The minimum twist mode is important because it allows creating normals without sudden changes in direction. The disadvantage of minimum twist normals is that the normals depend on all control points. So changing one control point can change the normals everywhere. The computed normals do not match the existing code exactly, although they do match quite well on non-cyclic and on some cyclic curves. I also noticed that the existing implementation has some fairly simple failure cases that I haven't found in the new implementation so far. Differential Revision: https://developer.blender.org/D11621