Use pivoted QR for replicate df and NaN inference for rank<=1 designs

Replace unpivoted QR with scipy pivoted QR (LAPACK dgeqp3) matching
R's qr(..., tol=1e-5) column-pivoting semantics for rank computation.

When a replicate design has undefined df (rank <= 1), force NaN
inference instead of falling through to z-distribution or model-based
df. Applies to LinearRegression.get_inference() and TripleDifference.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: igerber

diff_diff/linalg.py

@@ -2029,6 +2029,16 @@ def get_inference(
             effective_df = df
         elif self.survey_df_ is not None:
             effective_df = self.survey_df_
+        elif (hasattr(self, '_survey_design') and self._survey_design is not None
+              and hasattr(self._survey_design, 'uses_replicate_variance')
+              and self._survey_design.uses_replicate_variance):
+            # Replicate design with undefined df (rank <= 1) — NaN inference
+            warnings.warn(
+                "Replicate design has undefined survey d.f. (rank <= 1). "
+                "Inference fields will be NaN.",
+                UserWarning, stacklevel=2,
+            )
+            effective_df = 0  # Forces NaN from t-distribution
         else:
             effective_df = self.df_
 

diff_diff/survey.py

@@ -590,10 +590,13 @@ def df_survey(self) -> Optional[int]:
                 analysis_weights = self.replicate_weights
             else:
                 analysis_weights = self.replicate_weights * self.weights[:, np.newaxis]
-            # Use QR decomposition with R-compatible tolerance (1e-5)
-            Q, R_mat = np.linalg.qr(analysis_weights, mode='reduced')
+            # Pivoted QR with R-compatible tolerance, matching R's
+            # qr(..., tol=1e-5) which uses column pivoting (LAPACK dgeqp3)
+            from scipy.linalg import qr as scipy_qr
+            _, R_mat, _ = scipy_qr(analysis_weights, pivoting=True, mode='economic')
+            diag_abs = np.abs(np.diag(R_mat))
             tol = 1e-5
-            rank = int(np.sum(np.abs(np.diag(R_mat)) > tol * np.abs(np.diag(R_mat)).max()))
+            rank = int(np.sum(diag_abs > tol * diag_abs.max())) if diag_abs.max() > 0 else 0
             df = rank - 1
             return df if df > 0 else None
         if self.psu is not None and self.n_psu > 0:

diff_diff/triple_diff.py

@@ -590,7 +590,12 @@ def fit(
                 survey_metadata.df_survey = self._replicate_n_valid - 1
             # df <= 0 means insufficient rank for t-based inference
             if df is not None and df <= 0:
-                df = None
+                df = 0  # Forces NaN from t-distribution
+        elif (resolved_survey is not None
+              and hasattr(resolved_survey, 'uses_replicate_variance')
+              and resolved_survey.uses_replicate_variance):
+            # Replicate design with undefined df (rank <= 1) — NaN inference
+            df = 0  # Forces NaN from t-distribution
         else:
             df = n_obs - 8  # Approximate df (8 cell means)
             if covariates: