Function glm_hc_pre 

pub fn glm_hc_pre(
    gpu: &dyn GpuBackend,
    kernels: &Glm5NextMhcKernels,
    streams: DevicePtr,
    w: &Glm5NextMhcSiteWeights,
    y_out: DevicePtr,
    post_out: DevicePtr,
    comb_out: DevicePtr,
    num_tokens: u32,
    hidden_size: u32,
    hc_mult: u32,
    sinkhorn_iters: u32,
    norm_eps: f32,
    hc_eps: f32,
    stream: u64,
) -> Result<()>

hc_pre: collapse the hc_mult FP32 streams to one BF16 sequence and emit this site’s post / comb mixing coefficients.

🪤 Named glm_hc_pre, not hc_pre: ops already exports DeepSeek-V4’s hc_pre/hc_post through a glob, and the two are NOT interchangeable — V4’s reads hc_head.{fn,base,scale} and uses a different mixing law. The rename is the compiler-enforced version of this module’s “never silently fall back onto the DeepSeek variant” rule; a glob collision here resolved the wrong way would be a silent architecture swap.

The residual that glm_hc_post mixes is the stream tensor as it entered here — snapshot it before calling, which is the skeleton’s ResidualStep::SaveResidual. Overwriting the streams in place before hc_post runs silently changes what the residual means.