But that usually applies more in the case of the classical BFT protocols (PBFT and alike), doesn't it?. Doesn't Steem follow the classical 51% threshold we see in most PoS and PoW?

I don't think there has ever been a real proof, and I vaguely recall some reasonable argument it was actually 33%. But I may be misremembering or wrong.

I would have to do the math behind it, but I feel 20 would be the worst possible threshold, and 30 also not being very great.

If you have further analysis I'm interested to see it. In my view 30 being well over 20 is probably somewhat okay because it allows coverage to all acceptable primaries (which should tend to overlap a lot among reasonable stakeholders) as well as some backups. But without a stronger argument than that, confidence can't be too high. If we could gain significantly by increasing 30 to say 50, I don't think the downsides would be that serious and we should consider it.