Also, total time to repaint seems to be equally fast whether we're recreating the entire 2K rows from scratch or changing the display property of 4K rows. That seems concerning to me.
But yes, a more fair comparsion would be to write a WebGL or Canvas based layout engine in JS. If its less than 4 times slower (its JS after all), then the DOM is bloated.
I certainly would expect that you could beat the DOM by specializing – e.g. a high-speed table renderer could make aggressive optimizations about how cells are rendered, like table-layout:fixed but more so – but the more general-purpose it becomes the more likely you'd hit similar challenges with flexibility under-cutting performance.
The most interesting direction to me are attempts to radically rethink the DOM implementation – the performance characteristics of something like https://github.com/glennw/webrender should be different in very interesting ways.
Similarly, React is significantly slower – usually integer multiples – than using the DOM. The only times where it's faster are cases where the non-React code is inefficiently updating a huge structure using something like innerHTML (which is much slower than using the DOM) and React's diff algorithm is instead updating elements directly. If you're using just the DOM (get*, appending/updating text nodes instead of using innerHTML, etc.) there's no way for React to be faster because it has to do that work as well and DOM + scripting overhead is always going to be slower than DOM (Amdahl's law).
The reason to use React is because in many cases the overhead is too low to matter and it helps you write that code much faster, avoid things like write/read layout thrashing, and do so in a way which is easier to maintain.