> is there any numbers about the actual impact on the user's perception
ThreadItJS[1] has some tests for this (scroll down to the "Dependency load and execute" section). On my desktop, loading React on a clean cache takes a little over 300ms. Ember is over 600ms. This is just to get to a point where you can start a hello world with the framework on a powerful desktop machine with fast internet.
Another anecdote: one project at my day job is using the Auth0 lock library (long story, don't ask). It's basically a login widget, but it bundles all of React and a few other libraries, and it's over 1MB of Javascript (uncompressed). Is the site still usable? Sure, I guess. Could it be faster? Absolutely.
Basically it comes down to a trade-off between developer comfort and a commitment to give users the best experience they can get. You just gotta draw a line somewhere. The thing is that many people justify the trade-off of React/Ember/Angular/whatever's code size by saying it provides benefits in maintainability, etc, but the reality is that for a lot of projects, you could compromise less on the user experience side and still have good maintainability with smaller libraries.
[1](https://koglerjs.com/verbiage/performance)