Function matvec_reference 

pub fn matvec_reference(
    weight_words: &[u32],
    x: &[f32],
    rows: usize,
    row_bytes: usize,
    n_blocks_per_row: usize,
) -> Vec<f32>

Host emulation of the Q8_0 matvec shader.

weight_words is pack_words applied to rows * row_bytes of GGUF Q8_0 data; x holds n_blocks_per_row * 32 activations. Returns rows floats. row_bytes is passed rather than derived, mirroring the push-constant block.

This is a correctness reference, not a fast path.