nx_intfp_train_gate.nx

nx_intfp_train_gate.nx -- THIRD brick of integer training: prove an integer model actually LEARNS. A Q16 2-layer MLP (W1->ReLU->W2) trained by INTEGER SGD to fit NEX input->target examples, loss L=sum(y-t)^2. Everything -- forward, backprop, AND the weight update -- is fixed-point integer, NO float. If the loss falls by >10x over the run, the sovereign integer optimizer works end-to-end (the last unknown before the full tape). Q16 (S=2^16): fwd as before; gy=2*(y-t); grads accumulate over examples; SGD W_q -= (lr_q * gW_q) >> 16. This is where fixed-point can bite: small gradients can UNDERFLOW the update (truncate to 0). We MEASURE whether it converges anyway -- honest data on integer training dynamics, not an assertion. license_tier: ORIGINAL