ECMA doesn't require efficient tail recursion. Caveat programmer.
If you're going to do real CPS programming in JavaScript, you'll want to use a trampoline or a timer mechanism to reset the stack.
But, in most cases, you do a little computation and then block waiting for an event.
When the continuation gets invoked by the event-handling loop, the stack is reset.
You're not compiling to C here: "the stack" may be a chain of heap-allocated activation records that are still live as far as GC is concerned. What you're describing might work but you'd want to test it to be sure.
EDIT: On reflection I agree that the setTimeout() approach should free the stack, however it's implemented.
You can always use setTimeout to call your tail, which has the neat effect that your CPS "threads" become cooperatively multitasked and can progress alongside each other. Just replace "foo()" tail calls with "setTimeout(foo, 0)".
And the downside that you're nuking your stack every time, making debugging quite a bit harder.
You can work around it with a trampoline (http://en.wikipedia.org/wiki/Tail_call#Through_trampolining), which is what I did with cfa-js. (http://dl.dropbox.com/u/6600185/cfa/cfademo.html)