Function cost_flash_decoding_cuda 

pub fn cost_flash_decoding_cuda(
    _shapes: &[Shape],
    dtypes: &[DType],
    params: &OpParams,
    _caps: &BackendCapabilities,
) -> CostEstimate

Cost for the CUDA flash_decoding decode arm (binding-table form).

Doubles as the static ranker gate for the decode kernel’s supported set. Fuel dispatch is fail-fast (a registered kernel that returns Err fails realize; there is NO runtime fallback to the decomposed base map — see pipelined.rs “Fail-fast dispatch”), so an unsupported shape must never be PLACED on this kernel. This returns an INFEASIBLE cost (saturates composite_ns to u64::MAX) for anything the kernel can’t do — seq_q != 1, or head_dim outside [1, 128] — so the ranker keeps the decomposed base map / CPU oracle. Inside the supported set it prices a real decode: work scales with the live prefix k_len, not the physical capacity sk.

NOTE (design): cost DEPRIORITIZES but does not HARD-EXCLUDE a candidate (composite_ns ranks; it never drops). For a decode node whose K/V are already CUDA-resident this is sufficient — the decomposed arm is cheaper once the infeasible cost saturates. The GUARANTEED clean decline for a KV-resident prefill (seq_q > 1) is the Op::Branch decode-arm emission (symbolic-extent persistent decode; D2/D3), which only ever creates the CUDA flash arm for seq_q == 1 shapes. See the gating note in fuel-cuda-backend/src/baracuda/attention.rs.