It "scales well" because of shouldComponentUpdate. That lets you avoid building huge chunks of virtual DOM as your app becomes very large.
The trick is that the same exact optimization is available in Elm and Angular 2 (with different names) so the same argument applies there as well.
The "Do these results generalize?" section makes an argument as to why the numbers you see on a simple TodoMVC app should generalize to apps of any size. When everyone has the equivalent of shouldComponentUpdate, the question becomes: when you finally DO need an update, how fast is it? That can be measured in small apps.