We say an inductive type T is reflexive if it contains at least one constructor that
takes as an argument a function returning T.
For reflexive types it doesn't seen to be possible to define a single
brec_on that can eliminate to Type.{>=1} and Prop.
The universe level expressions get too complicated.
Even if we extend the universe constraint solver in the kernel, the
additional complexity might be a problem.
We workaround this issue by defining two versions of brec_on:
- One (brec_on) that eliminates to Type.{>=1}, and
- binduction_on that eliminates to Prop.
For non-reflexive types, we can combine both of them.
