PmCheck: Long-distance info and where clause "refactorings" are strange

Consider the following two programs:

{-# OPTIONS_GHC -Wincomplete-patterns -fforce-recomp #-}

module Lib where

f :: Maybe Int -> Int -> Int
f x y = case x of
  Nothing -> 1
  _       -> g y
  where
    g y = case x of
      Just n -> n + y

{-# OPTIONS_GHC -Wincomplete-patterns -fforce-recomp #-}

module Lib where

f :: Maybe Int -> Int -> Int
f x y = case x of
  Nothing -> 1
  _       -> g y
    where
      g y = case x of
        Just n -> n + y

The first warns that the case match within g is incomplete with GHC HEAD, the second doesn't.

It's a really delicate interaction between where clause syntax and very accurate long-distance information, maybe even too accurate for its own good: The first program syntactically associates the where clause with the FunDef of f (I think), while the second one associates the where clause with the _ case alt. In the latter case, we know that x can't be Nothing, but not so in the former case. Indeed, it's not unreasonable to call g from the Nothing case in the first program:

module Lib where

f :: Maybe Int -> Int -> Int
f x y = case x of
  Nothing -> g 1
  _       -> g y
  where
    g y = case x of
      Just n -> n + y

And hence it's correct to warn there. I know Haskell is indentation-sensitive, but this kind of interaction was very surprising to me. I'll just open this issue for future reference, as I'm sure that I won't be the last person stumbling over this.