Tracking issue for temporary test skips/xfails to revert

[ev-br]

(As suggested in #17273 (comment))

Once in a while, we skip or xfail tests to quickly restore the CI to the green status, when we either lack the immediate bandwidth to resolve an issue properly, or a fix takes a while etc. Once a proper fix lands, the skip needs to be reverted.
This is a tracking issue to keep track of these skips and their reverts.

• TST: stats: silence warnings temporarily #17272 silences warnings produced by stats distributions with implementations from Boost. It needs to be reverted once BUG: stats: distribution methods emit unnecessary warnings from Boost #14901 is resolved.
• TST: linalg: temporarily silence failure in test_solve_discrete_are #17273 skips a test of a discrete Riccatti equation which fails on Cirrus CI. TST: linalg: temporarily silence failure in test_solve_generalized_discrete_are #17525 skips a similar test which fails on some machines. The skips need a revert once TEST/BUG: Tolerance violation in test_solvers::test_solve_discrete_are #16926 is resolved.
• CI: On Azure, pin pytest-xdist to version 2.5.0 #17331 Unpin a CI package dependency of pytest-xdist because of an incompatibility with pytest-cov
• BUG: failing sqrtm regression test #20219
• BUG: scipy.linalg.blas.dnrm2 may return error result when incx < 0 #16930 adds an xfailed test of linalg.blas.nrm2(..., including=-1). Revisit when minimum openblas is >= 0.3.27

[andyfaff]

Thanks @ev-br should we pin it?

[andyfaff]

Unpin the pin in #17331

[lucascolley]

Hi all, here are the temporary skips from gh-19005 and why they are needed:

• scipy/fft/tests/test_basic.py::TestFFT1D has 7 skips for torch backend. These can be removed once array-api-compat has implemented torch.fft. There is no tracker for that yet, but it will probably be the next step after ENH: Implement the fft extension in numpy.array_api numpy/numpy#23892 is closed. There are 2 more torch skips towards the end of the file which will be able to be removed for the same reason. (The specific problem is that torch does not accept the axes keyword.)
• scipy/fft/tests/test_helper.py::TestFFTShift has 4 skips for the same reason.
• scipy/fft/tests/test_helper.py has 2 skips for numpy.array_api and cupy. These can be removed after ENH: Implement the fft extension in numpy.array_api numpy/numpy#23892 and array-api-compat implementing cupy.fft respectively. The problem here is that these libraries don't accept the device keyword of fftfreq and rfftfreq. numpy.array_api PR: ENH: Add fft optional extension submodule to numpy.array_api numpy/numpy#25317, numpy issue: Tracking issue: Add array API standard support to the main namespace numpy/numpy#25076

There are also lots of skips for GPU arrays only, but I don't think that this issue is the place to document them. They will be removed eventually once we implement GPU support where compiled code would otherwise be hit.

[lucascolley]

#17331 Unpin a CI package dependency of pytest-xdist because of an incompatibility with pytest-cov

I think this point can be ticked off now, right? pytest-xdist is not pinned in the prerelease_deps job anymore.

[ev-br]

Ticked off Riccati equation solvers in the OP, based on #20335 having removed the xfails. We might want to revisit the tolerance bumps at some point after openblas 3.26 becoming the lowest supporter version.