Add survey real-data validation against R (API, NHANES, RECS)

.github/workflows/ai_pr_review.yml

@@ -148,7 +148,11 @@ jobs:
             git --no-pager diff --name-status "$BASE_SHA" "$HEAD_SHA"
             echo ""
             echo "Unified diff (context=5):"
-            git --no-pager diff --unified=5 "$BASE_SHA" "$HEAD_SHA"
+            # Exclude large generated data files from the full diff to stay
+            # within the model's input limit. The --stat above still lists them.
+            git --no-pager diff --unified=5 "$BASE_SHA" "$HEAD_SHA" \
+              -- . ':!benchmarks/data/**/*.json' ':!benchmarks/data/**/*.csv' \
+                   ':!docs/tutorials/*.ipynb'
           } >> "$PROMPT"
 
       - name: Run Codex

.gitignore

@@ -55,6 +55,7 @@ Thumbs.db
 
 # Benchmarks - generated data and results (can be regenerated)
 benchmarks/data/synthetic/*.csv
+benchmarks/data/real/raw/
 benchmarks/results/
 
 # Rust build artifacts

CHANGELOG.md

@@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+- **Survey real-data validation** (Phase 9) — 15 cross-validation tests against R's `survey` package using three real federal survey datasets:
+  - **API** (R `survey` package): TSL variance with strata, FPC, subpopulations, covariates, and Fay's BRR replicates
+  - **NHANES** (CDC/NCHS): TSL variance with strata + PSU + nest=TRUE, validating the ACA young adult coverage provision DiD
+  - **RECS 2020** (U.S. EIA): JK1 replicate weight variance with 60 pre-computed replicate columns
+  - All metrics (ATT, SE, df, CI) match R to machine precision (< 1e-10 differences)
+
 ## [2.8.4] - 2026-04-04
 
 ### Added