Okay, but as I said before, for specific cases we can still do things the old way. I.e., upon decoding we detect that the url points to a known format, and we run the fast+safe decoder. For unknown formats, we download the slow decoder and use that. This way, we have more than what we would otherwise have (flexibility). And in the future, WASM will be faster anyway. I only see benefits. Of course, the sandbox should be formally proved correct first.
The computational power you need for image decoding is extremely narrow and easy to make safe. You need some mathematical operations and some loops. You don't need any APIs or data structures. Mask off all the pointers and you can have have a provably safe interpeter/compiler that runs pretty fast.
And that's accepting that WASM is safe, which I do not axiomatically accept. History suggests that I am very safe in claiming that most implementations will end up with some catastrophic-level security vulnerabilities in them before all's said and done.