I agree that for it to be generally useful, it’ll need to be easy to port existing software to. But I don’t think I’ll need a translation layer at the syscall level (unless I want to expose a Linux-compatible ABI as you suggest, which is totally a non-goal).
You can get surprisingly far in this area with a decent libc implementation - once you have that porting a lot of things becomes possible. There will always be harder problems (e.g. anything that hard-depends on fork) but with enough work pretty much anything in user-space should be possible to port eventually.
I’m using newlib for libc, with a custom Anos-specific libgloss that mostly talks to SYSTEM via IPC to get stuff done, and uses syscalls where needed (and a process has the appropriate capabilities). I’m filling out that low-level interface as I go, and will be using porting of exiting BSD or Linux software to drive that implementation :)
I agree that instead of implementing the Linux syscalls implementing just those libc functions that are OS dependent is sufficient for a large fraction of the existing programs.
However, this means implementing a POSIX layer, similarly to Windows NT. So it is a solution more distant from a "non-POSIX operating system".
While a large part of the Linux syscalls have been motivated by the necessities of implementing POSIX, there are also significant differences and the Linux applications that require the best performance must deviate significantly from POSIX, so they are no longer based on libc for I/O.