ci: add dependency caching and split lint from test matrix

[framsouza]

Dependency caching: setup-uv has built-in cache support keyed on uv.lock but it wasn't enabled in the test or docs jobs. Added enable-cache: true and cache-dependency-glob: "uv.lock" to all jobs. The docs workflow already had its own caching setup so this makes things consistent.

Lint runs once: ruff and mypy were running inside the 4-version test matrix, which means they were running 4 times and producing the same result every time. Moved them into a separate lint job on Python 3.12. The test job now has needs: lint so if linting fails nothing else starts.

Coverage artifact: pytest was generating coverage.xml on every run but the file was never saved anywhere. Added an artifact upload so it's accessible from the Actions tab per Python version.

Each PR commit currently triggers 4 redundant package downloads and runs linting 4 times for no gain. With caching and the lint split, subsequent runs skip the downloads entirely and linting consumes one job instead of four. This translates directly to faster feedback and fewer GitHub Actions compute minutes burned, which at scale has a real cost impact.

Testing

Tested locally with act pull_request. All jobs pass.

| ================================ tests coverage ================================
| _______________ coverage: platform linux, python 3.12.3-final-0 ________________
| 
| Coverage XML written to file coverage.xml
| ================ 506 passed, 16 skipped, 38 warnings in 57.61s =================

NOTE: test_from_url will show as failing on this PR in CI until #188 is merged into main first.

[juliendenize]

Added enable-cache: true and cache-dependency-glob: "uv.lock" to all jobs.

Default settings should be correct here and take care of this. Might be worth to update the action but not the rest.

ruff and mypy were running inside the 4-version test matrix, which means they were running 4 times and producing the same result every time.

Don't think it's true, there are issues rising only for some Py versions so we need to ensure everything works as expected for all.

[framsouza]

thanks @juliendenize , you've raised two valid points that I had overlooked.

You're right that setup-uv handles caching well by default. I've removed the explicit cache-dependency-glob since the default already covers uv.lock, pyproject.toml, and other dependency files, which is actually broader and better than what I had. I kept enable-cache: true to be explicit, though on github-hosted runners auto would also work.

About linting, I moved mypy back to the test matrix so it runs across all four python versions, the runtime python does matters here and different versions can surface different type errors as you said.

Ruff stays in the single-version lint job because AFAIU (please correct me if i'm wrong) it works differently, it determines its target python version from requires-python in pyproject.toml (https://docs.astral.sh/ruff/configuration/#python-version), not from the runtime interpreter. It's a Rust binary that parses source files directly and never loads version-specific stubs or interacts with the python runtime. The output is identical regardless of which python runs it.

oh, and I also moved uv lock --check to the lint job since it just verifies that uv.lock is in sync with pyproject.toml, it doesn't depend on the python version, so running it 4 times in the matrix was redundant.

[abdelhadi703]

Great improvement! Separating lint from the test matrix is a solid optimization — running ruff 4 times across Python versions when the result is identical wastes CI minutes.

The coverage artifact upload is also valuable — it makes coverage data available for future integration with tools like Codecov.

I reviewed the changes and they look clean. +1 for merging.

[juliendenize]

There are still adjustments to make as now we support 3.14 and i don't think we should split linting from tests as we want the linter to work on all versions.