Function ssm_ba_gates_hopper_reject 

pub fn ssm_ba_gates_hopper_reject(
    requested: bool,
    kernel_present: bool,
    m: u32,
    n: u32,
    k: u32,
    k_stride: u32,
    sm_count: u32,
) -> Option<&'static str>

2026-09-25: Why the twin does not run, as one of BA_GATES_REJECTS; None means it runs. Pure, so tests cover it without a GPU. The shape guards:

• K % 8 != 0: the uint4 sweep would skip the tail (the parent has the same limit and does not check it);
• K_stride < K: the row is shorter than the reduction;
• N == 0 or K == 0: nothing to compute;
• fewer than ba_gates_min_tokens tokens: the decode guard in the module header.