I find it a bit strange that there is an implementation of a generic heap data structure for the weighted sample elimination algorithm but the point cloud data structure used for nearest neighbor searches uses the standard C++ <algorithm> provided make_heap, push_heap, and pop_heap functions. It's especially strange (to me) since the commit that introduced the weighted sample elimination also introduced this max heap data structure. I assume the intent was to use it for both algorithms.
With some local testing here, converting the point cloud data structure to using cyHeap by creating a side-channel heap structure is slightly faster than using the standard C++ max heap algorithms in <algorithm>.