feat: réduction FP + gazetteers adresses FINESS + batch parallèle + corrections multi-axes

- Token min length relevé de 2-3 → 4 chars (élimine FP EPO, IRC, SIB...)
- Stop-words enrichis : acronymes médicaux 3 lettres, termes pharma, soins infirmiers
- BDPM stop-words : ~7300 noms commerciaux + DCI/substances actives
- Gazetteers adresses FINESS : 63K patterns Aho-Corasick (position-preserving normalization)
- Filtre contextuel anatomique pour FINESS établissements
- Nouvelles regex : RE_CIVILITE_COMMA_LIST, RE_EXTRACT_NOM_UTILISE, RE_EXTRACT_PRENOM,
  RE_NUM_EXAMEN_PATIENT, RE_ADRESSE_LIEU_DIT, RE_CIVILITE_INITIALE, Dr X.NOM
- URLs complètes (RE_URL) + détection multiline
- N° venue inversé (layout-aware) + EPISODE/NDA dans _CRITICAL_PII_TYPES
- HospitalFilter désactivé pour ADRESSE/TEL/VILLE/EPISODE (identifient le patient)
- Batch silver export parallélisé (multiprocessing spawn, N workers)
- Seuil sur-masquage relevé à 8%, server.py enrichi (source regex/ner)
- Blacklist villes : COURANT, PARIS ; contexte villes étendu (UHCD, spécialités)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

server.py

@@ -210,17 +210,34 @@ async def anonymize_text(
     final_text = selective_rescan(final_text, cfg=cfg)
 
     elapsed = time.time() - t0
-    audit_list = [
-        {"kind": h.kind, "original": h.original, "placeholder": h.placeholder, "page": h.page}
-        for h in anon.audit
-        if h.page != -1  # exclure les propagations globales
-    ]
+
+    # Inclure tous les hits (regex page≥0 + NER page=-1) avec source
+    ner_prefixes = ("NER_", "EDS_")
+    audit_list = []
+    ner_count = 0
+    regex_count = 0
+    for h in anon.audit:
+        is_ner = h.kind.startswith(ner_prefixes) or h.page == -1
+        entry = {
+            "kind": h.kind,
+            "original": h.original,
+            "placeholder": h.placeholder,
+            "page": h.page,
+            "source": "ner" if is_ner else "regex",
+        }
+        audit_list.append(entry)
+        if is_ner:
+            ner_count += 1
+        else:
+            regex_count += 1
 
     return {
         "text_anonymized": final_text,
         "audit": audit_list,
         "stats": {
             "pii_detected": len(audit_list),
+            "regex_count": regex_count,
+            "ner_count": ner_count,
             "elapsed_seconds": round(elapsed, 3),
             "ner_active": use_ner and _eds_manager is not None,
         },