Squashed three or four more bugs today. Unanswered message backlog is down
to 27.

The most interesting problem today is that the git-repair code was using
too much memory when `git-fsck` output a lot of problems (300 thousand!). I
managed to half the memory use in the worst case (and reduced it much more
in more likely cases). But, don't really feel I can close that bug yet,
since really big, really badly broken repositories can still run it out of
memory. It would be good to find a way to reorganize the code so that the
broken objects list streams through git-repair and never has to all be
buffered in memory at once. But this is not easy.