For every spline, *all* of the normals and tangents in the output were normalized. The node is multithreaded, so sometimes a thread overwrote the normalized result from another thread. Fixing this problem also made the node orders of magnitude faster when there are many splines.