Add cp314t build

[bashtage]

closes #208

[rgommers]

The 3.14t-specific failures on macOS, Linux and x86-64 Windows are all the same:

  ================================== FAILURES ===================================
  _________________ test_negative_binomial_default_alpha_param __________________
  [gw1] win32 -- Python 3.14.5 C:\Users\runneradmin\AppData\Local\Temp\cibw-run-23n5l6d3\cp314t-win_amd64\venv-test\Scripts\python.exe
  
      def test_negative_binomial_default_alpha_param():
          with pytest.warns(
              UserWarning, match="Negative binomial dispersion parameter alpha not set"
          ):
              sm.families.NegativeBinomial()
          with pytest.warns(
              UserWarning, match="Negative binomial dispersion parameter alpha not set"
          ):
              sm.families.NegativeBinomial(link=sm.families.links.NegativeBinomial(alpha=1.0))
          with warnings.catch_warnings():
              warnings.simplefilter("error")
  >           sm.families.NegativeBinomial(alpha=1.0)
  
  ..\venv-test\Lib\site-packages\statsmodels\discrete\tests\test_discrete.py:2681: 
  _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
  
  self = <statsmodels.genmod.families.family.NegativeBinomial object at 0x0000026651B65290>
  link = None, alpha = 1.0, check_link = True
  
      def __init__(self, link=None, alpha=1.0, check_link=True):
          self.alpha = 1.0 * alpha  # make it at least float
          if alpha is self.__init__.__defaults__[1]:  # `is` is intentional
  >           warnings.warn(
                  "Negative binomial dispersion parameter alpha not "
                  f"set. Using default value alpha={alpha}.",
                  ValueWarning,
                  stacklevel=2,
              )
  E           statsmodels.tools.sm_exceptions.ValueWarning: Negative binomial dispersion parameter alpha not set. Using default value alpha=1.0.

That seems relatively benign, and is possibly related to the different default for warnings behavior under free-threading: https://docs.python.org/3.14/whatsnew/3.14.html#concurrent-safe-warnings-control.

The Windows on Arm jobs show crashes, but on all of cp313, cp314 and cp314t jobs. May be useful to try to diagnose one of the GIL-enabled builds first, and then see if anything is left.

https://github.com/MacPython/statsmodels-wheels/actions/runs/27669875530/job/81831635348?pr=212

[bashtage]

The windows arm failures are expected. They have been failing since they were first started. There appears to be OpenBlas bug(s). These have been reported to NumPy.

I agree that the warnings failure is pretty important. The bigger issue is that other parts segfault under free threading when actually run in parallel. But I suppose we can still have a cp314t wheel even if it isn't actually threadsafe.

[kumaraditya303]

The bigger issue is that other parts segfault under free threading when actually run in parallel. But I suppose we can still have a cp314t wheel even if it isn't actually threadsafe.

The statespace tests segfault when run in parallel because the testa are written in an unsafe way. The class TestDynamicFactor shares a single model across multiple threads and when tests call filter on it in test_params it crashes. Fixing this would require porting all of those tests to keep the model either thread local or create fresh model for each thread.

[bashtage]

I'm happy with more models. Cold you point me to an example where a model is shared? I'm happy to use some tokens to tackle this.

[kumaraditya303]

I'm happy with more models. Cold you point me to an example where a model is shared? I'm happy to use some tokens to tackle this.

See https://github.com/statsmodels/statsmodels/blob/cfae27cdd96c2fda81eeac682308092432d211eb/statsmodels/tsa/statespace/tests/test_dynamic_factor.py#L28-L63 During the setup of test it assigns a model to a cls.model which gets used concurrently by all threads. The test_params test also modifies the model by changing enforce_stationarity which also needs to be fixed.

[kumaraditya303]

I submitted statsmodels/statsmodels#9877 to fix the test_negative_binomial_default_alpha_param failures.