I think changing the order here is fine as a fix for this particular issue, but we may be squashing bugs similar to this for 4.3 unfortunately; ideally I think we would do something like bke::object::pbvh_ensure instead of bke::object::pbvh_get inside undo::push_begin_ex so that we're not dependent on external ordering, but unfortunately that requires a depsgraph object, which would mean changing the signature of push_begin_ex and push_begin to take it.
Instead of checking for the subdiv_ccg, looking at the PBVH type is better.
This comment is unnecessary, I'd remove it.
Writing my current thoughts here: seems like the most straightforward way to solve this with geometry_push is to extend the NodeGeometry struct with two more CustomData layers for MDisps and…
StepData into structs
