At the surface level, I could see how your position would appear logical.
However, the nature of the replacement VM doesn't really matter in the context of what we're describing here.
The problems of compatibility remain the same regardless of whether the target VM is a JS-specific VM, or a general purpose VM; reliance on undefined behavior, accidental changes in semantics, optimization bugs, and so on.
If browser makers were able to swap-out the VM, they've already demonstrated the ability to solve the exact same kind of compatibility problems they'd face here.
A VM that provides a standardized bytecode and execution model capable of executing most modern dynamic and static turing complete languages with sufficient performance. This requires more general facilities for common functionality required; refer to tail call support, JVM's invokedynamic, etc.
Or, shorthand, a general purpose browser VM.
Unless, of course, you're arguing that the above is not truly possible, in which case I'd argue that's exactly what Google is exploring via PNaCL, and has (if not completely efficiently) been implemented across existing VMs. The PNaCL choice of LLVM bitcode was likely in error, but regardless, "no such thing" is a stretch.
Which comes full circle to my original point, which is that other than Google, nobody with a vested interest is trying, and worse yet, they actively oppose progress on this front.
However, the nature of the replacement VM doesn't really matter in the context of what we're describing here.
The problems of compatibility remain the same regardless of whether the target VM is a JS-specific VM, or a general purpose VM; reliance on undefined behavior, accidental changes in semantics, optimization bugs, and so on.
If browser makers were able to swap-out the VM, they've already demonstrated the ability to solve the exact same kind of compatibility problems they'd face here.