Does it actually work?

That situation happens for me too, even with no swap, only zram. Presumably it would happen even without any kind of swap too, because the issue is probably due to constantly paging in binaries from disk as they run?

Probably there isn't any easy way to prevent this, but it could be mitigated if systemd could lock into RAM itself and everything needed for a rescue login console, the files needed for a VT switch, login, a root shell session and command-line and interactive tools needed to find and kill processes, so that when it happens I can instantly kill offending processes using all my RAM. Debian has a memlockd package for this, but it doesn't seem to help any more unfortunately. Perhaps the systemd folks can come up with a better design and or maintain it better.

Another thing that could help is being able to automatically shut down more systemd user/system services when memory pressure begins to build, IIRC this is implemented, but probably not widespread within service configs.