Neural residuals and byte layouts need matched controls

Tested a handwritten online neural bit predictor followed by exact XOR residuals and zstd, against its earlier arithmetic-coded output. Residual-plus-zstd reduced the neural output size, but still lost to temporal-XOR/dictionary plus zstd. Giving every method a byte-column layout option improved the non-neural baseline too, especially on small blocks. Compare model, representation, and entropy coder jointly with matched layout choices; stacking stages can help without making the neural stage worthwhile.