t9106 failure, bisect weirdness

[carbonated beverage <ramune@net-ronin.org>]

Hi all,

Just checked out v1.5.3.6 from git, built it on a Debian/stable i386 box,
and got a failure in one of the test suites:

*** t9106-git-svn-dcommit-clobber-series.sh ***
*   ok 1: initialize repo
*   ok 2: (supposedly) non-conflicting change from SVN
*   ok 3: some unrelated changes to git
* FAIL 4: change file but in unrelated area

                test x"`sed -n -e 4p < file`" = x4 &&
                test x"`sed -n -e 7p < file`" = x7 &&
                perl -i -p -e 's/^4$/4444/' file &&
                perl -i -p -e 's/^7$/7777/' file &&
                test x"`sed -n -e 4p < file`" = x4444 &&
                test x"`sed -n -e 7p < file`" = x7777 &&
                git commit -m '4 => 4444, 7 => 7777' file &&
                git svn dcommit &&
                svn up tmp &&
                cd tmp &&
                        test x"`sed -n -e 4p < file`" = x4444 &&
                        test x"`sed -n -e 7p < file`" = x7777 &&
                        test x"`sed -n -e 58p < file`" = x5588 &&
                        test x"`sed -n -e 61p < file`" = x6611

*   ok 5: attempt to dcommit with a dirty index
* failed 1 among 5 test(s)
make[1]: *** [t9106-git-svn-dcommit-clobber-series.sh] Error 1
make[1]: Leaving directory `/home/ramune/src/git/git/t'
make: *** [test] Error 2
ramune/lycaeum:git:

Bisecting gives me:

fb159580a1628947f0a088e24cfe6fe4c81d99d0 is first bad commit
commit fb159580a1628947f0a088e24cfe6fe4c81d99d0
Author: Eric Wong <normalperson@yhbt.net>
Date:   Mon Nov 5 03:21:48 2007 -0800

    git-svn: t9114: verify merge commit message in test

    It's possible that we end up with an incorrect commit message
    in this test after making changes to fix the clobber bug
    in dcommit.

    Signed-off-by: Eric Wong <normalperson@yhbt.net>
    Signed-off-by: Junio C Hamano <gitster@pobox.com>

:040000 040000 1c4f3776c031f9531869e736b661a0559fab5ff6 4f4c5ecd6c79f7e9732b554e2f8dd536a8a8f2a6 M      t

Which seems odd -- the previous commit (marked good, because make test didn't
fail) was the one that modified t9106-git-svn-dcommit-clobber-series.sh.

I tried running the test manually and noticed something weird as well:

ramune/lycaeum:t: sh t9106-git-svn-dcommit-clobber-series.sh
*   ok 1: initialize repo
*   ok 2: (supposedly) non-conflicting change from SVN
*   ok 3: some unrelated changes to git
*   ok 4: change file but in unrelated area
* passed all 4 test(s)
ramune/lycaeum:t:
ramune/lycaeum:t: pwd
/home/ramune/src/git/git/t
ramune/lycaeum:t:
ramune/lycaeum:t: sh t9106-git-svn-dcommit-clobber-series.sh
*   ok 1: initialize repo
*   ok 2: (supposedly) non-conflicting change from SVN
*   ok 3: some unrelated changes to git
* FAIL 4: change file but in unrelated area

                test x"`sed -n -e 4p < file`" = x4 &&
                test x"`sed -n -e 7p < file`" = x7 &&
                perl -i -p -e 's/^4$/4444/' file &&
                perl -i -p -e 's/^7$/7777/' file &&
                test x"`sed -n -e 4p < file`" = x4444 &&
                test x"`sed -n -e 7p < file`" = x7777 &&
                git commit -m '4 => 4444, 7 => 7777' file &&
                git svn dcommit &&
                svn up tmp &&
                cd tmp &&
                        test x"`sed -n -e 4p < file`" = x4444 &&
                        test x"`sed -n -e 7p < file`" = x7777 &&
                        test x"`sed -n -e 58p < file`" = x5588 &&
                        test x"`sed -n -e 61p < file`" = x6611

* failed 1 among 4 test(s)

So it succeeds once, then fails.

/bin/sh is dash 0.5.3-7 from Debian.

Any additional information needed?

-- DN
Daniel