ENH: compute hash without PYTHONHASHSED

[redeboer]

Closes #443

[redeboer]

@Zeyna777 it would be good to extract a separate maintenance PR from this PR that only updates the lock files and implements the formatting and fixes by Ruff (29b2d40...ffe52cc).

git fetch origin
git checkout -b update-lock origin/main
git cherry-pick 70de299 9836f66 ffe52cc
git push origin update-lock -u

Could you create that PR? Then it's easier to review this PR after that one has been merged.

[redeboer]

I'm currently investigating why the hashes are different on CI
https://github.com/ComPWA/ampform/actions/runs/12396403891/job/34604373028?pr=444

I tried on two different Linux machines (Almalinux 9 and WSL Ubuntu 24.04), but there the tests work. To quickly check, run:

pytest tests/sympy/test_cache.py

Commit	Algorithm	Failing tests
29b2d40	xxh3_64	both formalisms
5b3a507	xxh64	only canonical-helicity
4d26866	xxh128	only canonical-helicity

See xxhash.com and the xxhash package for the differences between the algorithms.

[redeboer]

Okay... so the hash is different depending on whether running all tests

pytest

or the specific test separately

pytest tests/sympy/test_cache.py

[redeboer]

So, not using a fixture fixes it. I guess the moment of creating the model while running pytest somehow influences the hash. Fixed by 8e96e69.

This is not good though... it suggests that the hash depends on the runtime of the program (?). It should only depend on the binary bytes content of the object that is being cached.

It seems that ReactionInfo itself does not have a stable hash, even if it is marked as immutable (see 7bcf84b). I suspect the hash of FrozenDict (8a9a171). We should try either MappingProxyType¹ (see PEP 416) or frozendict as an alternative.

Footnotes

1. Ah okay... this type is not pickable. ↩

[redeboer]

Alright, so frozendict has a stable hash with xxhash 🎉 See a1d4c94. FrozenDict does not...