👷 ci: test on Python 3.10, 3.14, and 3.15

[nstarman]

The test matrix is now 3.10, 3.14, 3.15 (replacing 3.13), on Linux, macOS and Windows. This is the first time CI tests the code path that uses the PEP 814 built-in frozendict with JAX's native pytree support. #82's parity tests also run against the real built-in on 3.15.

Changes

• ci.yml
  • The matrix is now ["3.10", "3.14", "3.15"].
  • The GitHub code-coverage artifact comes from the ubuntu 3.14 job (it was 3.13).
  • The tests-latest canary, which installs the newest dependencies, now runs on 3.15. It exists to catch upstream breakage early, which matters most on the newest Python.
• uv.lock
  • jax/jaxlib → 0.11.2 for Python ≥ 3.14. The previous lock had jax 0.7.1, which has no Python 3.15 wheels, so uv sync --locked would have failed.
  • Every other package whose locked version had no cp314/cp315 wheels is upgraded too: numpy, scipy, ml-dtypes, opt-einsum, markupsafe, pyyaml, watchdog, librt, and pywin32 (311 → 312). Without this, CI would compile them from source; scipy would likely fail on Windows, and pywin32 can't be built from source at all. pywin32 arrives via ipykernel → jupyter-core in the dev group, which the nox test session installs, and it is what failed the first 3.15 Windows run. Every numeric package now has cp315 wheels for Linux x86_64, macOS arm64 and Windows amd64.
  • Two packages still build from source, because upstream hasn't published wheels: pyyaml 6.0.3 on 3.15, and watchdog 6.0.0 on macOS 3.14. Both come from zensical in the docs group, which is part of dev. Both compiled cleanly in local runs. pyyaml also falls back to pure Python if its C extension fails to build.
  • Python 3.10 resolutions are unchanged (jax 0.6.2, numpy 2.2.6).
• pyproject.toml: adds the 3.14 and 3.15 classifiers and removes a duplicated 3.10 line.
• _core.py: the comment on the sys.version_info >= (3, 15) branch said CI didn't exercise it, which is no longer true. The pragma: no cover stays because each job's coverage report only sees one side of the branch.

Verification (local, macOS arm64)

For each Python version I ran the CI steps in a fresh environment: uv sync --locked --no-default-groups --group nox --group test followed by nox -s test.

Python	jax / numpy	Result
3.10.15	0.6.2 / 2.2.6	135 passed
3.14.3	0.11.2 / 2.5.4	135 passed
3.15.0	0.11.2 / 2.5.4	135 passed, using the built-in frozendict

uv sync --locked --all-groups --no-build --dry-run --python-platform … succeeds for every OS × Python combination in the matrix, except for the pyyaml and watchdog source builds noted above. The docs workflow's locked install and zensical build also still work, and prek run --all-files passes.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@8a48b5e). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff            @@
##             main       #83   +/-   ##
========================================
  Coverage        ?   100.00%           
========================================
  Files           ?         3           
  Lines           ?        98           
  Branches        ?         0           
========================================
  Hits            ?        98           
  Misses          ?         0           
  Partials        ?         0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.