An odd bug in OS X 10.4 causes open(2) on a non-existent,
invalid-encoded filename to return errno 22, EINVAL: Invalid argument,
instead of the expected errno 2, ENOENT: No such file or directory,
*if* the containing directory is not empty.  That caused frequent
failures when running the buildbot tests on 10.4 depending on the state
of the test working directory.  The failure is easy to reproduce on
10.4 by running the test directly (not with regrtest), first in an empty
directory, then after adding a file to it.  The fix is to check for and
pass if either errno is returned.

Backported from patch by Jeremy Kloth.

when running scripts that terminated by raising SystemExit
(i.e. unittest and turtledemo).

Also add running of test_email_renamed to the email regrtest.  It contains
tests that the base email/tests/test_email.py does not, which I discovered
while trying to backport this test for confirmation of the behavior.

a stream larger than 2 GiB on 64-bit systems.

* avgpp() and maxpp() no more crash on empty and 1-samples input fragment. They now work when peak-peak values are greater INT_MAX.
* ratecv() no more crashes on empty input fragment.
* Fixed an integer overflow in ratecv().
* Fixed an integer overflow in add() and bias() for 32-bit samples.
* reverse(), lin2lin() and ratecv() no more lose precision for 32-bit samples.
* max() and rms() no more returns negative result for 32-bit sample -0x80000000.
* minmax() now returns correct max value for 32-bit sample -0x80000000.
* avg(), mul(), tomono() and tostereo() now round negative result down and can return 32-bit sample -0x80000000.
* add() now can return 32-bit sample -0x80000000.

input buffer.

The main docs were fixed to remove mention of None long ago,
but the docstring was not.  Reported by Armin Rigo, patch
by Ian Cordasco.

parses nested mutating sequence.

The customize_compiler function moved many times during the 2.7 series;
in 2.7.3, setup scripts importing this function from ccompiler were
broken.  This commit restores compatibility without reintroducing the
issue that #13994 originally fixed (duplication of the function).

A unit test makes little sense here, as distutils tests never do imports
in functions, and the fix is very simple.

stream or a decoder produces data of an unexpected type (i.e. when
io.TextIOWrapper initialized with text stream or use bytes-to-bytes codec).

deployment target prior to 10.6.

Issue #15633: httplib.HTTPResponse is now mark closed when the server sends less than the advertised Content-Length.