feat: TurboQuant 4-bit vector quantization

[damahua]

Summary

• Add TurboQuant (Zandieh et al. 2025) 4-bit scalar quantization to reduce per-vector memory from 3072 to 384 bytes (8x compression for 768-dim vectors)
• Enabled via quantization_bits=4 parameter in HnswIndexInitConfig (default 0 = unchanged)
• Companion to feat: TurboQuant 4-bit vector quantization for HNSW index chroma#6794

Changes (5 files)

• hnswlib/turbo_quant.h (new): TurboQuantizer class — random sign rotation + 4-bit Lloyd-Max codebook + symmetric distance; TurboQuantL2Space and TurboQuantIPSpace wrappers compatible with SpaceInterface
• hnswlib/hnswalg.h: quantizer_ field, set_quantizer(), quantized storage in addPoint, quantized query in searchKnn, dequantize in getDataByLabel
• hnswlib/hnswlib.h: Include turbo_quant.h
• src/bindings.cpp: create_index_quantized() FFI entry point, set_quantizer() wiring after index init
• src/hnsw.rs: quantization_bits field in HnswIndexInitConfig, conditional create_index_quantized call

How it works

1. Random sign rotation (diagonal Rademacher matrix) decorrelates dimensions, making independent scalar quantization near-optimal
2. 4-bit Lloyd-Max codebook (16 centroids for standard normal) encodes each dimension
3. Symmetric distance: both vectors dequantized to float32 for exact L2/IP/cosine — used in graph construction (addPoint) and search (searchKnn)
4. getDataByLabel: dequantizes stored codes back to float32 for Chroma's brute-force fallback path

A/B Test Results (N=3, 50K × 768-dim, Chroma single-node)

	Run 1	Run 2	Run 3	Mean
Baseline (float32)	317 MB	315 MB	318 MB	316.7 ± 1.5 MB
TurboQuant (4-bit)	207 MB	176 MB	190 MB	191.0 ± 15.6 MB

39.7% peak RSS reduction. Zero errors across all API requests.

Test plan

• Existing hnswlib test suite passes with quantization_bits=0
• Existing hnswlib test suite passes with quantization_bits=4
• Measure recall@10 accuracy vs float32 baseline
• Test persistent index save/load with quantized data

🤖 Generated with Claude Code