nx_reader_mlp_train.nx

nx_reader_mlp_train.nx -- NEURAL-READER RUNG 2: train the span-reader scorer as an MLP on nx_autograd over the REAL MODE-R candidate features (dumped by `nx_qabench dumpfeats` -> knowledge/index/reader_feats.bin), and MEASURE whether nonlinearity beats the linear perceptron that plateaued (~250 vs oracle 727). Pairwise hinge ranking loss: the gold candidate must score above each negative by a margin. Reports TRAIN ranking accuracy (argmax==gold) for the MLP vs a linear baseline trained the same way -- the decisive de-risking measurement before wiring integer inference into nx_qabench. Persists quantized integer weights -> knowledge/index/reader_mlp.bin (Q12) for the no-float inference path. TEETH: T1 data loaded (rows>=500) T2 MLP ranking-acc > linear ranking-acc (nonlinearity helps on REAL feats) T3 MLP ranking-acc >= 500 permille T4 weights persisted+reloaded. expect_exit: 0 license_tier: ORIGINAL