I don't see a lot of reasons to differentiate the category based on how the instruction is encoded for other than re-encoding which is already handled by different data and logic, from an exploration perspective its better if those instructions are categorized by their signature/class/menmonic.
Instructions like vpclmulqdq might report a different instruction category based on the encoding rather then semantics.
Example: https://gist.github.com/ZehMatt/8e52b8d01fc110b113e57ca401bf18a1/revisions , another instruction I found is also vpclmulqdq, there might be more.
I don't see a lot of reasons to differentiate the category based on how the instruction is encoded for other than re-encoding which is already handled by different data and logic, from an exploration perspective its better if those instructions are categorized by their signature/class/menmonic.