We took a similar approach to what you described. Then we called it Hack.
This is what we mean by seamless inter-operation: the HHVM runtime understands both syntaxes and runs both <?php and <?hh code in the same process. Whether Hack integrates into the runtime at the parsing (current) or bytecode (future?) layer is an implementation detail.
If you mean that future versions of Hack will evolve to have a different syntax, while still targeting the runtime... then you'll still have code around from the intermediate era, and you'll have to interoperate with that too, won't you? You'll have PHP, PHP-looking-Hack, and actually-nice-to-code-in-Hack.
That's not to say that syntax doesn't matter, but semantics and pragmatics tends to matter more, despite Wadler's Law of Language Design stating that most people don't understand this.
I also wouldn't count Clojurescript here. Its a whole different language that just happensto compile down to JS.
You don't need to rewrite anything (at least, not all at once.) Personally, I'd have expected you to make something akin to CoffeeScript or ClojureScript that targets PHP, and can "link with" your existing PHP modules (or rather, with their HHVM bytecode representations.) Then treat the PHP code as a constantly-dwindling Big Ball of Mud (http://laputan.org/mud/).