This notorious topology change and shrinkage per iteration does not occur with the Limit Surface based SubD approach (as already used inside several CAD systems and planned by McNeel). Here one always edits either the control cage or the smooth Limit Surface.
Looking at these smooth limit surfaces in isolation they should be precise enough, even for you :o).
But several roadblocks for precision work do remain in place. One can not trim SubD-models without spoiling their topology. One for the same reason may not combine several SubD objects (as Booleans would also involve trimming). If one wants clean looking shape intersections SubD Modelers often completely redo the control cage in order to conform the new shape – a very time-consuming manual process.
In order to deal with this huge limitation one has come up with various strategies:
Inside mesh modelling applications there’s products like Mesh Fusion in Modo (there’s similar tools for Blender, Maya etc.). This finally gives access to Trimming / Booleans on sets of SubD input meshes, without time-consuming workarounds. Output however is a quad dominant mesh similar to a Nurbs render mesh. Other strategies are to use SubD for laying out base geometry and to switch over to very high resolution meshes in later stages, which again no longer obey the Catmull Clark scheme. That happens in Zbrush for instance.
Inside very expensive, parametric feature base CAD-systems one runs two separate geometry Layers, the SubD Cage plus a Live-converted Nurbs version where one may trim, apply fillets and the like. This approach is kind of cool, but one can say that it is out of reach for McNeel for foreseeable future. Also patch layout and overall surface quality of the auto converted Nurbs certainly do never match serious production needs,
as pointed out by you. One uses this in all ways crazy expensive dual workflow in order to yield conceptual geometry.
The other large limitation I see is transition control. With meshes that’s simple – one adds control loops (more faces) and slides them around until things look good. But as also pointed out by you – that’s display only and only looks good in conjunction with openGL smoothing tricks.
With Limit Surface based approaches one rather would not want to add needless complexity to faces. The only thing one has at disposal here is point weighting (creasing) but this alone would be way too limited for serious work. New Tools like Xirus propose alternative interfaces to control surface flow with their Tangency Handles but this all still looks extremely rough to me.
Even if one had all of this solved and had a precise limit surface based SubD model: One still had to convert to Nurbs – and that’s a great problem as well. Thus far I have not seen a tool which outputs clean, human editable patch layouts: Output is overly complex, hard to edit. While compatible for milling now its no good input either – in the sense that every other crappy Nurbs model may be a pain (think creating helper geometry, curve and surface offsets etc – this all sucks with weird surface layouts).
Likely it will be the Tool maker who redoes at least parts of the model to make things work.