The PBVH is given the evaluated positions from the end of the modifier
stack, but then updates normals from the original mesh based on those
positions. This causes a feedback loop, even when in vertex paint mode.
As mentioned in a comment, calculating these normals and duplicating
the arrays is quite wasteful. But it's necessary since the PBVH always
assumes we are interested in modifying the positions, and also always
retrieves the normals from the original mesh too. Those things would
be good to untangle at some point too, but for now this change is small
and helps to clarify the situation as well.
Caused by: b339e3937d
Pull Request: https://projects.blender.org/blender/blender/pulls/114118
5af8b839cf