Once trimming is in play, it’s probably best to stick with NURBS surfaces.
Why:
One of the main properties of SubDs that makes them powerful modeling tools is the control net does not have to be a rectangular grid and they are generally curvature continuous. This permits a great deal of freedom in creating smooth surface shapes with a wide range of surface topology. However, the arbitrary topology means there is no canonical surface parameterization over an entire SubD.
Part of robust trimming requires parameter space and 3d trimming curves that work in tandem. The need for easy to create and maintain accurate parameter space curves tips the scales heavily in favor of using an underlying surface with a standard global and easily accessible parameterization (like NURBS surfaces).
It is true that trimming can be implemented on Catmull-Clark subdivision surfaces and this has been tried many times in a variety of applications and research projects. (In fact, this is one way to circumvent the zero curvature issues at boundaries.) When extensive trimming is required, I’m not aware of a SubD based method that would be any better (overall) than simply using trimmed NURBS (like Rhino polysurfaces/breps).
It may be that a SubD could be used in an early phase of putting some type of surface over the hole, converting that SubD to NURBS, and then trimming the NURBS. To get robust results for complicated holes would take lots of work, testing, and refinement to provide a high quality user experience and a result that is at least as good as other NURBS patching techniques.
In short, SubDs are not a magic wand that will solve general case hard surface patching problems.