nx_reader_quantize_gate.nx

nx_reader_quantize_gate.nx -- R4 of the neural-reader arc: QUANTIZE-AFTER. Takes the f32-TRAINED reader (reader_neural_f32*.bin, our own from-scratch model), quantizes its weights to Q16, and runs PURE-INTEGER (nfa_*) inference -- the train-float / infer-int split, completed for a SOVEREIGN-TRAINED model. Proves the no-float reader is FAITHFUL to the f32 model by running BOTH forwards on the same real held-out SQuAD split and comparing span predictions + gold-F1. (The nfa_* Q16 forward has NO collapse risk -- collapse was a TRAINING dynamics artifact; inference of trained weights is exact-enough in Q16.) T1 no-float (Q16) held-out span-F1 within 60 permille of the f32 reader's F1 (quantization faithful) T2 Q16-vs-f32 span-prediction AGREEMENT > 800 permille (same model, integer vs float) T3 the Q16 path is PURE INTEGER (structural: only nfa_* Q16 ops, zero nx_f32 in the inference) T4 deterministic Consumes: reader_neural_f32.bin (f32 W), semppmi_v1.bin (vocab hashes), reader_rows.bin (same rows/vocab as training so token ids address the trained embedding rows). expect_exit: 0 license_tier: ORIGINAL