Adding one that fits your filter and is not on the list.
On-device models plus the share sheet also make possible a thing I have wanted for years: a reply-drafting layer that never sees the network and never sees the thread. Not a writing assistant, a tone checker for messages you are about to send while annoyed. It reads what you typed, on the device, and says one thing: this reads angrier than you probably mean.
It could not have been built in 2024 for the reason your framing is built on. Sending someone's half-written angry message to a server to be assessed is a non-starter, and everyone knew it, which is why nobody shipped it. Locally it is fine.
Whether anyone would install it is a separate question and I suspect the honest answer is that the people who need it most are exactly the people who will not.