[pattonw]

uv run pytest now measures coverage

[tool.pytest.ini_options] addopts carries --cov --cov-report=term-missing, so the README's uv run pytest prints the coverage table and enforces the committed fail_under floor exactly as CI does. The CI step drops the flags it no longer needs. A single file can be run without the floor:

uv run pytest --no-cov tests/<file>.py

testpaths = ["tests"] replaces the tests argument on the command line.

Type stub and py.typed

ty was green only because two # ty: ignore[unresolved-import] comments hid the compiled module, so every call into Rust was Unknown. mwatershed/mwatershed.pyi now declares agglom_rs, cluster, cluster_weighted and cluster_arrays with the signatures from src/lib.rs (usize as np.uintp), py.typed marks the package as typed, and [tool.maturin] include ships both in the wheel. The two ignores are gone, so ty checks the wrapper's calls into the extension and the tests' calls to it. The lint job and README run ty check --error-on-warning.

Verified locally: uv run ty check --error-on-warning clean, unzip -l of the built wheel lists mwatershed/mwatershed.pyi and mwatershed/py.typed, 76 tests pass at 100% coverage.

Import time and test duration, tracked per commit

The tests job times its own uv run pytest and measures the cold import of mwatershed (benchmarks/run.py), and uploads the two numbers as an artifact from the floor-interpreter, highest-resolution cell. A new benchmark workflow runs when a tests run on main completes, downloads that artifact, creates the gh-pages branch if it is missing, and publishes with github-action-benchmark to gh-pages:dev/bench, with a comparison table in the job summary and a commit comment when a metric gets 1.5x worse. The docs gain a Benchmarks page that charts every metric from that data; the docs workflow now runs after each benchmark run (instead of on push) and copies dev/bench from gh-pages into the deployed site, so the charts carry the newest point and the action's own dashboard stays reachable under the docs site. Nothing runs the test suite twice.

The same setup runs end to end in pattonw/prototype-lib.

🤖 Generated with Claude Code