[bashtage]

Branch a90-fix-macos-intel-tests, two commits on main e677722.

The macos-15-intel wheel job fails on main, with a different test in each run:

• Run at e677722:
  test_valid_results_are_unchanged_by_numpy_scalars fails. It asserts that
  ZivotAndrews(Y, trim=np.float32(0.25)).stat ==
ZivotAndrews(Y, trim=0.25).stat, and the two differ in the last place
  (-4.3285219553671475 vs -4.328521955367147).
• Run at a93ff73: the same test fails, and so does
  test_float32_is_computed_in_float64[arch_model]. That second test passed in
  the later run.

Cause. Not float32 itself: float32 to float64 is exact, and both trims give
the same trim index. On macOS on x86-64, two calls with the same inputs are not
always bit-identical. This is most likely because the BLAS kernels' summation
order depends on memory alignment, which changes from call to call.

• The Zivot-Andrews comparison used ==.
• The float32 test of arch_model fits a GARCH(1,1) to i.i.d. data. The
  estimate is on the boundary (alpha = 0, beta = 1), where the likelihood is
  flat, so the optimizer turned a difference in the last bits into a relative
  difference of 3e-8 in mu and omega, against a tolerance of 1e-12.

Fix (tests only):

• Separate calls are compared with assert_allclose(..., rtol=1e-12): the
  ADF, Zivot-Andrews, mackinnonp and mackinnoncrit comparisons, and the
  float32 against float64 long-run covariance in test_covariance_dtypes.py.
• The two GARCH targets of the data type tables (arch_model and
  arch_model(mean="ARX")) evaluate the model at fixed parameters (the
  log-likelihood and the conditional volatility) instead of fitting it. They
  still exercise the float64 conversion and still detect a calculation in
  float32 (an error of about 1e-7), without the optimizer.

Fast suite 26,955 passed. No library code changes.

Also fixed: an order-dependent test on the Windows legs. The first CI run
of this PR failed on the two Windows legs:

• test_model_forecast_negative_start and test_align_horizon_is_target, from ENH: Accept 0 and 1 as flags, negative positions and the horizon alias of align #956.

Cause:

• The tests shared one model fixture for the whole module, and fit and fix
  store the sample on the model.
• When the result fixture had been created before the test that fits with
  first_obs=-20, the model kept that sample. Then forecast(start=-25)
  failed the check that start is not before the first fitted observation.

The tests now use a new model each, and result fits its own. Reproduced
locally with that order and fixed. The file passes in 8 random orders, and the
full fast suite passes in 2 random orders (26,955 passed). The other modules
with shared models passed in 5 random orders.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.97%. Comparing base (e677722) to head (c71ed59).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #959   +/-   ##
=======================================
  Coverage   98.97%   98.97%           
=======================================
  Files         136      136           
  Lines       30021    30028    +7     
  Branches     2327     2327           
=======================================
+ Hits        29712    29719    +7     
  Misses        255      255           
  Partials       54       54           

Flag	Coverage Δ
adder	98.91% <100.00%> (+<0.01%)	⬆️
subtractor	98.91% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.