The solution is to always call

async { … }

from a

coroutineScope { … }

lambda, so it throws the exception that led to failing, and you can catch it there without having the upper parent scope being cancelled. I finally understood the whole structured concurrency behavior!

I think there's not enough docs on that