I think, unfortunately, that there was a complex of issues that was causing this particular problem and it's not clear which fix or fixes fixed this one in particular. I do know that at least two problems were likely relevant:
* screen in the gate has been notoriously flakey about randomly not starting process and not reporting why. Screen is no longer being used on those runs, so that's probably helping
* in earlier devstacks the ordering of the compute agent and the collector were bad leading to notifications sometimes being missed
I think, unfortunately, that there was a complex of issues that was causing this particular problem and it's not clear which fix or fixes fixed this one in particular. I do know that at least two problems were likely relevant:
* screen in the gate has been notoriously flakey about randomly not starting process and not reporting why. Screen is no longer being used on those runs, so that's probably helping
* in earlier devstacks the ordering of the compute agent and the collector were bad leading to notifications sometimes being missed
Both of these things have been changed.