research(nightly): 2026-05-08 Locally-Adaptive Vector Quantization (LVQ)

[CrossGen-ai]

Summary

Adds a working Rust PoC for Locally-Adaptive Vector Quantization (LVQ) —
the per-vector 8-bit scalar compression scheme from Intel's Scalable Vector
Search (Aguerrebere et al., VLDB 2024) — as a new workspace crate
`crates/ruvector-lvq`. Includes a two-level residual variant (LVQ-8x8) that
matches fp32 recall at half the memory.

• Crate: `crates/ruvector-lvq/` (pure Rust, `#![forbid(unsafe_code)]`, all files <500 LOC)
• ADR: ADR-193
• Research doc: docs/research/nightly/2026-05-08-lvq-locally-adaptive-vq/README.md
• Public gist (SEO writeup, real benches): https://gist.github.com/CrossGen-ai/3e266d36b77394ff5d770c9ee229c66e

Real benchmarks (Apple M4 Max, n=200K, d=128, k=10)

Variant	Memory	Recall@10	QPS
fp32 baseline	97.66 MB	1.000	148
LVQ-8	26.70 MB (27.3%)	0.942	120
LVQ-8x8 reranked 10×	53.41 MB (54.7%)	1.000	121

Test plan

• `cargo build -p ruvector-lvq --release` — succeeds
• `cargo test -p ruvector-lvq --release` — 10/10 pass (3 unit + 4 module + 3 integration)
• `cargo run -p ruvector-lvq --release --bin ruvector-lvq-bench` — prints real numbers
• `LVQ_N=200000 cargo run …` — confirms recall holds at 4× scale
• Wire `lvq8_l2sq` into `ruvector-core::hnsw` (follow-up)
• DiskANN block format integration (follow-up)
• `ruvector-lvq-wasm` and `-node` mirror crates (follow-up)

Notes

This branch is on a fork (`CrossGen-ai/RuVector`) because the nightly
research bot does not have push access to `ruvnet/RuVector`. PR is opened
draft so the upstream maintainers can decide whether to merge or rebase.