nx_vec_embed.nx

nx_vec_embed.nx -- R-VEC-1 of the onsite-search S-class ladder: SOVEREIGN text->vector embeddings (LIBRARY), the CLASSICAL (count-based distributional) kind -- "methods to generate this mapping include neural networks" (cited srch_embedding.raw); this is the pre-neural, fully-sovereign, no-external-weights path (ladder option A). Principle (distributional semantics, cited srch_word2vec.raw): words in SIMILAR CONTEXTS get SIMILAR vectors. Realized with zero ML training instability: a term x term co-occurrence matrix from the corpus; each term's vector is its co-occurrence row; a query/doc is the bag-sum of its term vectors; relatedness = cosine (R-VEC-0 kernel). KAT lives in nx_vec_embed_gate.nx (library+gate split, matching nx_vec_kernel/nx_vec_fuse). HONEST: count-based, NOT neural understanding -- the neural cross-encoder stays census-BEHIND (R-VEC-6). Production scale = fill the matrix from the real manifest + reduce dims (R-VEC-4 quantization) -- wiring, not math. exports: vr_embed (term->vec), ve_docvec (bag->vec), ve_build_co. license_tier: ORIGINAL