Bench exact KNN ground truth

[hevbot]

gate:impl

Closes #30

Summary

• Change hevsearch-bench recall sweep to collect ground truth with in-engine exact: true queries on the same namespace used for indexed queries.
• Score indexed ivf_pq recall/ndcg against those exact engine row IDs instead of external .ivecs or in-process brute force.
• Leave the filtered ground-truth strategy as an explicit hev/search#25 seam; this implementation runs the unfiltered query set first.

Evidence

Smoke bench on isolated MinIO (search-bench-gt-minio, host port 19002) with seeded synthetic namespace:

HEVSEARCH_STORAGE_URI=s3://search-bench-gt \
HEVSEARCH_S3_ENDPOINT=http://127.0.0.1:19002 \
HEVSEARCH_S3_ACCESS_KEY=minioadmin \
HEVSEARCH_S3_SECRET_KEY=minioadmin \
HEVSEARCH_BENCH_ROWS=2000 \
HEVSEARCH_BENCH_DIM=16 \
HEVSEARCH_BENCH_QUERIES=5 \
HEVSEARCH_BENCH_PARTITIONS=8 \
HEVSEARCH_BENCH_SUB_VECTORS=1 \
HEVSEARCH_BENCH_NPROBES_SWEEP=1,4 \
HEVSEARCH_BENCH_OUT_DIR=bench/results/recall-smoke \
./scripts/cargo run -p hevsearch-bench --bin recall_sweep

Output included:

exact reference (unfiltered) qps=78.1 p50=5.4ms p95=42.7ms
unfiltered nprobes=   1  recall@1=0.2000 @10=0.2600 @100=0.3180  ndcg@10=0.3038
unfiltered nprobes=   4  recall@1=0.2000 @10=0.3600 @100=0.4740  ndcg@10=0.3791

Generated JSON recorded ground_truth_source: "in_engine_exact", same_namespace: "recall-ivfpq-1783554566869249552", and each indexed point had scored_against: "in_engine_exact".

Verification

• ./scripts/cargo fmt --all --check
• ./scripts/cargo check -p hevsearch-bench --bin recall_sweep
• ./scripts/cargo test -p hevsearch-bench recall --lib
• ./scripts/cargo clippy -p hevsearch-bench --bin recall_sweep -- -D warnings

[hevbot]

Factory verified: bench scores indexed recall against the in-engine exact reference (RFC 0011↔0014 target state), #25 strategy seam left open as required, evidence run in body. Merge-ready pending checks.