feat: dp_finalizer — arbitrage Trackare vs CRH-only avec traçabilité audit

Nouveau module src/medical/dp_finalizer.py :
- 5 règles d'arbitrage (R1-R5) : CRH CONFIRMED override, Trackare corroboré,
  symptôme R* override/review, ambigu REVIEW, Z-code/R-code interdits auto-confirm
- Traçabilité : dp_trackare, dp_crh_only, dp_final sur DossierMedical
- quality_flags dict (merge sans écraser) + alertes_codage (append)

Modèles config.py :
- DPCandidate, DPSelection (NUKE-3)
- get_dp_ranker_llm_enabled(), check_adversarial_model_config()
- Champs DossierMedical : dp_trackare, dp_crh_only, dp_final, quality_flags

Intégration :
- main.py : appel finalize_dp() après vetos/GHM (individuel + fusionné)
- benchmark : finalizer dans _rebuild_and_select(), dp_final dans output

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

scripts/benchmark_nuke3_compare.py

@@ -560,21 +560,37 @@ def _rebuild_and_select(data: dict) -> dict:
     }
 
     selection = select_dp(dossier, synthese, config={"llm_enabled": False})
+    dossier.dp_selection = selection
+
+    # Finalizer DP (arbitrage Trackare vs CRH, traçabilité)
+    try:
+        from src.medical.dp_finalizer import finalize_dp
+        finalize_dp(dossier)
+    except Exception:
+        pass
+
+    # Utiliser dp_final si disponible, sinon dp_selection
+    final = dossier.dp_final or selection
 
     # Convertir en dict compatible analyze_dp_selection
-    cands = [c.model_dump() for c in selection.candidates]
-    return {
+    cands = [c.model_dump() for c in final.candidates]
+    result = {
         "dp_selection": {
-            "verdict": selection.verdict,
-            "confidence": selection.confidence,
-            "chosen_code": selection.chosen_code,
-            "chosen_term": selection.chosen_term,
+            "verdict": final.verdict,
+            "confidence": final.confidence,
+            "chosen_code": final.chosen_code,
+            "chosen_term": final.chosen_term,
             "candidates": cands,
-            "evidence": selection.evidence,
-            "reason": selection.reason,
-            "debug_scores": selection.debug_scores,
+            "evidence": final.evidence,
+            "reason": final.reason,
+            "debug_scores": final.debug_scores,
         }
     }
+    if dossier.dp_final:
+        result["dp_final"] = dossier.dp_final.model_dump(exclude_none=True)
+    if dossier.quality_flags:
+        result["quality_flags"] = dossier.quality_flags
+    return result
 
 
 def _run_debug_reports(