Function cublas_bf16_proj_dense 

pub fn cublas_bf16_proj_dense(
    act: DevicePtr,
    weight_bf16: DevicePtr,
    out: DevicePtr,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Route a projection out[M,N] = act[M,K] @ weightᵀ through cuBLASLt BF16 for a weight that is already BF16 [N,K]. Two kinds of caller: models whose attention/shared-expert weights ship unquantized (e.g. Laguna), and the row-wise GDN prefill arms, which hand it the ledgered BF16 dequant qwen3_ssm/rowwise_bf16.rs writes once per layer.

There is deliberately NO cublas_bf16_proj beside it any more — the dequant-and-cache variant that used to own the FP8→BF16 expansion is the #917 off-ledger allocation (see dequant_fp8_bf16_into). Splitting “who owns the BF16 bytes” from “multiply them” is what keeps the ledger honest: this function cannot allocate.