stacked_did: address CI codex R1 P2 + P3 polish

CI codex R1 verdict was ✅ with 2 polish items:

P2 (Maintainability): StackedDiDResults.summary() didn't surface the
analytical variance family (vcov_type), even though hc1 vs hc2_bm are
now materially different user-facing inference modes. Other result
classes (DiD/MPD/TWFE) already print a Variance: line via the shared
`_format_vcov_label` helper at results.py:49-89. Fix: added a Variance:
line to StackedDiDResults.summary() using the shared helper, suppressed
under survey_metadata (the survey block already surfaces inference
source). My earlier plan note ("mirror SA: don't add Variance: row")
was based on a wrong assumption — SA-results uses the same shared
helper, and StackedDiD inherits the same precedent.

P3 (Performance): on `vcov_type="hc2_bm"` fits with `aggregate=None`
or `"simple"`, the code built per-event Bell-McCaffrey contrast DOFs
that were never surfaced (only overall ATT is exposed). Avoidable
CR2 cost on every overall-only hc2_bm fit. Fix: gated the per-event
contrast construction (es_keys, es_cols_full) on
`aggregate == "event_study"`. The overall ATT contrast still gets
computed unconditionally (it's the user-facing scalar).

89 tests still pass — no behavioral regression on the event-study
path; overall-only fits are now slightly faster.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: igerber

diff_diff/stacked_did.py

@@ -550,14 +550,18 @@ def fit(
             # Only build contrasts whose target column is identified; if a
             # delta_h column itself was dropped, that event-time will get
             # NaN inference (left to safe_inference's df=None path).
+            # Per CI codex R1 P3: skip per-event contrast DOFs when the
+            # event-study surface is not user-visible (aggregate != "event_study").
+            # The overall ATT contrast still gets computed below.
             es_keys: List[int] = []
             es_cols_full: List[np.ndarray] = []
-            for h in event_times:
-                if h in interaction_indices and _identified[interaction_indices[h]]:
-                    c = np.zeros(k_design)
-                    c[interaction_indices[h]] = 1.0
-                    es_keys.append(h)
-                    es_cols_full.append(c)
+            if aggregate == "event_study":
+                for h in event_times:
+                    if h in interaction_indices and _identified[interaction_indices[h]]:
+                        c = np.zeros(k_design)
+                        c[interaction_indices[h]] = 1.0
+                        es_keys.append(h)
+                        es_cols_full.append(c)
             # Overall ATT contrast: average of post-period delta_h columns
             # (the same 1/K * ones contrast used for overall_se below). Only
             # construct if ALL post-period delta_h are identified — otherwise

diff_diff/stacked_did_results.py

@@ -176,6 +176,28 @@ def summary(self, alpha: Optional[float] = None) -> str:
             "",
         ]
 
+        # Variance family label (per CI codex R1 P2): surface the analytical
+        # vcov_type when the survey path didn't override. Mirrors the shared
+        # `_format_vcov_label` helper from results.py used by DiD/MPD/TWFE.
+        # Suppress under survey_metadata since the survey block above already
+        # reports the inference source.
+        if self.survey_metadata is None and self.vcov_type:
+            from diff_diff.results import _format_vcov_label
+
+            # StackedDiD is intrinsically clustered; the cluster name is
+            # either "unit" or "unit_subexp" (config-time, not stored
+            # explicitly on results — derive from the stacked data if
+            # available, else fall back to generic label).
+            label = _format_vcov_label(
+                self.vcov_type,
+                cluster_name=None,  # cluster identity is config; print plain label
+                n_clusters=None,
+                n_obs=self.n_stacked_obs,
+            )
+            if label is not None:
+                lines.append(f"{'Variance:':<30} {label:>50}")
+                lines.append("")
+
         # Add survey design info
         if self.survey_metadata is not None:
             sm = self.survey_metadata