The advantage of the scheme that zcorpan proposed is that there is no magic
proxy; we just add a capability to <img> to select its source using more
than just a src attribute. This has better fallback than your design and is
easier to implement.
I believe that from testing perspective both approaches are equivalent.
The spec I propose *is* only another way to control src of an image.
The only difference is that I don't expose the <img> to scripts.
That may make it even simpler, because you can't have odd cases like author
moving/removing the controlling img or setting values directly on img that
conflict with picture's definitions.
--
regards, Kornel
