nx_vec_vq.nx

nx_vec_vq.nx -- R-VEC-4 of the onsite-search S-class ladder: SOVEREIGN vector quantization (LIBRARY). Raw f32 embedding vectors blow the RAM budget at 366k scale (Rule 21); VQ compresses each vector to ONE codebook index -- "a codebook" of representative centroids (cited srch_vq.raw, Vector_quantization). k-means (Lloyd): assign each vector to its nearest centroid, move each centroid to the mean of its members, repeat. Integer, no-float, deterministic init (evenly-spaced seeds). HONEST SCOPE: single-codebook VQ; PRODUCT quantization (split the vector into sub-spaces, a codebook per sub-space) is the named extension for finer compression. exports: vr_vq_train, vr_vq_encode, vr_vq_decode. license_tier: ORIGINAL