There’s no need to access the full file system to download to wherever the user wants. In fact the user might not want to use the local file system, but instead a “cloud” storage provider app!
The Storage Access Framework is built precisely for this.
The real core difference is that XMPP just passes messages around (and history is just bolted on as an extra thingy between you and your server), while Matrix is literally a federated database of message history.
You can actually literally watch marcan/lina/(more rarely)alyssa from asahi work on livestreams :) No, they don’t actually do that kind of clean-room. There’s no issue with learning about how the hardware works and directly applying that to writing drivers! Hardware and drivers aren’t even comparable, they’re complementary. One basically cannot infringe on the other. They mostly avoid touching Apple drivers, but even when they’ve had to poke at them it kinda just wasn’t a big deal IIRC.
Clean-room techniques come from actually cloning very similar things. Basically the original Phoenix BIOS cloning IBM. As a more recent example Wine/ReactOS.
Google’s extension of RCS does do e2ee, which raises the question of “what happens to security when you talk to a non-Google user”…