(warning) Tonight's 'pu' may fail its tests

I usually try to make sure all four integration branches (maint, master,
next and pu) pass the selftest before pushing the results out, but it
seems that there is a subtle breakage somewhere in 'pu' that breaks
t4203-mailmap.sh.  The breakage is not reliably reproducible but happens
very often when tests are run in parallel (e.g. "make -j4").

I originally was planning to merge ms/mailmap to 'next' before tonight's
pushout, but I ran out of time and had to yank the topic out of 'next' and
kicked it back to 'pu'.

Does the following log ring a bell to anybody?

----------------------------------------------------------------
Initialized empty Git repository in /git/git.git/t/trash directory.t4203-mailmap/.git/
* expecting success:
        echo one >one &&
        git add one &&
        test_tick &&
        git commit -m initial &&
        echo two >>one &&
        git add one &&
        git commit --author "nick1 [off-list ref]" -m second

[master (root-commit) 3a2fdcb] initial
 1 files changed, 1 insertions(+), 0 deletions(-)
 create mode 100644 one
[master 7de6f99] second
 1 files changed, 1 insertions(+), 0 deletions(-)
*   ok 1: setup

* expecting success:
        git shortlog >actual &&
        test_cmp expect actual

--- expect      2009-02-06 10:47:25.000000000 +0000
+++ actual      2009-02-06 10:47:25.000000000 +0000

@@ -1,6 +0,0 @@
-A U Thor (1):
-      initial
-
-nick1 (1):
-      second
-

* FAIL 2: No mailmap

                git shortlog >actual &&
                test_cmp expect actual

make[2]: *** [t4203-mailmap.sh] Error 1
----------------------------------------------------------------

The frustrating thing is that

    GIT_SKIP_TESTS='t[0-35-9]??? t4[01]??' \
    GIT_TEST_OPTS='-i -v' \
    make -j4 test

fails very reliably with the above log, but:

 (1) running the same "../../git-shortlog" in  "t/trash directory.t4203-mailmap/"
     immediately after the test failed does emit the expected result;

 (2) going down to t/ and manually running t4203-mailmap.sh alone succeeds;

 (3) running the tests serially (i.e. without -j4) allows the test to succeed.