Trajectory Planner using Ruckig Lib

Yes that is correct. Three ta knot's can create a plane.

Note, the tool orientation vector and the current TA knot already define a plane:
 

On this plane we could record the tcp-ta intersection along the plane.

This crossed my mind as well but will only work for small changes in orientation angles. If we have a simple move like this:
G1 X0 A0
G1 x100 A180

there is no single plane that records the intersection path of the tcp-ta vector. As the vector rotates from pointing straight up to pointing straight down the resulting curve would disappear in infinity.

This 2020 publication proposes the use of 3d clothoid fillets for space corner smoothing in tool paths:
www.sciencedirect.com/science/article/abs/pii/S0736584519306222

So it seems that 2D fillets are still the norm in tool path segment blending, which gives me the impression that even modern controllers can't blend non-planar segment corners including anything that involves a helix (which is often used for tool entry moves). I have no idea really but this might explain why 3D CAM toolpaths are often made up of short straight segments as two straight segments will always form a planar corner.
All this makes me think that using the current (planar) clothoid library would require a segmentation of the recorded tool orientation path and then calculate the plane and fillet for every segment corner. I wouldn't know how to do the segmentation of the recorded orientation path much less if this is computationally feasable. Also it seems kind of silly to have to segment an entire helix when it is really just the beginning and the end that need fillets.

I had a look at your proposed method of fitting helixes:
forum.linuxcnc.org/38-general-linuxcnc-q...lib?start=420#303355
This would probably be ok for many cases, however I'm somewhat skeptical about the predictability of the path deviation as it seems to deform the helix not only locally at the beginning and the end but also quite noticeably in the middle. If this was done to our helical path of the tool orientation the deviation of the orientation angle might well be beyond acceptable.

So for the moment I have run out of ideas, unfortunately.

Looks very nice!
When you say 'tooldir blend 5mm' does that mean the blended curve has a maximum deviation of 5mm from the recorded path?