Comment 3 for bug 1431649

One thing about the test here is that it can be misleading - we need to measure the actual wall time for the use case that matters to assess performance, because deferring importing code that is eventually used is just waste heat :) - its deferring loading of code that isn't needed that gives a win.

'keystone help' might be a useful metric, but I've in the past seen that optimised hugely without any change on things that actually matter like (for a concrete example) 'bzr st', which needed separate optimisation work.