I'm worried about complexity. Traditional Linux package management isn't perfect. Atomic distros try to address this in various ways, including but not limited to a clear separation of OS and user packages.

For example, on Fedora Silverblue you're not meant to modify the base OS with rpm-ostree. You're supposed to use flatpak and toolbox (homebrew and appimages are also options, but not preinstalled). But these have limitations that rpms do not.

On the BSDs, this isn't really a problem. It manages a clear OS and user package split without the sandboxing limitations. That's not to say that sandboxing is bad, I fully support it, but when you run into the currently unfixed limitations, it's a pain. As an example, in the snap and flatpak versions of Chromium-based browsers, you can't use Android's adb tools, which I used to need when using custom OSs on my Pixel. You either have to overlay the rpm or use a non-atomic distro.

Systemd sysext is being thrown around as the solution to this problem as the way to install software that needs to be unsandboxed but without modifying the base OS. But I don't see why instead we can't just still be able to install RPMs on the base system, but to places like /usr/local/bin. The systemd sysext method seems like an unnecessary reliance on systemd and additional complexity. Take this for example: https://github.com/mmcnutt/Bazzite-Discover-Sys-Ext. There's so much work being done, what what is basically just taking the existing Plasma Discover rpm and converting it into a new format.