The only flexibility is that is easier to implement, whence why it is usually the first GC algorithm one learns, and among the initial chapters of GC algorithms overview books.

The algorithm used here is a combination of both rc and tracing, it's not simple.

Thus making the point that in the end it always ends up being some form of tracing GC when performance matters.

Right. To make RC work like a tracing GC (performant and without failing on circular references); it's going to be a combination. That said, there are a lot of interesting ideas in this that might be quite worth it. Like the article referenced, doesn't do a full heap scan, but makes a note of what is worth looking at.