… (R1 P0)

Local codex R1 caught a P0: StackedDiD(vcov_type="hc2_bm") computed CR2
vcov correctly but never propagated the Bell-McCaffrey Satterthwaite DOF
into safe_inference() calls. event_study_effects[h]['p_value']/['conf_int']
and overall_p_value/overall_conf_int silently fell back to normal-theory
inference (df=None ⇒ scipy.norm), contradicting the registry contract.

Fix mirrors the SunAbraham aggregated-inference pattern from PR #472
(sun_abraham.py:997-1097). After solve_ols(), if vcov_type=="hc2_bm" and
not on the survey replicate-refit path, build contrast matrix:
  - Per-event-time: unit vector at each interaction_indices[h]
  - Overall ATT: 1/K average across post-period interaction columns
Call _compute_cr2_bm_contrast_dof(X, cluster_ids, bread, contrasts,
weights=composed_weights). Apply per-event-time DOFs to event_study_effects
inference, overall DOF to overall_* inference. Wrap in try/except so any
rank-deficient or linalg failure emits a UserWarning and falls back to
normal-theory (visible deviation, not silent).

R fixture extended with the post-period-average ATT contrast DOF via
Wald_test(constraints=row_avg, vcov=CR2, test="HTZ")$df_denom (mirrors PR
#465's MPD avg_att approach). New goldens at both cluster=unit and
cluster=unit_subexp.

Test additions / strengthening (addresses R1 P2 + P3):
- test_hc2_bm_per_event_dof_matches_coef_test_df_satt_unit_cluster:
  uses brentq inversion of CI half-width to recover the DOF safe_inference
  actually used. If propagation failed (the R1 P0 bug), the inversion
  raises ValueError → test FAILS instead of silently skipping (replaces
  the prior `continue`-on-failure pattern which could vacuously pass).
  Hard-asserts validated_count == len(event_times).
- test_hc2_bm_overall_att_dof_matches_wald_test_htz_unit_cluster (NEW):
  pins overall ATT DOF at atol=1e-6 against R Wald_test(HTZ)$df_denom.
- test_hc2_bm_overall_att_dof_matches_wald_test_htz_unit_subexp_cluster
  (NEW): symmetric coverage at alternate cluster level.
- Renamed CR1 → CR1S throughout docs/tests/REGISTRY for consistency
  (diff-diff's HC1+cluster uses Stata-style G/(G-1)*(n-1)/(n-p); plain
  CR1 omits the (n-1)/(n-p) term and diverges by ~1.4%).

192 tests pass (74 stacked + 19 wls_cr2 + 47 SA + 52 estimators_vcov_type).

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

Local codex R2 caught a P1: on rank-deficient stacked designs, the BM
contrast DOF block called `_compute_cr2_bm_contrast_dof()` on the singular
full bread matrix, which raised LinAlgError and triggered the catch-and-
fallback path — downgrading the aggregated inference to normal-theory
even when the target event-time delta_h coefficients were still identified.

Fix mirrors the MultiPeriodDiD rank-deficient pattern (PR #465,
estimators.py:1860-1913):
- Derive `_identified = ~np.isnan(coef)` (solve_ols emits NaN for dropped
  columns under R-style rank handling).
- Subset X, bread, and contrasts to the identified-column block BEFORE
  calling `_compute_cr2_bm_contrast_dof`.
- Only build per-event-time contrasts for event-times whose target delta_h
  column is identified; only build the overall ATT contrast when ALL
  post-period delta_h are identified (otherwise the contrast is undefined).
- The remaining try/except is a defensive guard for genuine singularities
  on the identified-column design (rare; the rank-deficient handling
  already removes the problematic columns).

The earlier behavior was too aggressive: a single dropped nuisance column
would knock out BM DOF for ALL contrasts. The new behavior NaN-guards
only contrasts that target a dropped column.

New regression test `test_hc2_bm_rank_deficient_design_keeps_bm_dof_*` —
verifies the fit doesn't emit fallback warnings on a well-conditioned
design + checks that CI half-widths use t(BM DOF) instead of z=1.96
(catches the R1/R2 normal-fallback failure mode end-to-end).

75 tests pass across stacked + methodology_stacked_did.

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

Local codex R3 caught a P1 on the same surface: when BM contrast DOF was
unavailable (helper raised OR noise-floor NaN guard fired), StackedDiD
either fell back to normal-theory inference (silent wrong CIs/p-values
under the hc2_bm contract) or passed NaN df through safe_inference (which
guards df<=0 but NOT NaN, producing finite t_stat with NaN p/CI — mixed
inconsistent inference fields). Both fail-open patterns contradict the
registry contract that hc2_bm aggregated inference uses BM DOF.

Fix mirrors the LinearRegression.get_inference() pattern from PR #475 R7
(linalg.py:3689-3706): on the hc2_bm path, when the per-contrast BM DOF
is None or non-finite, emit ALL-NaN inference fields for that contrast
rather than falling back to safe_inference(df=None or NaN). Effect and SE
remain finite — only the inference fields downstream of the DOF are
suppressed.

Applied at both inference sites:
- Event-study per-event-time inference (event_study_effects[h]):
  inline check before safe_inference call.
- Overall ATT inference (overall_t/p/conf_int): same guard.

New regression tests:
- test_hc2_bm_nan_dof_fails_closed_with_all_nan_inference: monkeypatches
  `_compute_cr2_bm_contrast_dof` to return a NaN-only vector and asserts
  ALL event-study + overall t_stat/p_value/conf_int are NaN while effect
  and se remain finite.
- test_hc2_bm_helper_raises_fails_closed_with_all_nan_inference:
  monkeypatches the helper to raise LinAlgError and asserts the
  fallback path NaN-closes inference (does NOT use normal-theory).

77 tests pass. Pattern is consistent with the analytical-surface
fail-closed contract added in PR #475 R7.

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

…drift lock)

R4 verdict was ✅ with 3 polish items:

P2: helper-exception warning text was stale (said "Falling back to normal
distribution" but the code now NaN-closes per the R3 P1 fix). Rewrote
the warning to accurately describe the fail-closed contract: "Aggregated
p-values, t-statistics, and confidence intervals will be returned as NaN
to preserve the hc2_bm contract".

P3: set_params(vcov_type=...) bypassed the __init__ validation guard —
sklearn-style mutation could set arbitrary values and fail later in the
linalg layer with a different message. Fix: factored the vcov_type
validation into `StackedDiD._validate_vcov_type(staticmethod)` and call
it from BOTH __init__ AND set_params. New test
`test_set_params_revalidates_vcov_type` exercises all 4 reject paths.

P3: test_replicate_refit_coef_bit_equal_vs_bake_w_baseline was described
as a "drift lock" but only checked finiteness — wouldn't catch numeric
drift. Captured baseline overall_att=2.0807833161293945 +
overall_se=0.14697256517699428 on the post-PR branch and asserted with
atol=1e-13 (coef) / 1e-10 (SE, compounded through replicate-refit
variance aggregation). Now the test actually catches drift.

78 tests pass (4 set_params reject paths + strengthened drift lock).

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

…eficient test)

R5 verdict was ✅ with 2 P3 polish items:

P3 (Maintainability): the new solve_ols(weights=...) call sites relied on
the default weight_type="pweight". Locked the WLS-CR2 weight convention
locally by passing weight_type="pweight" explicitly in BOTH solve_ols
calls (main fit + _refit_stacked replicate closure). Prevents silent
variance-contract drift if solve_ols's default ever changes.

P3 (Docs/Tests): test_hc2_bm_rank_deficient_design_keeps_bm_dof_on_
identified_contrasts didn't actually create a rank-deficient design — it
only verified the t-distribution was used on a well-conditioned fixture.
Two-part fix:
  1. Renamed the existing test to
     test_hc2_bm_uses_t_distribution_not_normal_on_well_conditioned_design
     (honest description: it's an end-to-end smoke check that BM DOF
     threading reaches safe_inference).
  2. NEW test test_hc2_bm_rank_deficient_keeps_bm_dof_on_identified_
     contrasts: monkeypatches solve_ols to inject NaN at the intercept
     column, simulating rank deficiency. Exercises the reduced-design
     code path at stacked_did.py:529-577 added in the R2 P1 fix.
     Verifies: (i) no fallback warning (reduced-design helper succeeds);
     (ii) identified delta_h coefficients keep finite t-distribution
     inference (CI half-width > z_0.975 * SE); (iii) overall ATT also
     keeps t-distribution.

79 tests pass.

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

R6 verdict was ✅ with 2 P3 test-coverage polish items:

P3 (test): prior R fixture didn't pin the overall delta-method SE for
the post-period-average ATT contrast — only BM DOF. Post-period
covariance regressions could slip through. Fix: added `se_overall_cr1`
and `se_overall_cr2` to the R fixture (computed as
`c @ vcov @ c.T` for the post-period 1/K row contrast at both
cluster levels). Added overall_se parity assertions at atol=1e-10 to
the three existing hc1/hc2_bm parity tests.

P3 (test): the Python parity test regenerated the panel via
`generate_staggered_data(...)` instead of cross-checking against the
committed CSV. Risk: future DGP-helper drift would silently invalidate
R-parity assertions. Fix: added a DGP-drift check in the `panel`
fixture that loads `benchmarks/data/stacked_did_test_panel.csv`,
merges with the regenerated panel on (unit, period), and asserts the
`outcome` column matches at atol=1e-12. Catches any drift in the DGP
helper with a clear error message rather than wrong parity output.
(first_treat column not checked because inf-vs-0 recoding differs
between CSV/pandas-read and regenerated panel; outcome is what
actually drives the R-side fit.)

79 tests pass.

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

…ning

R7 verdict was ⚠️ Needs changes with one P1 + one P3:

P1: BM-DOF logic was only pinned on default (anticipation=0,
weighting='aggregate'). The BM-DOF code path uses anticipation in
post-period selection (`h >= -self.anticipation`) and consumes the
composed Q-weights directly — both surfaces were untested. Added 4
parameter-interaction tests:

- test_anticipation_plus_hc2_bm_threads_bm_dof: anticipation=1, asserts
  shifted reference period (e=-2 has SE=0), overall ATT contrast uses
  t(BM DOF) (CI half-width > z_0.975 * SE), at least one event-study CI
  uses t-distribution.
- test_population_weighting_plus_hc2_bm_finite_threads_bm_dof:
  weighting='population' (different Q-weight formula), verifies finite
  output + BM DOF propagation via t-distribution check.
- test_sample_share_weighting_plus_hc2_bm_finite_threads_bm_dof:
  weighting='sample_share' (third Q-weight formula), symmetric coverage.
- test_hc1_vs_hc2_bm_differ_under_anticipation: sanity check that hc1
  vs hc2_bm SEs actually differ under anticipation=1 (proves the
  leverage/DOF adjustment fires under non-default params).

These tests don't require R goldens (no methodology change vs default
case; just verifies the BM-DOF code path remains active under the
parameter variants). R parity on the default path is already strong.

P3: fixture-independence concern from R6 was only partially closed.
Now also pins the `first_treat` column equality after normalizing
inf/0 between CSV (pandas reads inf for never-treated) and regenerated
(0 for never-treated in some pandas versions). Catches cohort-assignment
drift independently of outcome drift.

83 tests pass (+4 new R7 tests).

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

…ion')

R8 verdict was ⚠️ still Needs changes: R7's parameter-interaction tests
were smoke checks (population test used constant pop=1.0 which is
numerically degenerate with aggregate weighting on a balanced panel);
no R goldens for non-default variants. Closes the gap with numerical
parity at atol=1e-10.

Source-tree additions:
- benchmarks/data/stacked_did_anticipation1_panel.csv (NEW): pre-stacked
  fixture with anticipation=1, generated from same base panel via
  StackedDiD(anticipation=1).fit(). Event times [-3,-2,-1,0,1,2] with
  ref=-2 (shifted by anticipation).
- benchmarks/data/stacked_did_population_panel.csv (NEW): pre-stacked
  fixture with weighting='population' AND heterogeneous unit populations
  (rng.uniform(1,100), seed=99). Q-weights range [0.68, 1.54] vs uniform
  ~1.0 under aggregate weighting; forces the population-formula branch.
- benchmarks/R/generate_stacked_did_golden.R: refactored to loop over
  3 variants (default, population, anticipation1), producing
  variant-keyed entries in `stacked_did_golden.json`. Post-period filter
  now generalized to `h >= -anticipation` (matches Python's
  `_post_event_times_preview` at stacked_did.py:546-548).
- benchmarks/data/stacked_did_golden.json: regenerated with all 3 variants.

Test additions (tests/test_methodology_stacked_did.py):
- new class TestStackedDiDParityRVariants (4 tests):
  - test_population_weighting_hc1_matches_clubsandwich_cr1s
  - test_population_weighting_hc2_bm_matches_clubsandwich_cr2
  - test_anticipation1_hc1_matches_clubsandwich_cr1s
  - test_anticipation1_hc2_bm_matches_clubsandwich_cr2
  All assert event-study SE AND overall_se at atol=1e-10.
- new fixture `population_panel` reproduces the heterogeneous-population
  seed (np.random.default_rng(99)) used by the R variant CSV generator.

87 tests pass (+4 variant parity).

Note: weighting='sample_share' parity not added because its Q-weight
formula structurally mirrors population weighting (both differ from
aggregate only in the Q-weight calculation, which feeds into the same
solve_ols(weights=) path). The population test exercises that code
path; sample_share remains covered by the smoke test from R7.

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

R9 verdict was ⚠️ still on parameter-coverage P1: sample_share had no
R-parity fixture, and population/anticipation1 tests asserted SE only
(not BM DOF).

Source-tree additions:
- benchmarks/data/stacked_did_sample_share_panel.csv (NEW): pre-stacked
  fixture for weighting='sample_share'.
- benchmarks/R/generate_stacked_did_golden.R: added sample_share variant
  to the variants list; loops over all 4 variants now.
- benchmarks/data/stacked_did_golden.json: regenerated with sample_share.

Test additions (tests/test_methodology_stacked_did.py):
- test_sample_share_weighting_hc1_matches_clubsandwich_cr1s
- test_sample_share_weighting_hc2_bm_matches_clubsandwich_cr2 (+ DOF)
- Extended test_population_weighting_hc2_bm_matches_clubsandwich_cr2:
  added per-event-time + overall BM DOF parity via CI-width inversion.
- Extended test_anticipation1_hc2_bm_matches_clubsandwich_cr2: same DOF
  parity additions.

89 tests pass. All 3 non-default variants now have full R parity:
event-study SE + overall_se + per-event BM DOF + overall BM DOF.

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