Marshal does not round-trip unicode surrogate pairs for wide unicode-builds:

marshal.loads(marshal.dumps(u"\ud800\udc00")) == u'\U00010000'

This is very annoying, because the size of unicode constants differs
between when you run a module for the first time and subsequent runs
(because the later runs use the pyc file).

I think this is unavoidable. Depending on whether you happen to be using
a narrow or wide unicode build of Python, \Uxxxxxxxx may be turned into
a pair of surrogates anyway. It's not just marshal that's not
roundtripping; the utf-8 codec has the same issue (and so does the
utf-16 codec I presume). You will have to code around it. I think that
the alternative would be more painful in other circumstances.

As Guido says: this is by design. The Unicode type doesn't really
support storage of surrogates; so don't use it for that.

I think you have a wrong understanding of round-tripping. 

In Unicode it is really irrelevant if you're using a UCS2 surrogate pair
or a UCS4 representation to describe a code point. The length of the
Unicode representation may change, but the meaning won't, so you don't
lose any information.