times per efSearch row. Cohere has only 1,000 queries; at several thousand QPS a row was over in under a second.
• Warm-up now uses the measurement's thread count and runs ≥ 1 full pass and ≥ 2 s. A fresh process soft-faults every page of a 3.4 GB mapping into its working set even when the file is OS-cached; the first row of a sweep was 2–3× slower than the second before this.

Measured (12 cores, throttling exemption, 12 build threads, --concurrency 12 --passes 10)

mode	build	file	efSearch	recall@100	QPS 1 thread	QPS 12 threads
float	419 s (2,385/s)	3,376 MiB	100	90.3 %	—	6,441
float			180	94.8 %	565	3,764
float			320	97.4 %	351	2,207
int8	181 s (5,516/s)	1,194 MiB	100	89.3 %	1,395	13,540
int8			180	93.2 %	839	6,841
int8			320	95.3 %	502	4,714

Zvec's chart for the same configuration on 16 vCPUs reads as roughly 8–9 k QPS at recall@100 ≈ 0.93–0.94 (image only, approximate). Per core the int8 row is in the same range — on different hardware, OS and day, which is as far as the comparison goes. The documentation says so explicitly. Note that Zvec's 1M run does not use its refiner (only 10M does); the earlier README sentence claiming otherwise is corrected.

Tests

Fast suite 333 + Slow 7 green, 0 warnings. The read-path change is covered by the existing metadata/Guid round-trip tests; no new test since the defect is a contention profile rather than a functional bug.