Rank-based replicate df, mse=False default, update docs

- Replicate df now uses numerical rank of replicate weight matrix
  (matching R's survey::degf()) instead of hardcoded R-1
- Default mse=False (matching R's survey::svrepdesign() package default)
- Update REGISTRY.md to reflect both changes
- WLS identifiability for zero-weight subpopulation already handled
  by existing zero-mass guards in ContinuousDiD/TripleDifference

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

Author: igerber

diff_diff/survey.py

@@ -69,7 +69,7 @@ class SurveyDesign:
     combined_weights: bool = True
     replicate_scale: Optional[float] = None
     replicate_rscales: Optional[List[float]] = None
-    mse: bool = True
+    mse: bool = False
 
     def __post_init__(self):
         valid_weight_types = {"pweight", "fweight", "aweight"}
@@ -557,7 +557,7 @@ class ResolvedSurveyDesign:
     combined_weights: bool = True
     replicate_scale: Optional[float] = None
     replicate_rscales: Optional[np.ndarray] = None  # (R,) per-replicate scales
-    mse: bool = True
+    mse: bool = False
 
     @property
     def uses_replicate_variance(self) -> bool:
@@ -566,9 +566,18 @@ def uses_replicate_variance(self) -> bool:
 
     @property
     def df_survey(self) -> Optional[int]:
-        """Survey degrees of freedom: n_PSU - n_strata, or R-1 for replicates."""
+        """Survey degrees of freedom.
+
+        For replicate designs: numerical rank of centered replicate weight
+        matrix, matching R's ``survey::degf()``. For TSL: n_PSU - n_strata.
+        """
         if self.uses_replicate_variance:
-            return self.n_replicates - 1 if self.n_replicates > 1 else None
+            if self.replicate_weights is None or self.n_replicates < 2:
+                return None
+            # Rank-based df from replicate weight matrix, matching
+            # R's survey::degf() for svrepdesign objects
+            rank = int(np.linalg.matrix_rank(self.replicate_weights))
+            return max(rank - 1, 1) if rank > 1 else None
         if self.psu is not None and self.n_psu > 0:
             if self.strata is not None and self.n_strata > 0:
                 return self.n_psu - self.n_strata

docs/methodology/REGISTRY.md

@@ -2011,7 +2011,8 @@ variance from the distribution of replicate estimates.
   contrasts are formed via weight-ratio rescaling:
   `theta_r = sum((w_r/w_full) * psi)` when `combined_weights=True`,
   `theta_r = sum(w_r * psi)` when `combined_weights=False`.
-- **Survey df**: `R - 1` for replicate designs (replaces `n_PSU - n_strata`)
+- **Survey df**: Numerical rank of replicate weight matrix minus 1,
+  matching R's `survey::degf()`. Replaces `n_PSU - n_strata`.
 - **Mutual exclusion**: Replicate weights cannot be combined with
   strata/psu/fpc (the replicates encode design structure implicitly)
 - **Design parameters** (matching R `svrepdesign()`):
@@ -2020,8 +2021,9 @@ variance from the distribution of replicate estimates.
     by full-sample weight before WLS.
   - `replicate_scale`: override default variance scaling factor
   - `replicate_rscales`: per-replicate scaling factors (vector of length R)
-  - `mse` (default True): center variance on full-sample estimate. If False,
-    center on mean of replicate estimates.
+  - `mse` (default False, matching R's `survey::svrepdesign()`): if True,
+    center variance on full-sample estimate; if False, center on mean of
+    replicate estimates.
 - **Note:** Replicate columns are NOT normalized — raw values are preserved
   to maintain correct weight ratios in the IF path.
 - **Note:** JKn requires explicit `replicate_strata` (per-replicate stratum