Let me summarize the issue as succinctly as possible without any formalism. If you want to read more details, see
https://arxiv.org/abs/1710.07212 (BW's paper), or
https://www.physicsforums.com/insights/wigners-friend/ (which I have updated accordingly) or
http://users.etown.edu/s/stuckeym/WignersFriendConsciousness.pdf (which we just submitted for a book on QM and consciousness). The source of inconsistency in FR (and the typical rendering of Wigner's friend) stems from contradictory assumptions, just as you suspected vanhees71. Classical systems obey Boolean algebra while quantum systems obey non-Boolean algebra. If you assume you have a quantum system obeying non-Boolean algebra that communicates with its classical environment (the universe) that obeys Boolean algebra, then you can end up with self-inconsistent classical information being shared among observers in the form of contradictory measurement outcomes. Simply put, this is precisely what FR did. Here is a bit more detail, if you're interested.
If one assumes that physics is used to model all shared, self-consistent classical information (we would call that "objective reality"), then any prediction of shared, self-inconsistent classical information would constitute a "scientific contradiction" in the language of BW and would invalidate the scientific theory making said prediction. There are two different QM formalisms -- the "standard" formalism with its measurement-update and Born rules (BW call this "objective collapse"), and the "relative-state" formalism with its universal unitary evolution (no measurement-update, which BW call "no collapse"). These formalisms produce the same predictions except in Wigner-friend-type experiments (observers measuring each other). In those experiments, the two formalism do make different predictions, so deciding which is correct is an empirical matter (good luck screening off macroscopic measuring devices to test it though). Neither of these formalisms produces a scientific contradiction for Wigner's friend and neither was used by FR (and neither is used in a typical rendering of Wigner's friend).
Instead, FR used what BW call "subjective collapse" (some would say that's an oxymoron of course and that's their point). In other words, FR assume a screened-off quantum system shares classical information with the universe and then shows how this leads to shared, self-inconsistent classical information between observers. Really?? Wow, how can that be?? [Sarcasm] Because you can't use non-Boolean algebra to model a system that shares classical information per Boolean algebra without generating contradictions. Go figure.
This is not intended to dismiss the importance of FR's paper, which has spawned many valuable discussions. BW unveiled inconsistencies in FR that are introduced in the typical rendering of Wigner's friend and are extremely important for understanding the foundations of QM. At the end of the day, if you want to avoid scientific contradiction, you have choices such as no collapse with its subjective reality (QM predictions are relative to the observer) or objective collapse with its objective reality (QM predictions are agreed upon by everyone). Or, you can do a hybrid where people's memories and written records are changed by quantum measurement (Bohmian relative-state approach per
https://dustinlazarovici.com/wp-content/uploads/comment_renner_new.pdf).
What FR do not show is that you must have Many Worlds, as Renner used to believe ("I have to admit that if you had asked me two years ago, I’d have said [our experiment] just shows that many-worlds is actually a good interpretation and you should give up" the requirement that measurements have only a single outcome, Renner said.
https://www.quantamagazine.org/frau...where-our-views-of-reality-go-wrong-20181203/). Or, that there is a limit to the applicability of QM (Renner, however, has changed his mind. He thinks the assumption most likely to be invalid is the idea that quantum mechanics is universally applicable. Same Quanta Magazine article).