[nicolamasotti]

What this PR fixes

After pip install xarray-ecmwf, opening any dataset fails:

ds = xr.open_dataset(request, engine="ecmwf")
# ImportError: chunk manager 'dask' is not available. Please make sure 'dask' is installed and importable.

and once dask is installed:

# ImportError: The zarr package is required for working with Zarr stores but could not be imported. ...

The engine needs both, for every client (cdsapi, ecmwf-opendata, polytope):

• dask: it opens each downloaded file with chunks={}, and chunks the schema sample of each variable;
• zarr: it caches the schema sample of each variable with to_zarr, and reads it back.

Neither was ever declared, not even in the old conda environment.yml. They were only added to the dev group in d962e26, to make make all-tests work.

With zarr 3, xarray must also be at least 2025.6. Older versions fail on every open:

# xarray 2025.4.0 or older, with zarr 3.1.6
# TypeError: Expected an integer or an iterable of integers. Got None instead.

Changes

• dask>=2026.8.0 and zarr>=3.1.6 move from the dev group to the package dependencies, with the same minimums.
• xarray>=2025.6, the first version that works with zarr 3.
• deptry ignores dask and zarr as unused: the package doesn't import them, xarray does.

Tests

All CI jobs pass locally. A plain install of the built wheel now includes dask and zarr, and opens a dataset.

Notes

• dask>=2026.8.0 is the minimum the dev group already tested. It could be lowered, if older versions are tested too.
• Once Fix reads of the first chunk when CDS returns it short #15 is merged, the skip for xarray < 2025.6 in tests/test_20_client_cdsapi.py can be removed.