I'd go with that option too. I don't think users want to be surprised with being experimented on. Some users could take it worse than others.

Getting one or a few new features is one thing, getting too many might be too much.

Some granularity and agency for the user is valuable. Maybe let them pick everything as a whole or a few features at a time.

I respectfully disagree. Depends on number and severity of experiments. Comparing two decimals is really not harder than checking a boolean, still a single "if". I do not see much over-engineering here.