The leak can be reproduced with the minimum attached test_os.py:

C:\haypo\python>PCbuild\amd64\python_d.exe -m test -R 3:3 test_os
Run tests sequentially
0:00:00 [1/1] test_os
beginning 6 repetitions
123456
......
test_os leaked [2, 2, 2] memory blocks, sum=6
test_os failed

1 test failed:
    test_os

Total duration: 172 ms
Tests result: FAILURE

The memory leak is in os_spawnv_impl and os_spawnve_impl in Modules/posixmodule.c. The call fails with a ValueError when the first argument in the argv list is an empty string, in which case these functions both mistakenly pass i (0) to free_string_array() as the count of strings to free. But the conversion was successful, so it needs to include the current string in the count, i.e. `i + 1`.

The fix for this could also address the following two issues, which are mostly cosmetic in nature. os_spawnve_impl needs its TypeError message to be special cased the same as in os_spawnv_impl, i.e. "spawnve() arg 2 must contain only strings". Currently it uses the default message from the failed conversion: "expected str, bytes or os.PathLike object, not %.200s". Also, its ValueError message needs to reference "spawnve()" instead of "spawnv()".

Eryk: do you want to work on a PR to fix it? It seems like you understand well the bug!

New changeset 526b22657cb18fe79118c2ea68511aca09430c2c by Victor Stinner in branch 'master':
bpo-30602: Fix refleak in os.spawnve() (#2184)
https://github.com/python/cpython/commit/526b22657cb18fe79118c2ea68511aca09430c2c

Oh, I fixed a leak in test_spawnve_noargs() of test_os, but test_spawnv_noargs() still leaks :-) os.spawnv() is very similar to os.spawnve(). https://github.com/python/cpython/pull/2212 fixes os.spawnv() as well.

I tested that "python -m test -R 3:3 test_os" doesn't leak anymore with this second fix.

New changeset 8acb4cf2b3436652568d7a70228b166316181466 by Victor Stinner in branch 'master':
bpo-30602: Fix refleak in os.spawnv() (#2212)
https://github.com/python/cpython/commit/8acb4cf2b3436652568d7a70228b166316181466

It seems like I introduced a bug:
https://github.com/python/cpython/pull/2184#pullrequestreview-44308536

The fixes must be backported to Python 3.6, test_os still leaks references on this branch:

test_os leaked [6, 6, 6] memory blocks, sum=18

New changeset c8d6ab2e25ff212702d387e516e258b1d8c52910 by Victor Stinner in branch 'master':
bpo-30602: Fix lastarg in os.spawnve() (#2287)
https://github.com/python/cpython/commit/c8d6ab2e25ff212702d387e516e258b1d8c52910

New changeset c472fb6b2744b36c7a0823c20e0d5ac9be3ea623 by Victor Stinner in branch '3.6':
bpo-30602: Fix lastarg in os.spawnve() (#2287) (#2357)
https://github.com/python/cpython/commit/c472fb6b2744b36c7a0823c20e0d5ac9be3ea623

Thank you Eryk Sun for the careful reviews! All known issues on os.spawn*() on Windows are now be fixed, so I closed this issue.

New changeset b78fbaaeab9df8cfbbdae3d5faf2d1537d73e43b by Victor Stinner in branch '3.6':
bpo-30602: Fix refleak in os.spawnv() (#2212) (#2486)
https://github.com/python/cpython/commit/b78fbaaeab9df8cfbbdae3d5faf2d1537d73e43b