Straight-up eval() would be the most blatant case of a remote code execution vulnerability the web has ever seen. I don't think it's easy or even possible to build a foolproof sandbox out of eval(). And even then you open yourself to a denial of service since you cannot interrupt the eval().
I think the link to Geordi is not quite relevant here (although it is super cool): the claim was that it's hard to build a foolproof sandbox out of eval(). If you're using Geordi, since it's a separate process in a chroot, you can't use eval(). If you're giving up on eval(), you can write a foolproof sandboxed JS evaluator in JS alone.
Besides, the hard part with all sandboxes is not making the sandboxing itself solid; that's straightforward. The hard part is giving it a sufficiently rich interface to the outside world to do useful things, while not giving it so rich an interface that it can escape through that hole. All the semi-recent vulnerabilities in qemu, for instance, have been in virtual devices, not in actually emulating a processor.
Geordi's story is relatively easy, since all it does is compile code and execute it in an intentionally-minimal interface. JS on a web page has thousands of things it can do to the web page or to the outside world, if you want to implement it according to spec.