Fix scalar-quantized CAGRA graph input encoding

[nvzm123]

Summary

This does not add scalar quantization: the accelerated scalar codec, service registration, and benchmark selector already exist. It fixes the unsigned byte encoding supplied to the intermediate CAGRA graph build.

• Initialize per-dimension maxima with negative infinity so all-negative dimensions are measured correctly.
• Emit monotonic unsigned 7-bit values in [0, 127], matching the unsigned CuVSMatrix.DataType.BYTE contract.
• Remove the ineffective signed-to-unsigned conversion and its redundant full byte-vector copy.
• Add focused regressions for negative-only dimensions and monotonic values across zero.
• Regenerate the affected Lucene API reference pages.

Why

The old helper emitted signed values in [-64, 63]. Its conversion used (byte) (value & 0xff), which is a bitwise no-op after casting the result back to a Java byte. Because the cuVS byte matrix is exposed as kDLUInt, negative codes were read as 192..255 while zero remained 0, introducing a discontinuity into CAGRA distances.

The direct unsigned formula is the intended centered code plus 64:

(round(127 * normalized) - 64) + 64 == round(127 * normalized)

For Euclidean distance, that constant shift preserves the intended pairwise distances while satisfying the unsigned matrix contract. Float.MIN_VALUE was also replaced because it is the smallest positive float, not a negative lower bound.

Usage

No benchmark-harness or YAML API change is required. The existing benchmark selector is:

"algoToRun": "CAGRA_HNSW_SCALAR"

This branch is based directly on the current NVIDIA/cuvs#2476 head, 8faae5092.

Compatibility and scope

• Lucene's persisted/search scalar quantization and on-disk format are unchanged.
• Existing indexes are unchanged; newly built scalar indexes may have different graph topology because the intermediate CAGRA input is corrected.
• The public helper's byte output intentionally changes from centered signed codes to unsigned 7-bit codes.
• This PR contains no post-ingest parallelism, graph-only persistence, Lucene 10.5 work, heuristic changes, or benchmark-harness changes.
• Benchmark quality has been validated only for Euclidean distance. Adding 64 is not inner-product or cosine invariant, so this PR does not claim quality correctness for those metrics.

Validation

Exact branch head:

• Focused scalar, quantized-format, and multi-layer round-trip tests: 111 tests, 0 failures, 0 errors, 14 intentional skips.
• Full cuvs-lucene module: 346 tests across 33 suites, 0 failures, 0 errors, 30 skips.
• ./build.sh java lucene --run-java-tests completed successfully, including Panama binding generation and local artifact installation.
• mvn -q spotless:check and git diff --check passed.
• API-reference generation was idempotent across consecutive runs.
• fern/build_docs.sh check validated all 284 MDX files with 0 errors. Its two warnings were the existing unauthenticated redirect check and light-mode accent contrast warning.

Historical Euclidean single-run ablations of this same three-file logical patch on a later integration base showed Deep1B-1M recall at efSearch=1500 increase from 92.4417% to 94.9120%, while Jasper-1M moved from 97.9313% to 97.9920%. These are supporting observations, not exact-head performance claims or evidence that scalar quantization should become a default.