We probably shouldn't leave the importer service running in a state
where a restart or reboot could change behaviour, so we should restart
the service as part of an update.
It turns out that "git format-patch --stdout ... | git am" does not
correctly round-trip in every case. If a commit message contains
something that looks like a patch, "git am" incorrectly tries to apply
it.
Instead, we can note the commits that need to be rebased at export time
with "git rev-list --reverse", and apply them at import time with "git
cherry-pick". This keeps everything "inside git" instead of
round-tripping through an mbox, working around the problem.
This does mean that the repository to which the import happens must have
the commits available, which commonly would mean that the export
destination repository must be the same as the import source repository.
As it happens, this is always the case for us anyway during a
--reimport, so this will work just fine.
Test rich history preservation unescaped patch bug
If a rich history commit _message_ contains something that looks like a
patch, the current implementation will attempt to apply it even though
it shouldn't. This test checks this case, but fails for now as the bug
exists, so we mark it xfail to verify that the test works.
We will shortly be changing rich history preservation to use
cherry-picks to fix a bug. This will require the same repository to be
used for rich history import as the commits being imported will need to
be present in the repository still. Tests will therefore need to use the
same repository for import as they did for export.
This does rely on the test cleanly and correctly removing all previous
relevant tags in order to properly test the import, making it a little
more error prone, but we have no choice given the new requirement.
Since this was previously the only use for our repo_factory fixture, we
can drop it from the code base, together with its documentation. If it
is needed again later, we can always pull it out of VCS again.
The review command was introduced experimentally during the pre-1.0
development period, however it is not believed to be widely used yet,
lacks test cases, and does not have a complete implementation.
