Previously, we had some unwrapped usage of git commit, which would
result in the current time being added to the commit object.
Our commit() shell function makes git commit deterministic. So
switch to that everywhere.
Because every call to commit() increments the counter for the current
time, inserting calls earlier in the test changes the hash of later
commits. So there was some test fallout associated with this. But no
meaningful behavior should have changed.
I'm not sure why the display order and the conversion order of some
things changed. If I had to venture a guess, it would be a) there
was a race condition in the previous code where commits with the same
second time were sorted differently from commits with different times
b) something somewhere was sorting by hash. I dunno.