If you do A/B testing you should be trying to control as many variables as possible.
It makes sense for twitch, given its skewed distribution, to compare the behaviour of users engaging with the different types of streamers.
Under this logic, I presume they might actually see that engagement is not substantially harmed for big time streamers while getting actually millions of views on their ads as opposed to engagement harmed for low viewership streamers who don't really give the platform much money anyway.
I think it's a flaw in your thinking to assume twitch might have the interests of small streamers in mind. They probably don't. Once a big streamer retires, another takes their place. The viewership remains.