feat(qa): recital detection, review split, duplicate comparison

Add _detect_recital() to QA pipeline — flags controls where
source_original_text contains Erwägungsgrund markers instead of
article text (28% of controls with source text affected).

- Recital detection via regex + phrase matching in QA validation
- 10 new tests (TestRecitalDetection), 81 total
- ReviewCompare component for side-by-side duplicate comparison
- Review mode split: Duplikat-Verdacht vs Rule-3-ohne-Anchor tabs
- MkDocs: recital detection documentation
- Detection script for bulk analysis (scripts/find_recital_controls.py)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

scripts/find_recital_controls.py

@@ -0,0 +1,79 @@
+"""Find controls where source_original_text contains Erwägungsgrund (recital) markers
+instead of actual article text — indicates wrong article tagging in RAG chunks."""
+
+import sqlalchemy
+import os
+import json
+import re
+
+url = os.environ.get("DATABASE_URL", "")
+if not url:
+    print("DATABASE_URL not set")
+    exit(1)
+
+engine = sqlalchemy.create_engine(url)
+
+with engine.connect() as conn:
+    conn.execute(sqlalchemy.text("SET search_path TO compliance,public"))
+
+    r = conn.execute(sqlalchemy.text("""
+        SELECT control_id, title,
+               source_citation::text,
+               source_original_text,
+               pipeline_version, release_state,
+               generation_metadata::text
+        FROM canonical_controls
+        WHERE source_original_text IS NOT NULL
+          AND source_original_text != ''
+          AND source_citation IS NOT NULL
+        ORDER BY control_id
+    """)).fetchall()
+
+    # Pattern: standalone recital number like (125)\n or (126) at line start
+    recital_re = re.compile(r'\((\d{1,3})\)\s*\n')
+
+    # Pattern: article reference like "Artikel 43" in the text
+    artikel_re = re.compile(r'Artikel\s+(\d+)', re.IGNORECASE)
+
+    suspects_recital = []
+    suspects_mismatch = []
+
+    for row in r:
+        cid, title, citation_json, orig, pv, state, meta_json = row
+        if not orig:
+            continue
+
+        citation = json.loads(citation_json) if citation_json else {}
+        claimed_article = citation.get("article", "")
+
+        # Check 1: Recital markers in source text
+        recital_matches = recital_re.findall(orig)
+        has_recital = len(recital_matches) > 0
+
+        # Check 2: Text mentions a different article than claimed
+        artikel_matches = artikel_re.findall(orig)
+        claimed_num = re.search(r'\d+', claimed_article).group() if re.search(r'\d+', claimed_article) else ""
+        different_articles = [a for a in artikel_matches if a != claimed_num] if claimed_num else []
+
+        if has_recital:
+            suspects_recital.append({
+                "control_id": cid,
+                "title": title[:80],
+                "claimed_article": claimed_article,
+                "claimed_paragraph": citation.get("paragraph", ""),
+                "recitals_found": recital_matches[:5],
+                "v": pv,
+                "state": state,
+            })
+
+    print(f"=== Ergebnis ===")
+    print(f"Geprueft: {len(r)} Controls mit source_original_text")
+    print(f"Erwaegungsgrund-Verdacht: {len(suspects_recital)}")
+    print()
+
+    if suspects_recital:
+        print(f"{'Control':<12} {'Behauptet':<18} {'Recitals':<20} {'v':>2} {'State':<15} Titel")
+        print("-" * 120)
+        for s in suspects_recital:
+            recitals = ",".join(s["recitals_found"])
+            print(f"{s['control_id']:<12} {s['claimed_article']:<10} {s['claimed_paragraph']:<7} ({recitals}){'':<{max(0,15-len(recitals))}} v{s['v']} {s['state']:<15} {s['title']}")