In the Python 2.x branch, os.path.normpath will sometimes return a str
even if given a unicode. This is not an issue in the Python 3.0 branch.
This happens specifically when it throws away all string data and
constructs its own:
>>> os.path.normpath(u'')
'.'
>>> os.path.normpath(u'.')
'.'
>>> os.path.normpath(u'/')
'/'
This is a problem if working with code which expects all strings to be
unicode strings (sometimes, functions raise exceptions if given a str,
when expecting a unicode).
I have attached patches (with test cases) for posixpath and ntpath which
correctly preserve the unicode-ness of the input string, such that the
new behaviour is:
>>> os.path.normpath(u'')
u'.'
>>> os.path.normpath(u'.')
u'.'
>>> os.path.normpath(u'/')
u'/'
I tried it on os2emxpath and plat-riscos/riscospath (the other two
OS-specific path modules), and it already worked fine for them.
Therefore, this patch fixes all necessary OS-specific versions of os.path.

Thanks for the patch, I tried it on Linux and it seems to solve the problem.
A few comments about it:
1) I'd change all the self.assertEqual(type(posixpath.normpath(u"")),
unicode) to self.assertTrue(isinstance(posixpath.normpath(u""), unicode));
2) a test for normpath(u'.') should probably be added;
3) in ntpath.py the 'slash' is actually a backslash, so the name of the
var should be changed;

Thanks Ezio.
I've updated the patch to incorporate your suggestions.
Note that I too have only tested it on Linux, but I tested both
posixpath and ntpath (and there is no OS-specific code, except for the
filenames themselves).
I'm not sure if using assertTrue(isinstance ...) is better than
assertEqual(type ...), because the type equality checking produces this
error:
AssertionError: <type 'str'> != <type 'unicode'>
while isinstance produces this unhelpful error:
AssertionError: False is not True
But oh well, I made the change anyway as most test cases use isinstance.

assertTrue() also accepts a 'msg' argument where to explain what went
wrong in case of failure [1].
[1]:
http://docs.python.org/library/unittest.html#unittest.TestCase.assertTrue

Also, assertTrue has an alias failUnless which I personally find more
descriptive (I don't know if either form is preferred for inclusion in
Python though).

failUnless is deprecated in Python3.1 [1]. The assert* methods are
preferred over the fail* ones that are now deprecated.
[1]:
http://docs.python.org/3.1/library/unittest.html#unittest.TestCase.failUnless

Fixed in r77442 (trunk) and r77443 (release26-maint), thanks!

stage: patch review -> resolved

messages: + msg95459

nosy: + sandberg, loewis, kcwu, ezio.melotti
messages: + msg95223

stage: patch review