nx_bm25_test.nx

nx_bm25_test.nx -- prove the team's integer BM25 reproduces the PRODUCTION ranking and corrects the old TF-IDF ranker's length bias. Corpus is engineered so length-normalization FLIPS the winner: doc0 "injury injury" (tf=2, dl=2 -- short + focused) doc1 "injury injury injury f f f f f f f f f" (tf=3, dl=12 -- more hits but diluted) Naive tf*idf (no length norm) picks doc1 (higher raw tf). BM25 picks doc0 (focused), which matches a float reference BM25 -- the triangulation the runner checks. Exit 0 if the team's logic holds 5/5; the runner then confirms the team's BM25_BEST == the Python reference's. license_tier: ORIGINAL