fix(rdd): cross-BLAS tolerance for constant-covariate exclusion tests

CI round 1 (Linux/Windows/pure-python; macOS passed): the two
constant-covariate exclusion tests asserted EXACT equality between the
fit with the excluded constant column and the fit without it. The
equality is exact mathematically (the excluded column's gamma row is
exactly 0) but only roundoff-level numerically: the response matrix
still carries the excluded column, and BLAS matmul kernels differ with
matrix shape across platforms (bit-equal on Accelerate, last-ULP
differences on OpenBLAS/Windows). Asserts relaxed to rel=1e-12 and the
"bit-for-bit" claims in the port docstring / REGISTRY / CHANGELOG
softened to "to floating-point roundoff".

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QGca52n6H8oDDXALjjrsp4

Author: igerber

CHANGELOG.md

@@ -44,9 +44,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   `covs_drop` echo on the results. Degenerate adjustments are GUARDED, not
   reproduced (documented deviation): R's `ginv(tol=1e-20)` inverts a float-noise
   singular value on constant covariates / full dummy sets, silently returning
-  platform-dependent estimates; diff-diff excludes degenerate columns, applies a
+  platform-dependent estimates; diff-diff excludes degenerate columns (a constant
+  covariate reproduces the fit without it to numerical precision), applies a
   scale-invariant stabilized cut for rank-deficient sets (a full dummy set
-  reproduces the drop-one-category fit exactly), and warns naming the columns.
+  reproduces the drop-one-category fit to numerical precision), and warns
+  naming the columns.
   **Canonical binding:** `att`/`se`/`t_stat`/`p_value`/`conf_int` are ONE
   coherent row - the robust bias-corrected row (`att = tau_bc`, CI centered on it,
   `t_stat == att/se`), preserving the library-wide field identities; the new

diff_diff/_rdrobust_port.py

@@ -439,8 +439,11 @@ def _covs_gamma(
 
     * per-column: ``diag(ZWZ)_j / (z_j' W z_j) < 1e-14`` means column j is
       numerically fully explained by the design -> excluded (gamma row 0;
-      a constant covariate then contributes exactly nothing, matching the
-      fit without it bit-for-bit);
+      a constant covariate then contributes exactly nothing and the fit
+      matches the one without it to floating-point roundoff - not
+      bit-for-bit, because the response matrix still carries the excluded
+      column and BLAS matmul kernels differ with matrix SHAPE on some
+      platforms);
     * set-level: equilibrated (scale-invariant) singular values of the
       remaining block with ``sv_min < 1e-12 * sv_max`` -> stabilized
       equilibrated pseudo-inverse with the noise directions cut

docs/methodology/REGISTRY.md

@@ -3854,7 +3854,10 @@ per-side and comb chains.
   (~0.5% in the smoke). The port instead (a) EXCLUDES per-column
   degeneracies - explained-ratio `diag(ZWZ)_j/(z_j'Wz_j) < 1e-14` -
   zeroing their gamma rows, so a constant covariate reproduces the fit
-  without it bit-for-bit; (b) cuts SET-level noise directions with an
+  without it to floating-point roundoff (tested at rel 1e-12; not
+  bit-for-bit - the response matrix still carries the excluded column
+  and BLAS matmul kernels differ with matrix shape across platforms);
+  (b) cuts SET-level noise directions with an
   equilibrated (scale-invariant) pseudo-inverse (`sv_min < 1e-12 *
   sv_max` -> `rcond=1e-12` cut), so a full dummy set reproduces the
   drop-one-category fit (span invariance, tested at rel 1e-9); and (c)

tests/test_rdd_methodology.py

@@ -518,8 +518,13 @@ def test_constant_covariate_excluded_warns(self):
             fit_c = RegressionDiscontinuity(h=0.3).fit(df, "y", "x", covariates=["zlong", "zconst"])
         fit_1 = RegressionDiscontinuity(h=0.3).fit(df, "y", "x", covariates=["zlong"])
         assert fit_c.covariate_coefficients["zconst"] == 0.0
-        assert fit_c.att == fit_1.att
-        assert fit_c.se == fit_1.se
+        # Equality holds mathematically (the excluded column's gamma is
+        # exactly 0) but only to float roundoff numerically: the response
+        # matrix still carries the constant column, and BLAS matmul
+        # kernels differ with matrix SHAPE across platforms (bit-equal on
+        # Accelerate, last-ULP off on OpenBLAS/Windows - CI round 1).
+        assert fit_c.att == pytest.approx(fit_1.att, rel=1e-12)
+        assert fit_c.se == pytest.approx(fit_1.se, rel=1e-12)
 
     def test_dummy_set_stabilized_equals_drop_one(self):
         # A full one-hot set passes the covariate-only QR but is

tests/test_rdrobust_port.py

@@ -864,8 +864,12 @@ def test_constant_covariate_excluded_equals_fit_without_it(self):
         assert fit_c.covs_excluded.tolist() == [False, True]
         # gamma row zeroed -> the constant contributes exactly nothing.
         assert fit_c.gamma_p is not None and fit_c.gamma_p[1, 0] == 0.0
-        assert fit_c.tau_bc == fit_1.tau_bc
-        assert fit_c.se_rb == fit_1.se_rb
+        # Mathematically identical; numerically only to float roundoff -
+        # the (n, 3) vs (n, 2) response shapes route through different
+        # BLAS matmul kernels on some platforms (CI round 1: last-ULP
+        # diffs on OpenBLAS/Windows, bit-equal on Accelerate).
+        assert fit_c.tau_bc == pytest.approx(fit_1.tau_bc, rel=1e-12)
+        assert fit_c.se_rb == pytest.approx(fit_1.se_rb, rel=1e-12)
 
     def test_dummy_set_stabilized_equals_drop_one(self):
         # A full one-hot set passes the intercept-free QR (rank 3) but the