Clear _replicate_df on refit, add CS zero-mass guards, document IF df

- Reset _replicate_df = None at start of LinearRegression.fit() so
  reused model instances don't carry stale replicate df
- Add zero-effective-mass guards in CS _att_gt_* for survey-weighted
  treated/control groups — return NaN for cells with zero weight
- Document in REGISTRY that IF-based replicate contrasts cannot fail
  (weighted sums are always finite), so n_valid == R and df propagation
  is not needed for IF paths

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

Author: igerber

diff_diff/linalg.py

@@ -1659,6 +1659,9 @@ def fit(
         X = np.asarray(X, dtype=np.float64)
         y = np.asarray(y, dtype=np.float64)
 
+        # Reset replicate df from any previous fit
+        self._replicate_df = None
+
         # Add intercept if requested
         if self.include_intercept:
             X = np.column_stack([np.ones(X.shape[0]), X])

diff_diff/staggered.py

@@ -604,6 +604,12 @@ def _compute_att_gt_fast(
         sw_treated = survey_w[treated_valid] if survey_w is not None else None
         sw_control = survey_w[control_valid] if survey_w is not None else None
 
+        # Guard against zero effective mass after subpopulation filtering
+        if sw_treated is not None and np.sum(sw_treated) <= 0:
+            return {"att": np.nan, "se": np.nan}
+        if sw_control is not None and np.sum(sw_control) <= 0:
+            return {"att": np.nan, "se": np.nan}
+
         # Get covariates if specified (from the base period)
         X_treated = None
         X_control = None

docs/methodology/REGISTRY.md

@@ -2029,9 +2029,11 @@ variance from the distribution of replicate estimates.
 - **Note:** CallawaySantAnna, ContinuousDiD, and EfficientDiD reject
   replicate weights with `n_bootstrap > 0`. Replicate weights provide
   analytical variance; bootstrap is a separate inference mechanism.
-- **Note:** When invalid replicates are dropped, `n_replicates` is updated
-  to the valid count so that `df_survey = n_valid - 1` and downstream
-  t-based inference uses the correct degrees of freedom.
+- **Note:** When invalid replicates are dropped in `compute_replicate_vcov`
+  (OLS path), `n_valid` is returned and used for `df_survey = n_valid - 1`
+  in `LinearRegression.fit()`. For IF-based replicate paths, replicates
+  essentially never fail (weighted sums cannot be singular), so `n_valid`
+  equals `R` in practice and df propagation is not needed.
 
 ### DEFF Diagnostics (Phase 6)