Another thought I had is that this node could support the case of a single instance quite easily to avoid realizing the target geometry instances in some cases (like the output of the join geometry node)

if target only has instances && target.instances_amount == 1 && target.instances[0].type == mesh:
   transform source geometry coordinates and ray directions into the transform space of the target mesh instance

That's probably a valid optimization in plenty of cases, so it could just be done in the future if it was worth it.

Another option is making a reduced version of realizing instances that only duplicates the vertex positions in that 1-instance case (at some point in the future).

Hit Index Output
I don't think we should include this in the node right now. I think mostly we should try to keep indices as an implementation detail while we can.
And also, inputting a list of attribute names will cover the case of fast transfer of multiple attributes when we implement it.

Yes, i wasn't sure about the index output socket, will remove it.

Naming and Category
Initially I was skeptical about putting this in the "Geometry" category, but I guess this node could work with more geometry types in the future.
The one I can think of is volume level sets when we add them in the future, I can't think of anything else though.

I agree, there doesn't seem to be a more fitting category atm. Easy to change though, and more future nodes will give us a better idea of how to organize them.

AttributeInterpolator Class
I like what you've done here, it looks quite useful, and simplifies interpolating multiple attributes. A few comments though:

• Naming: This is really specific to meshes, so I would include that in the name.
• Location: We already have mesh_sample.cc created recently for this purpose, I think it makes more sense to have this class there, since it's fairly low level code useful in many places.

Done, it lives in BKE_mesh_sample.hh as MeshAttributeInterpolator now.