This seems to be because of on inefficient implementation of gather_attributes_for_propagation in realize_instances.cc in the case when there are many instances of the same geometry. Should be…
LGTM and works fine in my tests. Just got a few inline comments.
I think this method is unnecessary. The main reason I prefer not having it is that it makes reasoning about the code on the call-site harder. With such a method I'm never sure whether it just changes the actual property, also also updates the node, or even the entire tree etc.
/** Name of the socket type (e.g. ...`). This can not be the integer socket type because custom socket types should be supported too. */
I like the x_to_y naming generally for matrices. It makes it obvious what they do and how to compose them, especially in more complex situations. But I'm also totally fine with using industry…
Site note, static_cast is preferred over reinterpret_cast when casting up/down a class hierarchy.
This wasn't really something we needed before, so there is no API for it yet. A new API could be added, but doesn't seem worth to make this part of this patch.