Given the limited set of required encodings for a conforming DFDL processor, I believe dfdl:utf16Width='variable' should be an optional feature.
That's just consistency with what is optional already. But it is also quite hard to implement.
There are other situations that are very hard to implement, probably never used by real users, yet which are non optional in the spec:
I would suggest that dfdl:lengthKind='explicit' for elements of complex type, with dfdl:lengthUnits='characters' and a variable-width encoding like utf-8 is very problematic to implement. I am pretty sure IBM DFDL has no implementation of this per email threads, and I know I don't want to implement this in Daffodil even though we're trying to be very comprehensive in the implementation eventually.
I think implementations should be free to just not implement this. These sorts of cases often exist just because we're trying to preserve some orthogonality of composition in the language. So it's possible to do quite a few things that probably aren't ever needed by anyone, that reflect ill-defined data formats, etc.