Loosen affine-equivariance p-value tolerance (BLAS-dependent CI flake) (#777)

Author: igerber

tests/test_rddensity.py

@@ -505,7 +505,12 @@ def test_affine_equivariance(self):
         r1 = _fit_quiet(RDDensityTest(cutoff=0.0), x)
         r2 = _fit_quiet(RDDensityTest(cutoff=10.0), 3.5 * x + 10.0)
         assert r1.t_stat == pytest.approx(r2.t_stat, rel=1e-9)
-        assert r1.p_value == pytest.approx(r2.p_value, rel=1e-9)
+        # p = 2*Phi(-|t|) inherits the t-stat's relative error amplified by
+        # 2*|t|*phi(|t|)/p (~5.9x at this seed's |t|~2.26), so its tolerance
+        # must sit above the t-stat's or BLAS-kernel rounding differences
+        # across CI runners can trip it (run #3384: rel diff 1.15e-9 on
+        # ubuntu py3.14 only).
+        assert r1.p_value == pytest.approx(r2.p_value, rel=1e-8)
         # densities scale by 1/s
         assert r1.f_left == pytest.approx(r2.f_left * 3.5, rel=1e-9)
         assert df1 is not df2