REG: Regression with changing items in subtrees

There are some tests in qmlbench that are supposed to test changing items in a subtree (isolated from the rest by a clip node) and verifying that this does not cause a full update on the rest of the scene.

In Qt 6, they are giving some strange and inconsistent results.

On Ubuntu, there is a regression for when the clip is rotated (no scissors), but an improvement for the common rectangle clip case. So this may be related to RHI in some way and not to the scenegraph:

On Windows, there is an almost 100% regression for both these tests, indicating that something may be going very wrong with the Direct3D backend:

Link to dashboard