nx_bm25_lib.nx

nx_bm25_lib.nx -- BM25 lexical retrieval + RECIPROCAL RANK FUSION, in pure integer fixed point. WHY THIS, WHY NOW (July 2026 SOTA, researched): every credible legal-retrieval stack this year is HYBRID -- a dense vector retriever AND BM25, fused with Reciprocal Rank Fusion, then reranked by a cross-encoder. Neither leg alone wins: vectors catch paraphrase, BM25 catches the exact defined term ("Force Majeure Event", a party name, a section cross-reference) that an embedder blurs away. In contracts the exact term is very often the whole question, which is why the lexical leg never gets dropped even in fully neural stacks. This builds the leg that needs NO MODEL, plus the fusion that any future embedder plugs straight into. The banked house order is embedder -> cross-encoder -> grep-tools -> graph LAST; BM25+RRF is the skeleton those slot into, and it is useful on its own from day one. u2605NO FLOAT. BM25 needs a logarithm and a saturating ratio, both normally float. Everything here is integer fixed point in MILLI units (1000 = 1.0), including a bit-position natural log with linear interpolation of the mantissa. Ranking only needs ORDER preservation, and integer fixed point preserves order exactly while staying reproducible bit-for-bit across machines -- a float BM25 can reorder two near-tied clauses depending on FMA and compile flags, which makes a retrieval regression impossible to reproduce. Determinism is worth more here than the third decimal place. u2605IDF IS THE HALF PEOPLE DROP. Scoring on raw term frequency ranks a clause highly for containing "the" fifty times. IDF is what makes a rare defined term outweigh filler, and BM25's length normalisation (b) is what stops a long clause winning purely by being long. Both are load-bearing. Parameters are DATA (rule 11): k1 and b as named constants in milli, the standard 1.2 / 0.75. license_tier: ORIGINAL LIB.