My specific use-case might be niche (I sometimes have to work on my phone's connection for a week or two), but I think everyone would benefit from being able to download the binary diff since the previous package version instead of downloading the entire package every time.

This would obviously increase the mirror's storage usage, but it would also greatly reduce their bandwidth usage, which I believe to be where most of the cost lies for a mirror provider, so I expect they would approve. This would need to be confirmed with them though, obviously.

Implementation-wise, I'm thinking that whenever a new package is generated, `makepkg` would uncompress the old & new packages and diff them (using `bsdiff` or equivalent), and that diff would be stored next to the new package.
This way, someone who doesn't have the previous package in their cache would download the full package, as before, nothing changes.
But if `pacman` sees a previous version in its cache (such as when performing an `-Syu` update), it would query the mirror for the diff between that version and the latest, and fall back to downloading the full package if that diff doesn't exist.
If it does exist, the diff is applied onto the package in the cache (or directly on the local install?)

To avoid having too big diffs (resulting in too much storage cost/not enough bandwidth saving), a threshold can be added, where the diff isn't uploaded if it's not <50% of the full package size, for instance for packages that contain mostly binary files that are recompiled into something completely different each time. The whole "reproducible builds" effort should also help with this, to avoid having compilation outputs that change from one version to the next when the source file hasn't changed.

In summary:
- The downside is (on top the of implementation effort) slightly[*] increased storage requirement and a slightly[*] longer package build & upload for package managers.
- The upside is faster download/updates for users, and less bandwidth consumed for users & mirror providers.

[*] Note that I /assume/ that the build time & storage space cost are small, but I haven't built a prototype of this and ran it against the existing repos' packages, which is the only way to actually know.
I'm willing to make these implementations, but I want first to make sure this is an idea that has a chance of being accepted, and also I'll need a contact person (or some access to gitlab.archlinux.org) to discuss the makepkg & pacman implementations :)