For root filesystems I am recently favoring EROFS which is read only, fast, and can run from ram.
On workstations I install almost all software to ~/.local as the home partition is writable. I also put all my systemd user units there, so I can run any binaries I compile as a system service to survive reboots as I like all without root.
The system root partition should contain a kernel, init system, and run any essential shared services unprivileged and fully/mostly stateless. Given that, I like to compile them all into a single UEFI uki image that contains efi shim, kernel, init all in a single binary that lives in the FAT32 UEFI partition.
The only reason to touch it is when you need to update your init system or kernel, which were always going to require a reboot anyway unless you get really creative with kexec. In an ideal situation the uki bundle is so generic that it is built deterministically in multiple locations and signed with a secure boot key. Then you can just straight up allow users to write to the boot partition, knowing any unsigned image that is not newer than the current one will be detected and also not allow access to the encrypted disk. The permission for that one update path can and should be external, and the result of a deterministic build matching a known hash, so you can prove it is not compromised. This could of course be automated by a distro with a UEFI wrapper or coreboot so users with no desire to customize their kernels do not have to think about it.
I am pursuing these ideas in stagex, first for secure enclaves and servers where we need it most, then for desktop. Until then Qubes is the least bad option.
stickynotememo: "How would you install packages (or update the kernel)?"
You: «You'd use sudo or su (of course) but I'm going to distract you from that with an advertisement for the work I'm doing on my personal project!» [0]
But, perhaps I misunderstood what you wrote. So... I'll ask stickynotememo's question in a different way:
How would a user of your system update the Systemd collection of software to apply a critical security update?
If the answer is something like "I'd rebuild the read-only root partition with the new code.", then I ask:
1) How does the data in that root partition get rebuilt? If it's on another computer, how does one control access to the root-partition-rebuild computer?
2) How does one instruct the user's computer to use that newly-rebuilt root partition? How does one control access to the system that permits one to load a new root partition and/or change which root partition to use?
3) How does a user recover when an update fails or is faulty in a way that your smoke tests didn't catch?
Please keep your answers concise.
[0] Personal projects are great, and I strongly encourage them. However, the way you've answered is what a greybeard would describe as "pulling a fast one". Pulling a fast one does not earn trust.
For root filesystems I am recently favoring EROFS which is read only, fast, and can run from ram.
On workstations I install almost all software to ~/.local as the home partition is writable. I also put all my systemd user units there, so I can run any binaries I compile as a system service to survive reboots as I like all without root.
The system root partition should contain a kernel, init system, and run any essential shared services unprivileged and fully/mostly stateless. Given that, I like to compile them all into a single UEFI uki image that contains efi shim, kernel, init all in a single binary that lives in the FAT32 UEFI partition.
The only reason to touch it is when you need to update your init system or kernel, which were always going to require a reboot anyway unless you get really creative with kexec. In an ideal situation the uki bundle is so generic that it is built deterministically in multiple locations and signed with a secure boot key. Then you can just straight up allow users to write to the boot partition, knowing any unsigned image that is not newer than the current one will be detected and also not allow access to the encrypted disk. The permission for that one update path can and should be external, and the result of a deterministic build matching a known hash, so you can prove it is not compromised. This could of course be automated by a distro with a UEFI wrapper or coreboot so users with no desire to customize their kernels do not have to think about it.
I am pursuing these ideas in stagex, first for secure enclaves and servers where we need it most, then for desktop. Until then Qubes is the least bad option.
To summarize:
stickynotememo: "How would you install packages (or update the kernel)?"
You: «You'd use sudo or su (of course) but I'm going to distract you from that with an advertisement for the work I'm doing on my personal project!» [0]
But, perhaps I misunderstood what you wrote. So... I'll ask stickynotememo's question in a different way:
If the answer is something like "I'd rebuild the read-only root partition with the new code.", then I ask:
Please keep your answers concise.
[0] Personal projects are great, and I strongly encourage them. However, the way you've answered is what a greybeard would describe as "pulling a fast one". Pulling a fast one does not earn trust.