Based on the description, I liked the idea. However, testing this with the Add > Curves > Random primitive (with experimental option "New Curve Tools" enabled), I don't really like the results more in practice. When dragging the slider, I'm noticing a lot more jumping than before.

I wonder if it's preferable to make sure it's possible to accomplish something like this after the trim process.
Or maybe it needs to be an option, because the results are better sometimes as well. It's just a bit hard to justify that without seeing people complain about the existing behavior.

Sampling within segments, interpolating, etc. seems like something to pursue more generally anyway. That makes me thing that it might be helpful to think about how the trim node could be split up to smaller components for this sort of behavior.
I do think that this is preferable to converting to Bezier curves implicitly, at least as a default. Otherwise people will end up paying for that later without realizing it. If this imperfect sampling is really bothering people, they can resample or convert the curves beforehand (though that does have a performance cost).

In short, I'm intrigued, but not totally convinced, and I think there's some opportunity to find a more general design here.

In D16066#438608, @Hans Goudey (HooglyBoogly) wrote:

Based on the description, I liked the idea. However, testing this with the Add > Curves > Random primitive (with experimental option "New Curve Tools" enabled), I don't really like the results more in practice. When dragging the slider, I'm noticing a lot more jumping than before.

Yes the snapping/jumping is a bit more obvious so for animation purposes I suppose it would be worse. Experimented a bit with other approaches and there was another approach that was intriguing where you shift the trim point toward the line intersecting the two neighboring points since once you have formed a line one can add/remove points however one wants. The issue with such an approach was that the trim point isn't really following the curve (which might be unintuitive) and one ends up with a linear segment which isn't really a natural or nice shape. On the positive side you could actually avoid jumping/snapping behavior as one is removing the point from a linear segment and the shape remains the same.

Thought about adding/removing multiple points but the issue with doing that is you essentially are:

a) Likely changing the shape of multiple segments.
b) At some point you are adding removing points and you end up with the same problem you started with but you are just doing it somewhere else.

In the end I thought this behavior was good because I thought the primary use case for this would be for hair in which case this method seemed the cheapest and best solution as long one is not considering animating the trim effect. But if animating the trim effect is of interest I think we should consider implementing a node for converting from Catmull-Rom -> Cubic Bezier as that seems like the proper solution rather then trying to invent some heuristic solution better then this one. Because I think the output curve is always better which I think is the issue with the current solution in certain cases the output is just bad. But this approach will be unintuitive for some as the curve jumps/snaps, but it already does that in but it could be 'more intuitive' when it hapends?

That makes me thing that it might be helpful to think about how the trim node could be split up to smaller components for this sort of behavior.

Could be possible but most likely such a node would impact more points on the curve then just the endpoints since it would have to be more general.