Function short_conv_step 

pub fn short_conv_step(
    x: &[f32],
    w: &[f32],
    state: &mut [f32],
    hidden: usize,
    kernel: usize,
) -> Result<Vec<f32>, EngineError>

Causal depthwise short-conv one-token step (LFM2).

w layout: [hidden * kernel], channel-major, w[c*kernel + 0] taps the oldest sample. state is [hidden * (kernel-1)] (oldest→newest); updated in place.