Shared operation for bf16 [M, K], axis-1 1MAD, T=256/L=12, k=2/3/4.
M=5..32. Borrow inputs and return owned [M, N]; caller controls evaluation.
Reconstruct the precise LUT value before bf16 weight rounding. The default
packed matvec's unrefined f32 code value is a different numerical contract.
Other geometries need a measured dispatch profile and model-level gates.
Shared operation for bf16 [M, K], axis-1 1MAD, T=256/L=12, k=2/3/4. M=5..32. Borrow inputs and return owned [M, N]; caller controls evaluation. Reconstruct the precise LUT value before bf16 weight rounding. The default packed matvec's unrefined f32 code value is a different numerical contract. Other geometries need a measured dispatch profile and model-level gates.