Content addressing can help here. If every asset is identified by a cryptographic hash, the user agent is free to fetch it from any available source. In the c

Content addressing can help here. If every asset is identified by a cryptographic hash, the user agent is free to fetch it from any available source. In the case of images, you can hash over the rendered pixels instead of the file encoding, so all transparent 1x1 images are concise red interchangeable. Alternatively/in addition, the user agent can treat embedded assets like textbooks do, and present them all as numbered, boxed, and captioned figures.