Implement exact KNN query mode

[hevbot]

Summary

• Adds exact: true to /query and routes vector/hybrid vector legs through LanceDB VectorQuery::bypass_vector_index().
• Rejects exact+nprobes, FTS-only exact, and enabled semantic cache with exact mode.
• Keys the exact result cache distinctly for exact vs indexed requests.
• Adds MinIO-gated RFC 0014 integration coverage for indexed-vs-exact recall, selective filters, validation, and hybrid fuzzy FTS composition.

Closes #29.

Bench integration follows as #30.

Evidence

• ./scripts/cargo fmt --all -- --check ✅
• ./scripts/cargo clippy --workspace --tests --benches -- -D warnings ✅
• HEVSEARCH_S3_ENDPOINT=http://127.0.0.1:19000 ./scripts/cargo test -p hevsearch-api --test api_exact_knn -- --ignored --nocapture ✅

Exact vs indexed seeded PQ namespace (k=20, IVF_PQ num_partitions=16, num_sub_vectors=4, num_bits=4, indexed nprobes=1):

query_id	indexed recall@20
3	0.750
19	0.700
41	0.450
73	0.550
127	0.850
211	0.850
307	0.350
401	0.700

Full ignored workspace attempt:

• HEVSEARCH_S3_ENDPOINT=http://127.0.0.1:19000 ./scripts/cargo test --workspace -- --ignored --skip _aws --skip _100_runs_ --nocapture did not complete in this local container because linking was killed by the OS (collect2: fatal error: ld terminated with signal 9 [Killed]) while building object_cache_s3_smoke / hevsearch-core test binaries.
• Retried with -j 1; same linker OOM class, so targeted MinIO coverage above is the passing evidence for this change.

[hevbot]

Factory verified: RFC 0014 implemented per accepted spec — exact/hybrid routing through bypass_vector_index, exact+nprobes rejected, shared prefilter path covered, exact-vs-indexed recall evidence in body. Required check green in 3m57s (the #28 fix paying for itself). Merge-ready — merging also flips #30 (bench ground-truth tie-in) to ready.