Hello, one of the paper authors here.
One of the jobs that our implementation has to do is close down all channels by which JavaScript could leak data: so this includes postMessage, the DOM, XHR, cookies, local storage... fortunately, these overt channels are well specified DOM APIs so from the implementor's perspective it's not too difficult to lock them all down. (And we just deny access, there's no "shadow" structure s involved.) There's nothing too fine grained going on: if you read private data, everything it isn't allowed to flow to is locked down.
Their example of a password strength checker -- say you download updated bitmaps (bloom filters) from a central location which alert for risky passwords being used. Now you can trust any code path which actually sees the password will be locked down through taint detection.
I wonder if there's a basic mechanism, or if they have to identify every possibly avenue for tainting? There are a lot of static locations you could stash bits, so are you just denied access to all of that? Or somehow is it all mirrored?
I want to make sure I really understand what they are saying can be trusted, what privacy guarantees are provided? What exactly is sandboxed?