In general, I agree with you. But in this specific case, we may not be able to extrapolate Linux/LLVM's success to web. Yes, you can create a toy kernel for yourself. You don't have to be POSIX-compliant from day one, or run OS/2 apps, or whatever. And you can create a more standard compiler that's not capable of compiling most programs that use gcc's esoteric features. They can have little niches for years, and slowly gain traction (from end-users and also developers).

But I don't think that you can do the same in browser space. If you want to create a new rendering engine, it absolutely, positively has to render 95+% of most-visited websites from early stages of development (before you "ship" a browser). Nobody would use a half-baked browser that's unable to render most websites. So, you have to also support WebKit's bugs-turned-into-standards.

In another words, you don't compile 500 different programs in a single day - if LLVM can compile the one program that your company is developing faster and better, it's a good fit for you. But you visit hundreds of websites a day. If a new engine can't display even 10% of them correctly, it'd a show stopper.

So, your choices are to either fork WebKit, or create a new engine that "simulates" most mainstream WebKit engines. Both result in WebKit becoming more and more of an standard.