Hi,
Have you tried to use a different approach and use something with strategy host pinned, serial and forks >=2x serial?
I read your post very fast but i think that something can be done in that way, you can simulate a sort of queue, where an host will performed all tasks in serial because is pinned.
When it finished will leave one slot in the queue and new host will proceed.
Fro example6 hosts, strategy host_pinned, serial 3, forks 6(to be sure that queue is availabe). Play will start in the first 3 hosts, all three hosts procede in each task indipendently. Suppese that 2 host require more time to oerfomr one task but the third hosta complete all tasks.
Instead wait all other 3 hosts, one slot will be availabe and fourfth host proceed.
In any case, for my opinion and expertise, awx need to be considered as “orchestration” only for not much complex scenario.
For flow that require a lot of complex strategy, probably is not the simplest tool and require a lot of 
to elaborate ad orchestrate everything.
Good luck