cluster_cull

The cluster-LOD cut + cull. One dispatch thread per (instance, cluster) pair: it picks the watertight continuous LOD cut (render a cluster iff its parent group projects above the pixel threshold and its own group projects at or below it), frustum- and backface-cone-culls the survivors, and atomically compacts their triangles into one visible_indices buffer plus a per-triangle instance mapping. The compaction cursor is the index_count word of the DrawIndexedIndirectArgs block in draw_args, so the cluster draw rasterizes the compacted buffer in a single INDIRECT indexed draw over exactly the triangles the cut kept — the count reaches the draw without a CPU round trip, so the drawn set never lags the camera by a frame. Camera, viewport, and the threshold ride in the cut uniform.