DeriveAnyClass regression involving a rigid type variable

I was writing up an example to show off how DeriveAnyClass has improved since 639e702b, and wouldn't you know it, the example doesn't actually compile anymore post-639e702b. Oopsie.

{-# LANGUAGE DefaultSignatures #-}
{-# LANGUAGE DeriveAnyClass #-}
{-# LANGUAGE DeriveGeneric #-}
{-# LANGUAGE FlexibleContexts #-}
{-# LANGUAGE ScopedTypeVariables #-}
{-# LANGUAGE TypeFamilies #-}
module TypeName where

import GHC.Generics

class TypeName a where
  typeName         :: forall proxy.
                      proxy a -> String
  default typeName :: forall proxy d f.
                      (Generic a, Rep a ~ D1 d f, Datatype d)
                   => proxy a -> String
  typeName _ = gtypeName $ from (undefined :: a)

gtypeName :: Datatype d => D1 d f p -> String
gtypeName = datatypeName

data T a = MkT a
  deriving (Generic, TypeName)

This compiles before that commit. After it, however, it fails with the error:

[1 of 1] Compiling TypeName         ( Bug.hs, interpreted )

Bug.hs:23:22: error:
    • Couldn't match type ‘f’
                     with ‘C1
                             ('MetaCons "MkT" 'PrefixI 'False)
                             (S1
                                ('MetaSel
                                   'Nothing 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy)
                                (Rec0 a))’
        arising from the 'deriving' clause of a data type declaration
      ‘f’ is a rigid type variable bound by
        the deriving clause for ‘TypeName (T a)’ at Bug.hs:14:38
    • When deriving the instance for (TypeName (T a))
   |
23 |   deriving (Generic, TypeName)
   |                      ^^^^^^^^

I'm not sure why it's complaining only about f and not, say, d.