The problem here was that we were dropping markers that don't move for a prolonged period of time, on the basis that they've probably fallen off the body or been incorrectly gap-filled. Antoine's example had tens of thousands of frames of markers not moving at all, which meant those markers were getting dropped and so IK was not constrained by anything, so it got some wonky solutions. By just removing that filter we've fixed his issue.
The problem here was that we were dropping markers that don't move for a prolonged period of time, on the basis that they've probably fallen off the body or been incorrectly gap-filled. Antoine's example had tens of thousands of frames of markers not moving at all, which meant those markers were getting dropped and so IK was not constrained by anything, so it got some wonky solutions. By just removing that filter we've fixed his issue.