Relax Bacon survey validation to exact-only and document fweight exclusions from PR #226 review (round 19)

- BaconDecomposition: move _validate_unit_constant_survey behind
  weights="exact" — approximate path uses obs-level means and doesn't
  need within-unit-constant survey columns
- REGISTRY.md: document StackedDiD fweight rejection and Bacon
  exact-only unit-constancy requirement
- survey-roadmap.md: update StackedDiD support to "Full (pweight/aweight)"

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

Author: igerber

diff_diff/bacon.py

@@ -478,8 +478,11 @@ def fit(
             _resolve_survey_for_fit(survey_design, data, "analytical")
         )
 
-        # Validate within-unit constancy for panel survey designs
-        if resolved_survey is not None:
+        # Validate within-unit constancy for exact survey weights only.
+        # The exact-weight path collapses to per-unit weights via groupby().first(),
+        # which requires constant survey columns within units. The approximate path
+        # uses observation-level weighted means and does not need this constraint.
+        if resolved_survey is not None and self.weights == "exact":
             from diff_diff.survey import _validate_unit_constant_survey
 
             _validate_unit_constant_survey(data, unit, survey_design)

docs/methodology/REGISTRY.md

@@ -1021,7 +1021,7 @@ The paper text states a stricter bound (T_min + 1) but the R code by the co-auth
 - [x] Anticipation parameter support
 - [x] Never-treated encoding (0 and inf)
 - [x] Survey design support (Phase 3): Q-weights compose multiplicatively with survey weights; TSL vcov on composed weights; survey design columns propagated through sub-experiments
-- **Note:** Survey weights compose multiplicatively with Q-weights for StackedDiD
+- **Note:** Survey weights compose multiplicatively with Q-weights for StackedDiD; `weight_type="fweight"` is rejected because Q-weight composition produces non-integer composed weights that break frequency-weight semantics
 
 ---
 
@@ -1576,7 +1576,7 @@ Weights depend on group sizes and variance in treatment timing.
 - [ ] TWFE coefficient ≈ weighted sum of 2×2 estimates
 - [ ] Visualization shows weight vs. estimate by comparison type
 - [x] Survey design support (Phase 3): weighted cell means, weighted within-transform, weighted group shares
-- **Note:** Bacon decomposition with survey weights is diagnostic; exact-sum guarantee is approximate
+- **Note:** Bacon decomposition with survey weights is diagnostic; exact-sum guarantee is approximate; `weights="exact"` requires within-unit-constant survey columns (approximate path accepts time-varying weights)
 
 ---
 

docs/survey-roadmap.md

@@ -18,7 +18,7 @@ Phases 1-3 are implemented. Phases 4-5 are deferred for future PRs.
 
 | Estimator | File | Survey Support | Notes |
 |-----------|------|----------------|-------|
-| StackedDiD | `stacked_did.py` | Full | Q-weights compose multiplicatively with survey weights; TSL vcov on composed weights |
+| StackedDiD | `stacked_did.py` | Full (pweight/aweight) | Q-weights compose multiplicatively with survey weights; TSL vcov on composed weights; fweight rejected (non-integer composition) |
 | SunAbraham | `sun_abraham.py` | Full | Survey weights in LinearRegression + weighted within-transform; bootstrap+survey deferred |
 | BaconDecomposition | `bacon.py` | Diagnostic | Weighted cell means, weighted within-transform, weighted group shares; no inference (diagnostic only) |
 | TripleDifference | `triple_diff.py` | Reg only | Regression method with weighted OLS + TSL on influence functions; IPW/DR deferred (needs weighted `solve_logit()`) |