GHCi accepts invalid programs when recompiling

This ticket may or may not be a duplicate of #9729, but I ran into it again in a different context, so I'm reporting it. My GHC version is 7.10.2.20151030. The example is as minimal as I could make it.

File Bar.hs

{-# LANGUAGE FlexibleInstances, MultiParamTypeClasses #-}

module Bar where

class Bar a b where
  (*^) :: a -> b -> b

File Foo.hs

{-# LANGUAGE MultiParamTypeClasses #-}

module Foo where

import Bar
import GHC.Prim

newtype Foo fp = Foo [fp]

instance {-# OVERLAPS #-}
  Bar (Foo fp) [fp]

File Main.hs

{-# LANGUAGE MultiParamTypeClasses #-}

import Foo
import Bar

newtype LW a = LW [a]

instance Bar (Foo fp) (LW fp) where
  r *^ (LW xs) = LW $ r *^ xs

If I start GHCi with ghci Main, everything compiles as it should. Then I remove the {-# OVERLAPS #-} pragma in Foo.hs, which should break Main.hs. But GHCi happily recompiles with :r.

A couple of notes: *with* the pragma, :i Bar gives the instances

instance Bar (Foo fp) (LW fp) -- Defined at Main.hs:8:10
instance Bar a b => Bar a [b] -- Defined at Bar.hs:8:10
instance [overlap ok] Bar (Foo fp) [fp] -- Defined at Foo.hs:22:3

while after (successfully, but incorrectly) recompiling without the pragma, I get the instances

instance Bar (Foo fp) (LW fp) -- Defined at Main.hs:8:10
instance Bar a b => Bar a [b] -- Defined at Bar.hs:8:10
instance Bar (Foo fp) [fp] -- Defined at Foo.hs:22:3

(i.e., without the [overlap ok]).

Another strange thing is that the bug is only triggered if I import specific (superfluous) modules in Foo.hs. For example, if I replace GHC.Prim with Control.Monad or Data.Maybe, the bug is not triggered (i.e., GHCi correctly detects when I remove the pragma).

Trac metadata

Trac field	Value
Version	7.10.2-rc2
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	GHCi
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture