Function hc_head 

pub fn hc_head(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    streams: DevicePtr,
    head_fn: DevicePtr,
    head_scale: DevicePtr,
    head_base: DevicePtr,
    y_out: DevicePtr,
    num_tokens: u32,
    hidden_size: u32,
    hc_mult: u32,
    norm_eps: f32,
    hc_eps: f32,
    stream: u64,
) -> Result<()>

Final collapse before the LM head: a single learned sigmoid-weighted sum over the hc_mult streams. One block per token.