nx_dr_fuse_cli.nx

nx_dr_fuse_cli.nx -- fit the learned combiner on REAL DRBench labels and report HELD-OUT performance against every individual signal (DR-12). nx_dr_fuse <feats.csv> [train_percent] feats.csv = flat comma-separated ints, 5 per candidate: task,label,lexical,semantic,semantic_idf Weights are fitted ONLY on the first train_percent of tasks and every reported number is computed ONLY on the DISJOINT remainder -- fitting and scoring on the same tasks would be rigged, so the split is explicit and reported. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0