Function select_tokens 

pub fn select_tokens(
    gpu: &dyn GpuBackend,
    kernels: &Glm5NextDsaKernels,
    cfg: &Glm5NextDsaConfig,
    geom: &DsaSelectGeometry,
    inputs: &DsaSelectInputs,
    scratch: &DsaSelectScratch,
    launch: DsaSelectLaunch,
    stream: u64,
) -> Result<()>

Run the four selection kernels, leaving [q_rows, out_width] token ids in DsaSelectScratch::tokens.

Launch geometry and argument order are transcribed from the GATE-4 arm of examples/dsa_indexer_microtest.rs, which is the numerically-proven reference. The pass is enqueued on stream and not synchronised — the caller sequences it with the attention that consumes the selection.