Ah, great counterexample! StrictData is another. Do you know any others that change language semantics non-trivially? I would roughly define "non-trivially" to mean "correct programs become incorrect".

By the way, I'm particularly interested because I am writing about Haskell myths. I have a very rough draft at http://h2.jaguarpaw.co.uk/posts/haskell-myths/

There’s also QuasiQuoter which can make valid code with list comprehensions not compile, but it doesn’t just compile it, it gets stuck looking for the end tag. Not sure if that qualifies :P