Function mlp_ws_bytes 

pub fn mlp_ws_bytes(cfg: &Glm5NextMlpConfig, max_rows: usize) -> [usize; 14]

Per-buffer device byte sizes of one Glm5NextMlpWorkspace, in allocation order.

🔴 Split out of new so the sizing can be unit-tested and LOGGED without a GPU. The load path prints the total per workspace and the stack total: the whole point of the shared workspace is a number that used to be multiplied by 45, and a number nobody can read is a number nobody checks. Every term here must stay in lockstep with new — the test ws_sizing_matches_new walks both.