LLVM IR is not platform independent, it's really just a bunch of (closely-related, ofc.) platform-specific IR's. What's wrong with WASM? Everything else seems broadly OK, of course. (Also, WASM directly supports "applets", via Canvas.)
What I'm driving at is - if we can dream, why not aim for maximal hardware performance and minimal overhead? Even if it comes down to just pre-compiling native code for common architectures like Android apps do. Right now, the browser is an idiosyncratic HAL burdened by its past... what would a really nice next generation be?
Interesting thing to know about the LLVM IR - I'm only casually familiar.
That said, I agree that performance may be a dealbreaker on that idea. HTML/DOM is also tightly integrated into a lot of browser behaviors, so I have no idea whether it's feasible to extract it into a swappable module. EDIT: It's still an idea that gets me pretty excited to try.
Once we are done with WASM we will look at it, compare it with Java Applets and say: Look, we already had that!
Wake me when it can free memory.