nx_ltr_lib.nx

nx_ltr_lib.nx -- LEARNING-TO-RANK reranker as a library so a gate drives it in-process on a synthetic fixture (search rung R0, 2026-09-14). COORDINATE ASCENT (Metzler-Croft) that directly maximises training-fold nDCG@10 one weight at a time by step-halving line search, INITIALISED at the BM25-only solution so it can never rank worse than BM25 on the folds it trains on; evaluated by K-FOLD CROSS-VALIDATION so no query's own label ever trains the weights that score it (the contamination-free path when only a test split exists). Integer/fixed-point end to end: features are min-max normalised to LTR_SCALE per query, weights are i64, ranking is an integer dot product, so the result is bit-reproducible. The harness (nx_beir_eval) stores per-candidate features + gold + idcg during its BM25 pass and calls ltr_cv over the SAME candidate pool, so the learned arm is the same controlled comparison as the semppmi and rrf arms. license_tier: ORIGINAL