So the rabbit whole lead me to a certain race condition in nodes communication. There was a lot of digging and eventually I was able to expose this an actually issue in a reproducible test. All tracked and already fixed in this PR. The fix would be deployed in the next release.
In advance: I have no proof this is what we are facing on main net. It is extremely hard to debug in such distributed environment, especially we don’t have any access to the interesting logs. What I am sure is that this is a legit bug scenario and it would be fixed on main net soon.