Inspired by looking at RnTypes.bindLHsTyVarBndr, and driven by #11592 (closed), Richard and I decided to make some improvements to the treatment of CUSKs.

Here's a scrappy Skype dump as a memory-jogger; it is a memory-jogger, not a full description. Richard has notes too

Richard Eisenberg: data F (x :: B a)
                   data G (x :: C b)
  You have written a complete user-suppled kind signature,
    but the following variable is undetermined: a0 :: A
    Perhaps add a kind signature.
    Inferred kinds of user-written variables:
      b :: B a0
      x :: C b

SPJ claim: CUSKs should be handled INDIVIDUALLY and in isoloation, just like term-level type signatures
SPJ claim: get rid of the "after the ::" side condition for data types (last bullet of HsDecls Note about CUSKs
All of this should mean (SPJ claims) no need for dep_vars returns by bindHsQTyVar, bindLHsTyVarBnrds etc
ToDo: clarify comments etc in bindLHStyVarBndr

Related to D2914