The following short program work in python 2.5, but does die with a
SystemError in python2.6:

~/ cat t.py
#! /usr/bin/env python

res = dict(__package__='foo')
exec "from os import path" in res
~/ python2.5 t.py 
~/ python2.6 t.py 
Traceback (most recent call last):
  File "t.py", line 4, in <module>
exec "from os import path" in res
  File "<string>", line 1, in <module>
SystemError: Parent module 'foo' not loaded

I think this has been introduced with svn revision 42649
(http://hgpy.de/py/trunk/rev/54c72e1678d68ebbf274)

Part of the diff reads:
        modules = PyImport_GetModuleDict();
        parent = PyDict_GetItemString(modules, buf);
        if (parent == NULL)
-               parent = Py_None;
+               PyErr_Format(PyExc_SystemError,
+                               "Parent module '%.200s' not loaded", buf);
        return parent;
        /* We expect, but can't guarantee, if parent != None, that:

Hmm, setting an invalid value for __package__ will definitely break
relative imports (see PEP-361), but it probably shouldn't be breaking
absolute imports.

One idea would be to change the import code to only produce a warning
for a missing __package__ entry instead of a SystemError (reserving the
SystemError for cases where the package name is derived from __name__
rather than being retrieved from the __package__ attribute).

Bumped priority - an existing module shouldn't crash in 2.6 just because
we started using __package__ as part of the import mechanism and that
module happens to already set an attribute by that name.

Adding to my personal to-do list for next beta.

Fixed in r64915.

The end result is that the import system now only emits a RuntimeWarning
instead of raising SystemError if it can't find the parent module when
attempting to perform an absolute import (regardless of whether the
parent module name was derived from __package__ or from __name__).
Explicit relative imports will still fail if __package__ is set
incorrectly and setting __package__ to a non-string value will still
result in a ValueError.

Thanks Nick for fixing this in a timely manner.
BTW I've seen this when trying to run doctests with py.test.