nx_vec_index.nx

nx_vec_index.nx -- R3 rung 1: a SOVEREIGN integer dense-vector store + exact cosine k-NN = the MILVUS/ANN replacement (deep-searcher rides Milvus; depstack census graded VDB=BEHIND). No-float: vectors are integer, normalized to fixed scale S with fixed-point precision P via integer isqrt (Newton), similarity = integer dot of unit vectors -> cosine in PERMILLE (0..1000, sign-preserving). Exact brute-force k-NN (correct; approximate HNSW is a SCALE optimization, deferred). HONEST SCOPE: this is the STORE+SEARCH half. The trained EMBEDDER that turns text->vector (the Jina-Embeddings replacement, EMB axis) is R3b -- distributional/trained, NOT built here. main() = KAT gate with a negative control (an orthogonal doc must NOT outrank a parallel one). license_tier: ORIGINAL expect_exit: 0