I'm on Ecma TC39 and in touch with JS hackers working for all the major browser vendors. FWIW, as far as I can foresee there will be no standard JS bytecode. Commenters here touch on good reasons:
* divergent, evolving VMs can't agree (not all NIH, see third bullet);
* IPR concerns;
* lowering from source over-constrains future implementation, optimization, and language evolution paths;
* view-source still matters (less, but still).
A binary encoding of JS ASTs, maybe (arithmetic coders can achieve good compression; see Ben Livshits' JSZap from MSR, and earlier work by Michael Franz and Christian Stork at UCI). But this too is just a gleam in my eye, not on TC39's radar.
Meanwhile, we are making JS a lot better as a target language, with things like ES5 strict mode, the ES-Harmony module system and lexical scope all the way up (no global object in scope; built on ES5 strict), and WebGL typed arrays (shipping in Firefox 4).
We have a Mozilla hacker, Alon Zakai, building an LLVM-based C++-to-JS compiler. Others are doing such things (along with good old SNES emulators and the like).
So being a good mid-to-high-level, memory-safe compiler target language is on TC39's radar. Not to the exclusion of other JS desiderata, and never in a way that compromises mass-market usability or buy-by-the-yard rapid-protoyping "scriptability". But one among several goals.