This is an interesting point. Imagine if you put a fake SSH agent on 22, it responds just like SSH but never allows a login. Would it make it even less likely that someone would bother trying another port?
Depends what we mean by sandbox. I wouldn't make a chroot the honeypot, but I don't see an issue with a program that just simulates a shell but doesn't allow exec or real fs access, for instance.
I don't think you would ever let them touch OS-level resources. There are plenty of third-party ssh server libraries where you just get a Reader and a Writer to the remote end. When they connect, you write "root@cool-computer# ". When they send bytes, you discard them, then print "root@cool-computer# " again.
While obviously accepting a TCP connection and allocating resources on your computer is more risky than just ignoring the connection, presumably it would be fun to do this, which is a good reason for doing something. You can set a memory limit, file descriptor limit, etc. and just crash if they're exceeded. You can run your little fake ssh daemon in gvisor and protect against attacks nobody even knows about yet. All in all, it would be pretty low risk, and also pretty interesting.
Moving SSH to a non-standard port is helpful only to reduce log noise coming from untargeted attacks. If an attacker is looking at your system, your attack surface is the same no matter where your SSH daemon is bound. I don't think it's worthy any extra effort to "distract" attackers like this.