free-threading: one test failure

[neutrinoceros]

I'm starting to study this lib's needs to support free-threading Python.
First things first; I found a failure in the existing test suite in free-threading python 3.13:

______________________________________ [doctest] README.rst _______________________________________
091   86400.0
092 
093 It is also possible to use the ufuncs directly, though then one has to
094 deal with the warning and error states explicitly.  For instance, compare::
095 
096   >>> erfa.jd2cal(-600000., [0, 1, 2, 3])
097   Traceback (most recent call last):
098   ...
099   ErfaError: ERFA function "jd2cal" yielded 4 of "unacceptable date (Note 1)"
100   >>> erfa.ufunc.jd2cal(-600000., [0, 1, 2, 3])
Expected:
    (array([-1, -1, -1, -1], dtype=int32),
     ...,
     array([-1, -1, -1, -1], dtype=int32))
Got:
    (array([0, 0, 0, 0], dtype=int32), array([0, 0, 0, 0], dtype=int32), array([0, 0, 0, 0], dtype=int32), array([0., 0., 0., 0.]), array([-1, -1, -1, -1], dtype=int32))

/Users/clm/dev/astropy-project/liberfa/pyerfa/README.rst:100: DocTestFailure
======================================== warnings summary =========================================
README.rst::README.rst
  /Users/clm/dev/astropy-project/liberfa/pyerfa/erfa/_dev/scm_version.py:21: UserWarning: liberfa/erfa not at a tagged release, but at <ScmVersion 2.0.1 dist=0 node=g9915ba3 dirty=True branch=HEAD>
    warn(f'liberfa/erfa not at a tagged release, but at {erfa_version}')

README.rst::README.rst
  <frozen importlib._bootstrap>:488: RuntimeWarning: The global interpreter lock (GIL) has been enabled to load module 'erfa.ufunc', which has not declared that it can run safely without the GIL. To override this behavior and keep the GIL disabled (at your own risk), run with PYTHON_GIL=0 or -Xgil=0.

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
===================================== short test summary info =====================================
FAILED README.rst::README.rst
================================== 1 failed, 2 warnings in 0.17s ==================================

So, there are two independent problems that surface here:

• the GIL is actually re-enabled at runtime (this can be worked-around with PYTHON_GIL=0, and there's no need to declare support until it's properly tested)
• this test actually fails. I haven't dug through it yet (but plan to), so it seems plausible at this stage that the bug isn't necessarily in pyerfa but could be in Cython or numpy.

[ColmTalbot]

I bumped into this when trying to make another project work with freethreaded Python and I don't think the test failure is a GIL issue.
It looks like a numpy 1 vs 2 issue.
The doctest always fails for me in numpy v2 and passes with numpy v1.
The doctest also doesn't seem to run in the CI tests, presumably some issue with the tox setup.

[neutrinoceros]

The doctest always fails for me in numpy v2 and passes with numpy v1.

I think this is a red herring, or a separate issue: I can still get the test to pass on numpy 2 with

uv run -p 3.14 --extra test pytest --doctest-rst

but

uv run -p 3.14t --extra test pytest --doctest-rst

fails as described above

[ColmTalbot]

I really disagree, I was able to reproduce what you got in a fresh codespace. Maybe this failure is hardware specific?

If I move this test further up the README (specially below the first import erfa) the test fails.

$ uv run -p 3.14 --extra test pytest --doctest-rst           
Using CPython 3.14.3
Creating virtual environment at: .venv                                                               
      Built pyerfa @ file:///workspaces/pyerfa
░░░░░░░░░░░░░░░░░░░░ [0/8] Installing wheels...                                                      warning: Failed to hardlink files; falling back to full copy. This may lead to degraded performance.
         If the cache and target directories are on different filesystems, hardlinking may not be supported.
         If this is intentional, set `export UV_LINK_MODE=copy` or use `--link-mode=copy` to suppress this warning.
Installed 8 packages in 665ms
======================================== test session starts ========================================
platform linux -- Python 3.14.3, pytest-9.0.2, pluggy-1.6.0                                          
rootdir: /workspaces/pyerfa
configfile: pyproject.toml
testpaths: erfa, docs, README.rst
plugins: doctestplus-1.7.1
collected 282 items                                                                                 

erfa/tests/test_erfa.py ...............................s.                                     [ 11%]
erfa/tests/test_ufunc.py .................................................................... [ 35%]
............................................................................................. [ 68%]
.......................................................................................       [ 99%]
README.rst F                                                                                  [100%]

============================================= FAILURES ==============================================
_______________________________________ [doctest] README.rst ________________________________________
066 -----
067 
068 The package can be imported as ``erfa`` which has all ERFA_ ufuncs wrapped with
069 python code that tallies errors and warnings.  Also exposed are the constants
070 defined by ERFA_ in `erfam.h
071 <https://github.com/liberfa/erfa/blob/master/src/erfam.h>`_, as well
072 as `numpy.dtype` corresponding to structures used by ERFA_.  Examples::
073 
074   >>> import erfa
075   >>> erfa.ufunc.jd2cal(-600000., [0, 1, 2, 3])
Expected:
    (array([-1, -1, -1, -1], dtype=int32),
     ...,
     array([-1, -1, -1, -1], dtype=int32))
Got:
    (array([         0, 1077149696,          0, 1078525952], dtype=int32), array([1996,    2,   10, 9999], dtype=int32), array([         0, 1072693248, -406640992, 1072495147], dtype=int32), array([0.0e+000, 4.9e-324, 9.9e-324, 1.5e-323]), array([-1, -1, -1, -1], dtype=int32))

/workspaces/pyerfa/README.rst:75: DocTestFailure
====================================== short test summary info ======================================
FAILED README.rst::README.rst
============================= 1 failed, 280 passed, 1 skipped in 2.51s ==============================

I also get the error if I remove the line above (the one that errors).
Maybe there is some state being stored somewhere from that evaluation and the freethreaded version is just more strict about this?

This looks consistent with running the relevant commands in REPL.

$ python
Python 3.14.3 (main, Mar 20 2026, 00:34:32) [Clang 22.1.1 ] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import erfa
>>> erfa.ufunc.jd2cal(-600000., [0, 1, 2, 3])
(array([681932393,         0,         0,         0], dtype=int32), array([686375409,         0,         0,         0], dtype=int32), array([686485536,         0,         0,         0], dtype=int32), array([0.0e+000, 4.9e-324, 9.9e-324, 1.5e-323]), array([-1, -1, -1, -1], dtype=int32))
>>> erfa.ufunc.jd2cal(-600000., [0, 1, 2, 3])
(array([686320110,         0,         0,         0], dtype=int32), array([167598,      0,      0,      0], dtype=int32), array([-1600300256,       30382, -1600300256,       30382], dtype=int32), array([0.0e+000, 4.9e-324, 9.9e-324, 1.5e-323]), array([-1, -1, -1, -1], dtype=int32))
>>> erfa.jd2cal(-600000., [0, 1, 2, 3])
Traceback (most recent call last):
  File "<python-input-3>", line 1, in <module>
    erfa.jd2cal(-600000., [0, 1, 2, 3])
    ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^
  File "/workspaces/pyerfa/erfa/core.py", line 612, in jd2cal
    check_errwarn(c_retval, 'jd2cal')
    ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^
  File "/workspaces/pyerfa/erfa/core.py", line 115, in check_errwarn
    raise ErfaError(f'ERFA function "{func_name}" yielded {emsg}')
erfa.core.ErfaError: ERFA function "jd2cal" yielded 4 of "unacceptable date (Note 1)"
>>> erfa.ufunc.jd2cal(-600000., [0, 1, 2, 3])
(array([-1, -1, -1, -1], dtype=int32), array([-1, -1, -1, -1], dtype=int32), array([0, 0, 4, 0], dtype=int32), array([0.0e+000, 4.9e-324, 9.9e-324, 1.5e-323]), array([-1, -1, -1, -1], dtype=int32))

[neutrinoceros]

The result is different though, so I think you're hitting a separate failure mode