Johannes Schindelin [off-list ref] writes:

What I have in mind is more along Elijah's suggestion at the Contributor
Summit to use the `tmp_objdir*()` machinery. But instead of discarding the
temporary object database, the contained objects would be repacked and the
`.pack`, (maybe `.rev`) and the `.idx` file would then be moved (in that
order) before discarding the temporary object database.

That may be more involved but does indeed sound like an approach
more generally applicable.  Back when the bulk-checkin machinery was
invented, I envisioned that we would be adding annotations to
various codepaths so that object creation machinery can say "now we
are plugged, in anticipation for creating many objects at once" and
"now the flood of new object creation is done, time to wrap up" for
that kind of optimization.  

The callsites to {begin,end}_odb_transaction() functions haven't
grown beyond the original "add" and "update-index" (because the user
can add the entire working tree worth of files to the object
database), "unpack-objects" (because a fetch can bring in many
objects), and "cache-tree" (because a tree creation can cascade up
to create many objects), but I agree "merge" and "replay" are prime
candidates to benefit from the optimization of the same kind (so is
"fast-import").  They are about creating many objects at once, and
give us an opportunity for such an optimization.