nx_nn_train.nx

nx_nn_train.nx -- sovereign FIXED-POINT TRAINING (batch gradient descent), the final piece of the neural-audio moonshot machinery. Together with nx_nn (the forward pass) this makes a net LEARNABLE entirely in integer arithmetic: no FPU, no GPU, no third-party autograd -- so a learned PLC / RVQ-codebook predictor can be trained sovereignly on affordable hardware. MSE loss, analytic gradient of a linear layer (dL/dW = (pred - target) . input), Q-scale step. lr_shift sets the learning rate (bigger = smaller, more stable steps). Composes the same matmul as nx_nn. license_tier: ORIGINAL