Test failures in pu: 4046, 950[012]

Failures as of yesterday (it alternates between next and pu, so today was next which was clean):

* 4046 has this valgrind failure:

  ==24237== Source and destination overlap in mempcpy(0x7feffed50, 0x7feffed50, 2)
  ==24237==    at 0x4C228DB: mempcpy (mc_replace_strmem.c:802)
  ==24237==    by 0x511AD6: copy_dirname (diffcore-rename.c:573)
  ==24237==    by 0x512071: check_one_bulk_move (diffcore-rename.c:744)
  ==24237==    by 0x5120C0: diffcore_bulk_moves (diffcore-rename.c:756)
  ==24237==    by 0x512988: diffcore_rename (diffcore-rename.c:961)
  ==24237==    by 0x4A48EF: diffcore_std (diff.c:4252)
  ==24237==    by 0x4974B5: run_diff_index (diff-lib.c:475)
  ==24237==    by 0x4267CD: cmd_diff_index (diff-index.c:50)
  ==24237==    by 0x404963: run_builtin (git.c:276)
  ==24237==    by 0x404AEE: handle_internal_command (git.c:435)
  ==24237==    by 0x404BD9: run_argv (git.c:479)
  ==24237==    by 0x404D3A: main (git.c:554)
  ==24237== 
  [snip suppression template]
  not ok - 14 diff-index --detect-bulk-moves on a move including a subdir.

* 950[012]: these are very strange.

Running them interactively appears to work ok, so maybe it is some environment bug. They appear to be "can't happen" style failures though, since they all look like (this one is from 9502)

  Initialized empty Git repository in /local/home/trast/git/t/trash directory.t9502-gitweb-standalone-parse-output/.git/
  expecting success: 
          test_commit first foo &&
          git branch xx/test &&
          FULL_ID=$(git rev-parse --verify HEAD) &&
          SHORT_ID=$(git rev-parse --verify --short=7 HEAD)

  fatal: This operation must be run in a work tree
  not ok - 1 setup

They are the only ones failing with this problem, though. They all bisect to

  commit e60cb38c7c4e0bdb8f9542f9925ba5cafd4cd33b
  Author: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
  Date:   Fri Oct 29 13:48:50 2010 +0700

      Remove all logic from get_git_work_tree()
      
      This helps break things some more. Eventually all repo setup logic
      should be centralized in setup_* functions. This helps push all the
      bugs back to setup_* functions.
      
      Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
      Signed-off-by: Junio C Hamano <gitster@pobox.com>

(Sorry, the UTF was garbled in cron-email transit...)