Function byte_balanced_row_chunk 

pub fn byte_balanced_row_chunk(cols: usize, n_rows: usize) -> usize

Row-block size that keeps each streamed rows × cols f64 chunk near the library row-chunk target, with a 512-row floor so a wide design still makes useful BLAS-3 progress per block, capped at the total row count and never zero (a zero-row chunk is not a chunk; every consumer strides by it). The ONE row-chunk rule for streamed dense kernels — CPU (gam-solve) and GPU (gam-gpu) alike — so the two cannot drift apart again (#2469, #2704).