Yeah, diagnostics would be really beneficial... possibly something that can generate a bundle from the polyfills that get added. Of course, would need to run across several browsers and aggregate the results.
Only including what you actually need, vs say the babel-env fills, that include everything missing for your target, even if you aren't using it.
Blocking requests can be seriously nasty in practice though... I had to use them for SCORM implementations at one point. And if you have a spotty connection as a user (think mobile) it can/will get particularly bad.