i mean python is 99% backwards compatible so as long as you tell your tooling you’re working with 3.10 it will warn you about using stuff that’s too new. that’s why the shipping version is usually enough. in general it’s not recommended to have multiple versions of python3 installed at the same time, but if you are a habitual venv user it’s usually not a problem. however i have also run into the issue of some versions being “too new” for a project, where the thing just would not work with newer versions.
basically, if your issue is only that you don’t want to “contaminate” an older codebase, that can be solved by configuring your tooling. but if your issue is that the thing just doesn’t work with the “wrong” version, you’re probably best of using a container. a user installation of the version you want will work but having multiple installations is annoying.
theoretically it can absolutely figure out how to do that without it being in the training data.
we know it’s in the training data because of google’s filters, but theoretically it could have been generated without having anything to draw on just due to how the thing works.