Yeah, That what I was worried for. And it was for nothing, it is only one extra fetch.
It is a good find in any case.
The approach however still suffers from the fact, that biasing is done based on shadow receiver, not shadow caster.
Biasing during shadow depth rendering based on shadow caster normal vs light vector does not distort the shadows that much, so it is still a warmly welcome feature, along with being able to set bias/ soft transition scale per each cascade.