Parallel bulk insert (AddEntries) and EcoQoS opt-out in benchmarks

[matengo]

Summary

Adds QvecDatabase.AddEntries(IReadOnlyList<QvecInsert>, int maxDegreeOfParallelism = 0) — multi-threaded HNSW index construction — and fixes a measurement problem that had been silently throttling every benchmark so far.

Parallel insert

• Chunks of 8,192 entries: phase A serial (slot, layer, vector/metadata/guid to disk, guid index), phase B Parallel.For linking nodes with hnswlib-style striped node locks (1,024 stripes, never nested) and a reader/writer lock around the entry point.
• The whole call holds the database write lock, so concurrent readers see either the old or the new state; nothing changes for AddEntry.
• dop = 1 produces a byte-identical graph to a loop of AddEntry (tested via RawGraph).
• Same semantics as AddEntry: ids returned in input order, duplicate ExternalIds skipped (within a batch and against existing data), dimension mismatch throws QvecDimensionException.
• Design doc: docs/design-insert-parallel.md.

Measured (same machine, back to back, throttling disabled)

Dataset	1 thread	12 threads	Speed-up	Recall
Cohere 100K float, M = 32	370.9 s (270/s)	58.6 s (1,707/s)	6.3×	recall@100 −0.09 pp @ ef 100
SIFT-1M, M = 32	1,145.6 s (873/s)	142.1 s (7,037/s)	8.1×	recall@10 within 0.1 pp at all ef

Trade-off, stated in README: parallel builds are not byte-reproducible (layers are seeded, link order is not). Scaling is sub-linear because ~20 % of thread time on Cohere is spent waiting on hub-node locks; the rest is memory latency.

EcoQoS discovery

12 threads initially got only ~4 cores' worth of CPU time with no kernel time, GC pauses or lock contention to explain it. Root cause: Windows 11 (Balanced plan) applies EcoQoS power throttling to console processes that are not in the foreground. The benchmark now opts out via SetProcessInformation(ProcessPowerThrottling) at startup (--allow-throttling restores the old behaviour). Serial Cohere 100K went from 589 s throttled to 371 s exempt, which also explains the historic ±40 % QPS variance. All earlier README figures were measured throttled; SIFT-1M and Cohere 100K tables are re-measured and marked as such.

Benchmark

--threads N (0 = all cores) builds through AddEntries in batches of 50,000. Report line shows build threads.

Tests

8 new ParallelInsertTests: stores every entry & recall vs serial (float + int8), byte identity at dop = 1, well-formed neighbour lists with bounded layer-0 orphan count vs serial, duplicates within batch and across calls, growth + deleted-slot reuse, empty batch no-op, wrong dimension throws.

Fast suite 333 tests + Slow suite 7 green locally; build 0 warnings.

Not in scope

• PartitionedQvecDatabase does not get AddEntries yet.
• Lock-free back-linking to recover the hub-lock wait.