The log-karaoke/stages.js is the most flaky part of the ATH.

It seems quite sensitive to timing, and isn't the easiest to work out what it is doing, and what exactly failed.

From what I can tell, jobs are created and run in parallel, and then the stages are run through inspecting the results. There should be no assumption about how jobs run in parallel in the ATH - they should be run stand alone as the number of executors may vary. I have seen tests (such as stage 04) that when it gets to the stage, it has already completed and thus the assertions are meaningless at that point and fail.

This is a most useful set of tests but needs a good cleanup as it reduces the confidence in the ATH and is impacting productivity badly.