Inferred type not principle/most general?

The type inferred for (#) is H a a -> H a a -> H a a (from Coroutining Folds with Hyperfunctions)

newtype H a b = Fn { invoke :: H b a -> b }

-- ( # ) :: H a a -> H a a -> H a a
f # g = Fn (\k -> invoke f (g # k))

but a more general type is actually

( # ) :: H b c -> H a b -> H a c
f # g = Fn (\k -> invoke f (g # k))

Is this expected or a blind spot in type inference

---

Same thing in ghci

$ ghci -ignore-dot-ghci 
GHCi, version 8.3.20170920: http://www.haskell.org/ghc/  :? for help
Prelude> :set prompt "> "
> 
> newtype H a b = Fn { invoke :: H b a -> b }
> 
> f # g = Fn (\k -> invoke f (g # k))
> :t ( # )
( # ) :: H a a -> H a a -> H a a
>
> let ( # ) :: H b c -> H a b -> H a c; f # g = Fn (\k -> invoke f (g # k))
> :t ( # )
( # ) :: H b c -> H a b -> H a c

Interestingly writing it in terms of fix does not type check with the more general type

> import Data.Function 
> :t fix $ \self -> \f g -> Fn (\k -> invoke f (self g k))
fix $ \self -> \f g -> Fn (\k -> invoke f (self g k))
  :: H a a -> H a a -> H a a
> :t (fix $ \self -> \f g -> Fn (\k -> invoke f (self g k))) :: H b c -> H a b -> H a c

<interactive>:1:2: error:
    • Couldn't match type ‘b1’ with ‘a1’
      ‘b1’ is a rigid type variable bound by
        an expression type signature:
          forall b1 c1 a1. H b1 c1 -> H a1 b1 -> H a1 c1
        at <interactive>:1:60-82
      ‘a1’ is a rigid type variable bound by
        an expression type signature:
          forall b1 c1 a1. H b1 c1 -> H a1 b1 -> H a1 c1
        at <interactive>:1:60-82
      Expected type: H b1 c1 -> H a1 b1 -> H a1 c1
        Actual type: H a1 a1 -> H a1 a1 -> H a1 a1
    • In the expression:
          (fix $ \ self -> \ f g -> Fn (\ k -> ...)) ::
            H b c -> H a b -> H a c

Trac metadata

Trac field	Value
Version	8.2.1
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture