nx_bm25.nx

nx_bm25.nx -- production-grade ranked retrieval: Okapi BM25 (Robertson & Sparck Jones), integer-only. This is the HONEST upgrade over the team's first TF-IDF ranker, which lacked LENGTH NORMALIZATION: a long document that mentions a term once would out-rank a short focused document with the same per-length relevance. BM25 fixes that with (1) tf SATURATION (k1) -- the 10th occurrence adds less than the 2nd -- and (2) LENGTH NORMALIZATION (b) -- penalize long docs by dl/avgdl. The point of building the real standard (not a weak baseline) is to TRIANGULATE: nx_bm25_test asserts the team's integer ranking reproduces a float reference BM25's ranking, so the "exceed" is measured against the production formula, not cherry-picked. license_tier: ORIGINAL BM25(q,d) = sum_t idf(t) * tf(t,d)*(k1+1) / ( tf(t,d) + k1*(1 - b + b*dl/avgdl) ) idf(t) = ln( 1 + (N - df + 0.5)/(df + 0.5) ) = ln( (2N+2)/(2df+1) ) [always positive]