Child camera results in wrong combined world transforms (rotates into opposite direction or something like that).
Now, looking at updateworldtransformjob.cpp we have:
The transform in this case is merely the following (from qcamera_p.h):
All in all just multiplying with worldTransform seems to be wrong for a lookAt type of camera.
At first glance what it should be doing instead is decomposing the worldTransform and apply translate/rotate to the camera like QCamera::translateWorld() and rotate() would. (the decomposed translation and rotation is readily available in nodeTransform in fact) This is also how the camera controllers in extras work. Assuming that the camera's view matrix can be treated like an ordinary transform matrix when it comes to translate/rotate/scale is wrong.
Not sure how this could be plumbed though since this is all on the backend side, with no access to the QCamera.