Sounds like the subdivision is using the catmull clark algorithm on the quaterion orientation attribute naively. This attribute defines the twist on the curves when skinning. If you just want noodles, you can probably remove the orientation attribute created by the simulation. Otherwise, the only fix is to not subdivide the orient as a quaternion. If you decompose the orient to a normal and up vector, the subdivision will still be mostly valid and won't result in flips. I would file a bug too.