Documentation

Group-domain size at or below which the direct-indexed accumulator path is taken; wider domains keep the group-range kernel. The admitted reductions are order-independent, so the per-worker accumulator merge is exact.

Run a recognised reduction through the direct-indexed path. Nothing (so the caller falls back to the order-preserving kernel) unless the reduction is order-independent at this element type AND the column is a clean unboxed Int/Double.