The old math wasn't handling inverted quaternions. In comparison,
Matrix.lerp is actually using internally exactly the same math code
that is used by constraint influence.