nx_vec_fuse.nx

nx_vec_fuse.nx -- R-VEC-5 of the onsite-search S-class ladder: SOVEREIGN hybrid-search rank fusion (LIBRARY). THE felt S-class jump: lexical BM25 nails exact terms, vector cosine (R-VEC-0/1) nails paraphrase -- fusing their two ranked lists gets the best of both. "Data fusion in information retrieval combines results from multiple search systems or retrieval models" (cited srch_ir.raw). Reciprocal Rank Fusion (Cormack/Clarke/Buettcher 2009): score(d) = SUM_lists 1/(k + rank_i(d)), k=60 canonical; sort docs by descending score. The defining behavior: a doc ranked MODERATELY in BOTH lists outranks a doc ranked #1 in only ONE -- agreement across retrievers beats a single strong signal. NO-FLOAT (operator doctrine): the reciprocal is computed in fixed point, weight = SCALE/(k+rank), SCALE=1e6. Determinism: strict-greater selection sort keeps ascending doc-id order on ties. exports: vr_rrf_w, vr_fuse_add, vr_fuse_rank. license_tier: ORIGINAL