[philipdkennedy]

Description

Complex-valued, dask-backed arrays can now be decomposed. Previously, the dispatch
matrix in the Decomposer had no branch for the complex + dask combination, so
Hilbert EOF/MCA on chunked data raised — the restriction existed because the code
path was never implemented, not because the numerics are unsound.

• xeofs/linalg/decomposer.py and xeofs/linalg/_numpy/_svd.py: the dask branch
  now catches complex-valued input alongside real-valued input, for both the
  full (tsqr) and randomized (svd_compressed) solvers.
• Model-level guards (assert_not_complex in GWPCA, SparsePCA, OPA, MultiCCA)
  are intentionally retained — those algorithms do not support complex input
  and should continue to reject it loudly.
• Note: solver="full" on dask input requires tall-and-skinny chunking
  (single-column-of-blocks), as before; documented in the test fixtures.

Behavioural change

HilbertEOF.test_compute expectations inverted: fitting on dask input
previously raised NotImplementedError; it now succeeds and returns
complex-valued components and scores. This is the intended capability.

Tests

• Fourth-quadrant test parity in tests/linalg/test_decomposer.py:
  test_fit_dask_full_complex, test_fit_dask_randomized_complex,
  test_dask_array_based_on_target_variance_complex, and a dask_complex
  case in the test_random_state parametrization.
• Red-green verified: with the dispatcher files reverted to main, the
  complex-dask tests fail; with this branch, 29/29 pass in tests/linalg/
  and the full suite is green apart from two pre-existing test_mca.py
  failures that also occur on main.
• New fixtures mock_complex_data_array / mock_dask_complex_data_array
  in tests/linalg/test_decomposer.py, mirroring the real-data fixtures.

Dependency note

Complex-valued dask input requires dask >= 2026.6.0: earlier versions
contained conjugation bugs in svd_compressed/tsqr for complex data
(dask/dask#7639, dask/dask#9421), producing silently incorrect
decompositions. Verified against 2026.6.0, where the numerics are correct
(round-trip tested in tests/linalg/).

Known issue

Complex-dask fits emit a cosmetic ComplexWarning from dask's meta
construction (hardcoded float64 meta in tsqr, upstream
dask/dask#10284). It is numerically inert since dask's complex conjugation
fixes. Suppressed locally in the affected tests with a comment, rather than
project-wide, so a future genuine cast is not hidden.

Agreed in advance with @nicrie to fix #197.

[nicrie]

@philipdkennedy thanks for the PR! I'll try to review it within the next day – please bear with me a little.

[nicrie]

Sorry @philipdkennedy that I keep you waiting for so long. The CI requires some updates that I'm working on, then we can bring in your changes cleanly. I'll keep you updated.

[nicrie]

Hi again @philipdkennedy – once #257 is merged, could you update your branch by pulling the latest changes. The CI should restart automatically and many (hopefully all?) failing tests will be gone.