test(r_parity): update parity summary tables

Author: brycewang-stanford

tests/r_parity/results/parity_table.md

@@ -211,7 +211,7 @@ Generated by `tests/r_parity/compare.py` on the `results/<module>_{py,R}.json` a
 | `att_sdid` | -15.948389 | -15.948389 | 4.09e-14 | 2.56e-15 | — | — | — | — |
 
 ## Module 13_causal_forest
-- **methodological_disclosure**: T3 combined-Monte-Carlo-error pass: sp.causal_forest and grf both report the doubly-robust AIPW ATE, so the row is like-for-like and graded against combined sampling error rather than a fixed relative band. On the clean-overlap DGP the two agree within ~0.05 combined SE (worst rel gap below 0.3%), the AIPW recovery test certifies truth-recovery within 4 SE, and the B=1000 Track B row confirms calibration -- the two criteria formerly held open are now both satisfied.
+- **methodological_disclosure**: T3 combined-Monte-Carlo-error pass: sp.causal_forest and grf both report the doubly-robust AIPW ATE, so the row is like-for-like and graded against combined sampling error rather than a fixed relative band. On the clean-overlap DGP the two agree within ~0.05 combined SE (worst rel gap below 0.3%), the AIPW recovery tests certify truth-recovery within 4 SE and across multiple clean-overlap seeds, and the B=1000 Track B row confirms calibration -- the two criteria formerly held open are now both satisfied.
 - **n_estimators**: `2000`
 - **random_state**: `42`
 - **covariates**: `['x1', 'x2', 'x3', 'x4', 'x5']`
@@ -220,7 +220,7 @@ Generated by `tests/r_parity/compare.py` on the `results/<module>_{py,R}.json` a
 - **true_ate**: `1.0034061636502045`
 - **true_att**: `1.0044253057506767`
 - **dgp**: clean overlap: e(X)=0.5+0.2*tanh(X1) in [0.30,0.70]; tau(X)=1+0.5*X2; Y=X1+0.5*X3+tau*T+N(0,1); N=4000.
-- **note**: Both sp.causal_forest.average_treatment_effect and grf::average_treatment_effect report the AIPW doubly-robust ATE/ATT. On this clean-overlap DGP the StatsPAI row uses clip=0.25, consistent with the known e(X) support [0.30,0.70], and agrees with grf within combined Monte Carlo error (worst rel gap below 0.3%). The former NSW-DW stress case is now tracked in the Track B robustness sweep.
+- **note**: Both sp.causal_forest.average_treatment_effect and grf::average_treatment_effect report the AIPW doubly-robust ATE/ATT. On this clean-overlap DGP the StatsPAI row uses clip=0.25, consistent with the known e(X) support [0.30,0.70], and agrees with grf within combined Monte Carlo error (worst rel gap below 0.3%). A multi-seed clean-overlap pytest guard checks AIPW truth recovery for both ATE and ATT. The former NSW-DW stress case is now tracked in the Track B robustness sweep.
 
 | stat | py est | R est | abs Δ | rel Δ | py SE | R SE | abs Δ SE | rel Δ SE |
 |---|---:|---:|---:|---:|---:|---:|---:|---:|

tests/r_parity/results/parity_table_3way.md

@@ -240,7 +240,7 @@ Generated by `tests/r_parity/compare.py` on the `results/<module>_{py,R}.json` a
 
 ## Module 13_causal_forest
 - **strictness_tier**: `moderate` (moderate (≤5e-2))
-- **methodological_disclosure**: T3 combined-Monte-Carlo-error pass: sp.causal_forest and grf both report the doubly-robust AIPW ATE, so the row is like-for-like and graded against combined sampling error rather than a fixed relative band. On the clean-overlap DGP the two agree within ~0.05 combined SE (worst rel gap below 0.3%), the AIPW recovery test certifies truth-recovery within 4 SE, and the B=1000 Track B row confirms calibration -- the two criteria formerly held open are now both satisfied.
+- **methodological_disclosure**: T3 combined-Monte-Carlo-error pass: sp.causal_forest and grf both report the doubly-robust AIPW ATE, so the row is like-for-like and graded against combined sampling error rather than a fixed relative band. On the clean-overlap DGP the two agree within ~0.05 combined SE (worst rel gap below 0.3%), the AIPW recovery tests certify truth-recovery within 4 SE and across multiple clean-overlap seeds, and the B=1000 Track B row confirms calibration -- the two criteria formerly held open are now both satisfied.
 - **n_estimators**: `2000`
 - **random_state**: `42`
 - **covariates**: `['x1', 'x2', 'x3', 'x4', 'x5']`
@@ -249,7 +249,7 @@ Generated by `tests/r_parity/compare.py` on the `results/<module>_{py,R}.json` a
 - **true_ate**: `1.0034061636502045`
 - **true_att**: `1.0044253057506767`
 - **dgp**: clean overlap: e(X)=0.5+0.2*tanh(X1) in [0.30,0.70]; tau(X)=1+0.5*X2; Y=X1+0.5*X3+tau*T+N(0,1); N=4000.
-- **note**: Both sp.causal_forest.average_treatment_effect and grf::average_treatment_effect report the AIPW doubly-robust ATE/ATT. On this clean-overlap DGP the StatsPAI row uses clip=0.25, consistent with the known e(X) support [0.30,0.70], and agrees with grf within combined Monte Carlo error (worst rel gap below 0.3%). The former NSW-DW stress case is now tracked in the Track B robustness sweep.
+- **note**: Both sp.causal_forest.average_treatment_effect and grf::average_treatment_effect report the AIPW doubly-robust ATE/ATT. On this clean-overlap DGP the StatsPAI row uses clip=0.25, consistent with the known e(X) support [0.30,0.70], and agrees with grf within combined Monte Carlo error (worst rel gap below 0.3%). A multi-seed clean-overlap pytest guard checks AIPW truth recovery for both ATE and ATT. The former NSW-DW stress case is now tracked in the Track B robustness sweep.
 - **stata_status**: bridge artifact not materialized: Stata 19's official cate command is the candidate causal-forest/AIPW reference, but the verified local runtime is Stata 18 and `which cate` fails.
 
 | stat | py est | R est | Stata est | rel py-R | rel py-Stata | py SE | R SE | Stata SE | rel SE py-R | rel SE py-Stata |