[jessegrabowski]

The actions/cache key for ~/.pytensor in run_tests.yml now ends in the commit SHA, and the old lockfile-hash key is the first restore key. A saved cache entry is never overwritten, so the old key kept whatever the first run on each pixi.lock compiled. Every graph changed after that run compiled cold on every later run.

Main at 142ad12 restored the optim cache on an exact key match and still took 12.4m, because #179 changed the L-BFGS graphs after that entry was saved. Locally the same group takes 444s with an empty cache and 84s with a warm one.

On Windows the cache step now caches ~/AppData/Local/PyTensor, which is where pytensor compiles there. ~/.pytensor doesn't exist on Windows, so the windows smoke job had never saved a cache.

The pytensor pin in pyproject.toml and conda_envs/ now requires >=3.3.8.

Tests in tests/optim/ that another test already covers are removed, and several tests now share compiled functions. test_rule_reduces_loss runs one case per rule. The L-BFGS kept-trial comparisons in test_rules.py now compile one step and run it twice from a restored snapshot.

test_a_single_precision_search_stays_in_single_precision now pins a float64 run that brackets over six trials. Its old case accepted the first trial, so the float32 zoom never ran.

Closes #109

---

📚 Documentation preview 📚: https://pytensor-ml--180.org.readthedocs.build/en/180/