It's early days, but I think that given published client-side source, enough users to provide vigorous peer review, and a verification mechanism like AppHash (which guards against exactly the kind of injection you talk about), it's possible to get it right.
What ever happened to all the other non-web-app crypto that has been doing much better for the past 20 years (PGP, SSL, /dev/urandom)? How does your solution compare to those?
As an aside, I would _love_ for browsers to expose a good, standardized set of crypto routines. They don't, so we do the only thing we can - implement them in Javascript. This makes me as queasy as anyone, but it's the best we can do at the moment.
This concept of "host-proof hosting" seems to encourage that we discard well-proven methods in favor of implementing new crypto schemes in the hostile browser-space.
Are web end-users really better equipped to ensure application security than web host administrators? </rhetorical>
It would be better to have a desktop application, with source code the user can verify. Then no matter what the host changes on their end, I can know that only the encrypted text is being sent. Even better, host-based applications can tie into existing frameworks like GNOME Keyring and GnuPG, for even better security than can be provided by a 128-byte passphrase.