2.) let me express a bit more precisely here QSceneGraph seem to have 2 uses or components
a.) is a scene interpreter needed for QML to render in here i agree with you totally named QDeclarativeEngine and here C++ devs have no complains or interest at all
b.) the rendering engine which is more tied to lighthouse project than QML that provide many nice advances in the render/texture management/shaders/object relation in the framebuffer/canvas operations in hot at the GPU,text|glyph rendering between others <--- this is the part we want to be exposed to the C++ API[seem they wanna to stabilize some more before allow this] and be as flexible as QtPainter/QGLpainter so we can reimplement selected render operation to adapt it to more complex projects and allow at the same time to speed QWidgets render.
The problem is QML is nice for the mobile space[<-- nokia ...] and graphic simple applications that need cool eye candy but for lets say a multi layered massively tiled rendering[GIS/CAD/etc] or lets say multi input HD video RT processing is deadly slow. i know this is better done in c++ side but you have to rewrite half QT render path to achieve it and many of this changes are already in this engine and using it properly will save you from a truckload of code and hacks, so if they make this changes later in another revision everyone with this kind of workload can make good use of this technology without reinvent the wheel everytime <-- that is all the issue here