Function hc_post_lowrank 

pub fn hc_post_lowrank(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    block_out: DevicePtr,
    residual: DevicePtr,
    inj: DevicePtr,
    out: DevicePtr,
    num_tokens: u32,
    hidden_size: u32,
    hc_mult: u32,
    stream: u64,
) -> Result<()>

Inject the block output back into every stream: out[t, s*H + d] = residual[t, s*H + d] + block_out[t, d] * inj[t, s].

Note there is no comb argument: DeepSeek mixes streams with a full [hc, hc] combine matrix on the way back, Qwen scales by one scalar per stream. Passing a combine matrix here would not type-check, which is the point of keeping the two launches separate.