[antonsoo]

MultivariateEof documents support for masked arrays, but np.concatenate discards their masks when merging fields. Changing values underneath a mask then changes the eigenvalues and includes excluded cells in the returned EOFs:

import numpy as np
from eofs.multivariate.standard import MultivariateEof

x = np.ma.array(np.arange(36, dtype=float).reshape(12, 3), mask=False)
x.mask[:, 1] = True
y = np.sin(np.arange(48)).reshape(12, 4)
before = MultivariateEof([x, y]).eigenvalues()
x.data[:, 1] = 1000 * np.arange(12)**2
after = MultivariateEof([x, y]).eigenvalues()
np.testing.assert_allclose(before, after)

Use np.ma.concatenate when a field is masked so the existing solver excludes those cells. Promote masked integral merged arrays to float for the solver's NaN representation. This also preserves masks in projectField; ordinary array concatenation stays unchanged.

Seven regressions cover mixed/both masked fields, weighting, projections and integral inputs. The full suite passes on Python 3.12 with Iris/xarray/Dask and the core suite passes on 3.8. Independent valid-cell SciPy SVD checks on the bundled SST data cover eigenvalues, projections, reconstructions and masks through the standard and Iris interfaces. Previously correct complete outputs remain exact, including additional float32/float64 and Dask controls. A clean installed wheel repeats the results; composition with #166 preserves its covariance workloads.