Mystery glitch in new pythonpath system under Python 2.6 in testsuite

[Max Bowsher]

I'm seeing weird behaviour from the testsuite when running under Python 2.6. Reproduce as follows:

bzr branch lp:~launchpad-committers/launchpad/python2.6
make check VERBOSITY=--list-tests #evilhack to just do the preparation for testing

Now run just the unit tests:
bin/test -vv -u

The following tests fail:
    make_test_tarball_2 (canonical.launchpad.tests.test_helpers)
    test_script (lp.translations.scripts.tests.test_validate_translations_file.TestValidateTranslationsFile)
    lib/lp/translations/tests/../doc/remove-upstream-translations-script.txt

The first is unrelated, the other two are failing due to RuntimeError: The wrong site.py is imported. Scripts should usually be started with Launchpad's bin/py, or with a Python invoked with the -S flag.

Now here's the really weird bit. Run exactly the same test run, but omit test_all_scripts:
bin/test -vv -u -t '!test_all_scripts'

When you do this, the two tests mentioned no longer fail. (Nor do they fail if run individually.)

[Max Bowsher (maxb)]

(
This bug was filed out of a #launchpad-foundations IRC conversation:
14:15 < maxb> gary_poster: Hello. I have a buildoutish conundrum for you :-) Under python 2.6, a couple of rosetta scripts throw the "bin/py or -S" RuntimeError during a `bin/test -u` but NOT if run individually or as `bin/test -u
              -t!test_all_scripts`. I don't expect you'll have a magical solution, but I thought I'd mention it nonetheless
14:19 < gary_poster> maxb, hm. I probably can tweak it one way or another. Would you mind making a foundations bug with the details for me?
14:20 < maxb> sure
)

[William Grant (wgrant)]

Fixed in devel r10711. run_script was accidentally popping PYTHONPATH out of os.environ. Before Python 2.6, this didn't actually unset the variable in the current environment.