This is a significant problem for even some of the more ubiquitous extensions. For instance, there are multiple compilers that implement RankNTypes, but I would not be surprised at all if programs using that extension were not portable across implementations (they're not really even portable across GHC versions).
The problem is that RankNTypes is not just about the fact that you are allowed to use such types; every compiler must decide on an inference algorithm that incorporates such types while defaulting to Hindley-Milner. But, there are several such algorithms, and they have different trade offs as far as where annotations must be placed, or even whether certain conceivably well-typed terms are type checkable (for instance, GHC used to do some level of impredicative instantiation; forall a. a -> a could be instantiated to (forall a. a) -> (forall a. a); but this no longer works).
So, even if we have ubiquitous agreement on the fact that RankNTypes are useful, and implementable, we don't have ubiquitous agreement on the algorithms for implementing them, and which set of trade offs to make. But any standard would have to nail all that down, or else programs won't be portable.
And this is not the only extension for which this kind of thing is an issue.
-- Dan