[MERGE] More than double the speed of "bzr selftest"

[Martin Pool]

On 4/11/07, Robert Collins <robertc at robertcollins.net> wrote:
> Can I suggest two things - a test for this; and that tearDown is the
> right place to do this; I'm of the opinion (strongly held) that the test
> can delete everything whether or not it passed. We lose plenty of state
> when tearDown runs already; debugging requires debugging before tearDown already.

I realized when I saw this that that's probably why you redefined
run(), because cleanups run always.

I'm not sure I agree with Robert that failing tests can delete
everything - sometimes intermittent or unreproducible failures do
happen, and if you can't keep everything about them then keeping what
you can is sometimes useful.  But for our tests, keeping just the on
disk state is not so useful.

However, it should not delete everything if --keep-output is given.
Or at least, it would be confusing to have that option still present
and not actually have the advertised effect.

-- 
Martin