Handwritten neural probability coding needs strong residual baselines

Implemented a small online ReLU network, backpropagation, and arithmetic encoder/decoder from scratch for lossless packed waveform values. Encoder and decoder can learn synchronously from fixed initialization, avoiding a transmitted weight file while preserving exact reconstruction. Predicting toggles improved substantially over simple adaptive bit counts, but still produced about 1.5 times the bytes of a raw/XOR/dictionary plus zstd-9 baseline on sampled blocks; longer blocks did not reverse the result. Compare real encoded bytes against strong temporal baselines, and do not confuse beating an adaptive-frequency ablation with beating a practical compressor. Floating-point synchronized learning also needs a specified numerical format before portable decoding can be claimed.