Function moe_lora_gather_bgmv 

pub fn moe_lora_gather_bgmv(
    gpu: &dyn GpuBackend,
    kernels: &LoraKernels,
    route: &MoeExpertRoute,
    x: DevicePtr,
    base_out: DevicePtr,
    indices: DevicePtr,
    row_adapter: DevicePtr,
    xa: DevicePtr,
    n_slots: u32,
    top_k: u32,
    x_gather: u32,
    stream: u64,
) -> Result<()>

SOLID Incr-4: launch the DECODE-path MoE expert down fold. The unsorted, slot-major analogue of moe_lora_grouped_down — instead of an expert_offsets prefix sum over sorted rows, each flat (token, slot) row gathers its expert from indices[row] (the same indices_dev the fused expert GEMV routed on) and its base/adapt decision from row_adapter[row / top_k] (< 0 = base skip, or DevicePtr::NULL to fold every row on the single-active-adapter path).

x = the post-swiglu activations (silu(gate)*up, produced by the caller’s moe_silu_mul launch into a packed [n_slots, k_in] BF16 scratch — the SAME kernel + BF16 round the prefill fold uses, so the delta is BF16-ULP identical to prefill). base_out = the slot-major expert_down_out ([n_slots, n_out] BF16, folded IN PLACE before moe_weighted_sum_blend, so the router weight multiplies base+delta). xa = the fixed-address [n_slots, max_rank] BF16 shrink scratch. The grid is EXACT (n_slots is a host constant per captured graph) — no worst-case tiles — and all args are pointer/value-stable, so the launch captures cleanly.

ARG ORDER is in lockstep with moe_lora_gather_bgmv.cu (cuLaunchKernel is type-blind; keep both in sync).