The problem with potential jumps due to Cartesian interpolation in principle remains, which is a dilemma of the current design:

- We do need Cartesian interpolation to reduce the risk of state rejection due to a failing path constraint, which currently cannot be checked during interpolation.
- Ideally, we would do joint interpolation and only resort to Cartesian interpolation if the path constraint is violated, i.e. follow a similar approach as in #3618. However, this would require access to the path constraint...

The tolerance of orientation constraints needs to be interpreted w.r.t. the given frame_id of the constraint instead of the target frame. Not doing so, a large tolerance is interpreted about the wrong axis and resolving constraint frames fails.
Unfortunately, the proposed approach only works for the ROTATION_VECTOR representation.