I think I am now 100% sure that there is no switch to the primary pool from the fall-back pool not 30 min later, not 10 (which I set in --set-reconnection-timer 10), not at all.
I will give that a check.
[Device 0] Average speed (5s): 9188.96 sol/s
[Device 1] Average speed (5s): 9918.42 sol/s
[Device 2] Average speed (5s): 10140.2 sol/s
[Device 3] Average speed (5s): 10173.8 sol/s
[Device 4] Average speed (5s): 9941.01 sol/s
[Device 5] Average speed (5s): 10024.8 sol/s
[Device 6] Average speed (5s): 6902.02 sol/s
[Device 7] Average speed (5s): 6636.27 sol/s
[Device 8] Average speed (5s): 6775.64 sol/s
[Device 9] Average speed (5s): 6401.12 sol/s
[Device 10] Average speed (5s): 7138.37 sol/s
[Device 11] Average speed (5s): 6737.25 sol/s
[Device 12] Average speed (5s): 6953.81 sol/s
Maybe a overall high CPU load? This seems to me that the runtime can not fill the cards fast enough with new work. What can help is increasing --set-work-batch. I hope to get is less demanding on CPU side timing in next major version.