Judging by where it fails, the filleting itself in OpenCascade has no problems keeping track of edges, where it seems to have difficulty is 1) saddle points, and 2) where the fillet size makes it intersect with a feature it's not responsible for.

A fillet breaking when a parameter update changes the topology is just a specific example of the topological naming problem. I don't think that's specific to OpenCascade. In fact, just thinking it through, I wouldn't be at all surprised if 2) above is the same thing: the fillet itself changes the object topology enough that the basis edges for the fillet get lost. If so that's a Freecad thing rather than an OpenCascade thing.

Maybe someone who's worked in that area can comment, it might be one of those things that gets better in 0.21.