Reject replicate designs in TwoWayFixedEffects, fix compute_deff gate

TWFE performs weighted within-transformation once with full-sample
weights, then lets LinearRegression compute replicate variance on
the frozen transformed design. This is methodologically incorrect —
replicate variance requires recomputing the within-transform per
replicate (same reason as SunAbraham rejection).

Also fix compute_deff() to gate on survey design presence instead of
survey_df_, allowing DEFF computation even when replicate df is
undefined.

Fix REGISTRY fweight note consistency (non-negative, not positive).
Update replicate support matrix to include TWFE rejection.

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

Author: igerber

diff_diff/linalg.py

@@ -1885,7 +1885,7 @@ def compute_deff(self, coefficient_names=None):
         DEFFDiagnostics
         """
         self._check_fitted()
-        if self.survey_df_ is None:
+        if not (hasattr(self, 'survey_design') and self.survey_design is not None):
             raise ValueError(
                 "compute_deff() requires a survey design. "
                 "Fit with survey_design= first."

diff_diff/twfe.py

@@ -127,6 +127,16 @@ def fit(  # type: ignore[override]
         resolved_survey, survey_weights, survey_weight_type, survey_metadata = (
             _resolve_survey_for_fit(survey_design, data, self.inference)
         )
+        # Reject replicate-weight designs — TWFE within-transformation must
+        # be recomputed per replicate (same reason as SunAbraham rejection)
+        if resolved_survey is not None and resolved_survey.uses_replicate_variance:
+            raise NotImplementedError(
+                "TwoWayFixedEffects does not yet support replicate-weight "
+                "survey designs. The weighted within-transformation must be "
+                "recomputed for each replicate. Use CallawaySantAnna or "
+                "TripleDifference for replicate-weight inference, or use a "
+                "TSL-based survey design (strata/psu/fpc)."
+            )
 
         # Use unit-level clustering if not specified (use local variable to avoid mutation)
         cluster_var = self.cluster if self.cluster is not None else unit

docs/methodology/REGISTRY.md

@@ -1907,9 +1907,9 @@ unequal selection probabilities).
 - **Note:** pweight HC1 meat uses score outer products (Σ s_i s_i' where
   s_i = w_i x_i u_i), giving w² in the meat. fweight HC1 meat uses
   X'diag(w u²)X (one power of w), matching frequency-expanded HC1.
-- **Note:** fweights must be positive integers; fractional values are
-  rejected by `_validate_weights()`. This matches Stata's convention and
-  avoids ambiguous frequency semantics.
+- **Note:** fweights must be non-negative integers; fractional values are
+  rejected by `_validate_weights()`. All-zero vectors rejected at solver
+  level. This matches Stata's convention.
 
 ### Absorbed Fixed Effects with Survey Weights
 
@@ -2049,11 +2049,11 @@ variance from the distribution of replicate estimates.
   - **Supported**: CallawaySantAnna (reg, no bootstrap), ContinuousDiD
     (no bootstrap), EfficientDiD (no bootstrap), TripleDifference (all
     methods), LinearRegression (OLS path)
-  - **Rejected with NotImplementedError**: SunAbraham (within-transformation
-    must be recomputed per replicate), DifferenceInDifferences,
-    MultiPeriodDiD, StackedDiD (use compute_survey_vcov directly),
-    ImputationDiD, TwoStageDiD (custom variance), SyntheticDiD, TROP
-    (bootstrap-based variance)
+  - **Rejected with NotImplementedError**: SunAbraham, TwoWayFixedEffects
+    (within-transformation must be recomputed per replicate),
+    DifferenceInDifferences, MultiPeriodDiD, StackedDiD (use
+    compute_survey_vcov directly), ImputationDiD, TwoStageDiD (custom
+    variance), SyntheticDiD, TROP (bootstrap-based variance)
   - CS/ContinuousDiD/EfficientDiD reject replicate + `n_bootstrap > 0`
     (replicate weights provide analytical variance)
 - **Note:** When invalid replicates are dropped in `compute_replicate_vcov`