homepage

I haven't worked out how yet, but importlib.machinery is managing to bypass the replacement of importlib._bootstrap with _frozen_importlib:
Python 3.3.0b1 (default:8bf691d0b004+, Jul 15 2012, 23:20:06) 
[GCC 4.7.0 20120507 (Red Hat 4.7.0-5)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import importlib._bootstrap
[74500 refs]
>>> import importlib.machinery
[74500 refs]
>>> importlib.machinery.FileFinder
<class 'importlib._bootstrap.FileFinder'>
[74505 refs]
>>> importlib._bootstrap.FileFinder
<class '_frozen_importlib.FileFinder'>
[74505 refs]
>>> importlib.machinery.FileFinder.path_hook()("")
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
 File "/home/ncoghlan/devel/py3k/Lib/importlib/_bootstrap.py", line 1350, in path_hook_for_FileFinder
 if not _path_isdir(path):
 File "/home/ncoghlan/devel/py3k/Lib/importlib/_bootstrap.py", line 117, in _path_isdir
 path = _os.getcwd()
NameError: global name '_os' is not defined
[74566 refs]
>>> importlib._bootstrap.FileFinder.path_hook()("")
FileFinder('.')

I would have said that it wasn't because it was just the class type not picking up on the __name__ re-assignment, but that global name failure states otherwise.

Yup. A simpler demonstration:
Python 3.3.0b1 (default:8bf691d0b004+, Jul 15 2012, 23:20:06) 
[GCC 4.7.0 20120507 (Red Hat 4.7.0-5)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from importlib import _bootstrap, machinery
>>> _bootstrap.FileFinder is machinery.FileFinder
False
My initial theory (a missing sys.modules entry) failed miserably, as import.__init__ already has the following line:
 sys.modules['importlib._bootstrap'] = _bootstrap
I'm guessing it has to be some weird interaction with the explicit relative import, but I would have expected that to pick up on either the parent module attribute or the sys.modules entry.

Checking with importlib.abc and importlib.util (which turned out not to have the problem) put me on the right path: the problem is the import of "imp" in importlib.__init__
What appears to be happening is that the interpreter sees the partially initialised importlib package, considers it good enough, and continues on with importing _bootstrap and machinery from source. Control eventually returns to __init__ and it overwrites the source version with the frozen version.
Changing the __init__ file to do "import _imp" instead (the same as what import_init does in the C code) was enough to fix it.
No patch yet - the actual fix is trivial, but it needs a test.

Nice catch! Obviously that needs a big comment to avoid that problem in the future.

New changeset 4431dc4bb770 by Nick Coghlan in branch 'default':
Close #15386: There was a loophole that meant importlib.machinery and imp would sometimes reference an uninitialised copy of importlib._bootstrap
http://hg.python.org/cpython/rev/4431dc4bb770 

That was a *lot* harder than I expected to create a test for - I had to import importlib right at the start of regrtest, as well as tweak the import order in runpy, and doing so actually caused test_import to *crash* completely without the bug fixed.
The trick was realising that *anything* importing imp before importlib would mask the bug.

New changeset 8b7db5a1114e2113a756bdf8877fbe366055c69a by Victor Stinner in branch 'master':
bpo-37473: Don't import importlib ASAP in tests (GH-14661)
https://github.com/python/cpython/commit/8b7db5a1114e2113a756bdf8877fbe366055c69a