You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Relax Wave B golden bit-identity to assert_allclose(rtol=1e-14)
CI Pure Python Fallback (Linux py3.14) drifted 1 ULP from the macOS Accelerate
capture machine on `test_event_study_false_matches_wave_b_golden` -- expected
-0.08620379515400438, got -0.08620379515400439. The 6 `==` checks against
_WAVE_B_GOLDEN_* are cross-machine pins, exactly the BLAS reduction-order
class that `feedback_assert_allclose_numerical_parity` warns about.
Switched all 6 golden assertions to `np.testing.assert_allclose(rtol=1e-14,
atol=1e-14)` -- tight enough to catch real aggregate-path drift, loose enough
to absorb cross-runner ULP differences. The same-machine determinism check
`test_event_study_false_bit_identical_to_wave_b_fixture` keeps `==` (both
fits run on the same runner).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
0 commit comments