GPU search hardcodes query batch size 1 — no multi-query batching

Summary

cuvs-lucene always passes a 1 × dim query matrix to cuVS. CagraSearchParams.maxQueries is never set. cuVS supports batched search natively; the Lucene layer does not expose it.

Reproducer

From a cuvs-lucene checkout:

rg 'deviceBuilder\([^,]+, 1,' src/main/java/com/nvidia/cuvs/lucene/

Both GPU search paths hardcode batch size 1:

• GPUKnnFloatVectorQuery.java (~line 195) — multi-segment path
• CuVS2510GPUVectorsReader.java (~line 458) — per-segment fallback

Neither calls CagraSearchParams.Builder.withMaxQueries(n) for n > 1.

Runtime check: run any GPUKnnFloatVectorQuery search; one cuVS cagraSearch call is issued per Lucene query. MultiPartitionCagraSearch batches segments, not query vectors — still one query per call.

Expected vs actual

• Expected: option to search N query vectors in one GPU call (as cuVS Java API allows via N-row CuVSMatrix + maxQueries).
• Actual: one query vector per search; throughput scaling in benchmarks uses queryThreads (N concurrent batch-1 searches).

Workaround

Use multiple search threads (queryThreads > 1). Not equivalent to in-kernel query batching.

Environment

cuvs-lucene 25.10+, CuVS2510GPUSearchCodec, GPUKnnFloatVectorQuery.