add python==3.13 and drop python==3.8

[Kallinteris-Andreas]

No description provided.

[RedTachyon]

Is there anything forcing us to break compatibility with 3.8? My general intuition is that we can keep it around for now, and once we need any features that were added in 3.9, we drop it without much worry in the next release

[Kallinteris-Andreas]

Generally we drop support for python versions once they become unsupported by the python developers.

[RedTachyon]

This isn't some dogma that we must always follow. My point is a bit more from first principles - there's a difference between actively and passively supporting some version.
With 1.0, we will actively support python>=3.9,<=3.13, as in we make sure that Gymnasium works with each of those versions.
On the other hand, I want to avoid deliberately "unsupporting" 3.8, i.e. we passively support 3.8 for as long as it makes sense.

FWIW I just checked and I'm pretty sure this PR actually does it the way I'm proposing:

Gymnasium/pyproject.toml

Line 11 in 50bf685

requires-python = ">= 3.8"

If we really want to drop 3.8, we'd bump this to 3.9, so that users cannot (without some weird workarounds) install gymnasium on 3.8. I don't think that makes sense to do if we know that Gymnasium actually runs alright on 3.8.

On the other hand, if we want to keep this "passive" support, removing it from the test suite makes it... awkward. Let's say in 1.1.0 we want to add a feature which uses, say, dict merging {1: 2} | {3: 4}, which was added in 3.9. We shouldn't hesitate to add this - but at that point we should also stop supporting 3.8, even passively, i.e. we actively mark it as incompatible by changing the aforementioned line in pyproject.toml. But if we don't have 3.8 in the test suite, then we won't know when this happens, so we might end up in a weird state where someone (for whatever reason) using 3.8, by default installs a version of Gymnasium that doesn't actually work.

All this to say - we should add 3.13 to the test matrix, and keep a policy that 3.8 compatibility can be broken, but consciously. There is no reason that I see to actually drop support right now, so let's keep it to maintain wider compatibility. We definitely shouldn't do anything halfway, i.e. allow it to be installed but untested.

[Kallinteris-Andreas]

this is done because 1.0 release

[Kallinteris-Andreas]

Most libraries such as numpy also stop supporting python versions as they become unsupported, so some issues may be caused long term, I am not opposed to making an exception for 1.0 specifically to have 3.8 supported, also depends on when 1.0 is actually being released

[pseudo-rnd-thoughts]

This would break what @RedTachyon is suggesting, but if we completely dropping 3.8 then we should update pyupgrade in pre-commit to --py39-plus which causes a significant number of changes, mainly reducing the number of type imports and to pyright in pyproject.toml with pythonVersion="3.9"

[pseudo-rnd-thoughts]

Python 3.13 has been delayed by a week due to unexpected performance regressions fyi

[Kallinteris-Andreas]

Waiting for pytorch as usual