Externalize dictionaries and add anonymization review corpus

anonymizer_core_refactored_onnx.py

@@ -44,6 +44,12 @@ try:
 except Exception:
     yaml = None
 
+from config_defaults import (
+    RUNTIME_DICTIONARIES_CONFIG_PATH,
+    load_effective_dictionaries_dict,
+    load_default_dictionaries_dict,
+)
+
 try:
     from doctr.models import ocr_predictor as _doctr_ocr_predictor
     _DOCTR_AVAILABLE = True
@@ -115,6 +121,29 @@ def _load_bdpm_medication_names() -> set:
         return set()
 
 
+def _load_wordlist_file(
+    path: Path,
+    *,
+    transform=lambda s: s,
+    label: str,
+    min_len: int = 1,
+) -> set:
+    """Charge un fichier texte, un mot par ligne."""
+    result: set = set()
+    if not path.exists():
+        log.warning("%s introuvable : %s", label, path)
+        return result
+    try:
+        for line in path.read_text(encoding="utf-8").splitlines():
+            word = line.strip()
+            if word and not word.startswith("#") and len(word) >= min_len:
+                result.add(transform(word))
+        log.info("%s chargé : %d entrées depuis %s", label, len(result), path.name)
+    except Exception as exc:
+        log.error("%s : erreur de lecture %s — %s", label, path, exc)
+    return result
+
+
 # ----------------- Gazetteers INSEE (prénoms + communes + noms de famille) -----------------
 # Prénoms et noms de famille sont utilisés sous deux formes :
 # - _INSEE_PRENOMS (lowercase) : check rapide "w.lower() in _INSEE_PRENOMS"
@@ -199,62 +228,24 @@ _FINESS_ADDR_AC = None             # Automate Aho-Corasick pour adresses (noms d
 _VILLE_AC = None                   # Automate Aho-Corasick pour villes (INSEE + FINESS)
 
 # Communes trop ambiguës (homonymes de mots courants, trop courts, etc.)
-_VILLE_BLACKLIST = {
-    # Directions / mots géographiques génériques
-    "SAINT", "NORD", "SUD", "EST", "OUEST",
-    "CENTRE", "SERVICE", "BOURG",
-    # Communes homonymes de mots courants français
-    "ORANGE", "TOURS", "NICE", "SENS", "VITRE",
-    "ROMANS", "MENTON", "SALON", "VIENNE",
-    "BREST",  # trop court et ambigu
-    "HYERES",  # proche de termes médicaux
-    "AGEN", "AUCH", "ALBI",
-    "BLOIS", "LAON", "LENS",
-    "GIEN", "GRAY",
-    "AIRE", "LURE", "SETE", "DOLE",
-    "VIRE", "LUNEL", "MURET", "MORET",
-    "COEUR", "FOIX", "GIVET",
-    "EVIAN", "MAURE", "MENDE",
-    "JOUE", "MEAUX", "REDON",
-    "CREIL", "CERGY",
-    # Communes de 4-5 lettres homonymes de mots très courants
-    "VERS", "MONT", "MARS", "PORT", "PONT", "FORT",
-    "BOIS", "ISLE", "LACS", "MURS", "OUST", "PREY",
-    "VAUX", "VERT", "FAUX", "REZE",
-    "BILLE", "PLACE", "VILLE", "COURS", "GRAND",
-    "ROUGE", "RICHE", "NUITS", "SORE", "SARE",
-    "TRANS", "RANS", "MARSA",
-    # Mots courants français (6+ lettres) aussi communes
-    "CHARGE", "SIGNES", "BARRES", "FOSSES", "GARDES",
-    "MARCHE", "LIGNES", "MOULIN", "PIERRE", "CHAISE",
-    "SOURCE", "VALLEE", "MAISON", "BEAUNE", "CORPS",
-    "PUITS", "CROIX", "LIGNE", "QUATRE", "PRISON",
-    # Prénoms très courants (aussi communes)
-    "MARIE", "PIERRE", "JEAN", "PAUL", "ANNE",
-    # Expressions composées ambiguës (aussi communes INSEE)
-    "LONG", "RECY", "PLAN", "MARCHE", "SALLE",
-    "CONTRE", "MERE", "ONDRES", "VEBRE",
-    # Mots structurels / médicaux
-    "PARIS",  # omniprésent, source de faux positifs
-    "FRANCE", "EUROPE",
-    # Termes ambigus (aussi communes INSEE) - trackare/DPI
-    "COURANT",  # "Médecin courant" ≠ ville
-    # Parties du corps homonymes de communes (FP "prurit invalidant (COU, décolleté)")
-    "COU", "DOS", "SEIN", "BRAS",
+_VILLE_BLACKLIST_FALLBACK = {
+    "PARIS",
+    "FRANCE",
+    "EUROPE",
+    "COURANT",
+    "COU",
+    "DOS",
+    "SEIN",
+    "BRAS",
 }
-# Enrichissement depuis fichier externe (modifiable sans toucher au code)
-_villes_bl_file = Path(__file__).parent / "data" / "villes_blacklist.txt"
-if _villes_bl_file.exists():
-    try:
-        for _line in _villes_bl_file.read_text(encoding="utf-8").splitlines():
-            _w = _line.strip()
-            if _w and not _w.startswith("#"):
-                _VILLE_BLACKLIST.add(_w)
-        log.info("Villes blacklist chargées : %d entrées", len(_VILLE_BLACKLIST))
-    except Exception as _exc:
-        log.error("Villes blacklist : erreur de lecture %s — %s", _villes_bl_file, _exc)
-else:
-    log.warning("Villes blacklist : fichier introuvable %s — défauts intégrés utilisés", _villes_bl_file)
+_VILLE_BLACKLIST = _load_wordlist_file(
+    Path(__file__).parent / "data" / "villes_blacklist.txt",
+    transform=str.upper,
+    label="Villes blacklist",
+)
+if not _VILLE_BLACKLIST:
+    _VILLE_BLACKLIST = set(_VILLE_BLACKLIST_FALLBACK)
+_BASE_VILLE_BLACKLIST = set(_VILLE_BLACKLIST)
 
 try:
     import ahocorasick as _ahocorasick
@@ -331,7 +322,7 @@ def load_medical_whitelists():
     global _MEDICAL_STRUCTURAL_TERMS, _MEDICATION_WHITELIST
     
     # 1. Charger les termes médicaux structurels
-    config_path = Path("config/medical_terms_whitelist.yml")
+    config_path = Path(__file__).parent / "config" / "medical_terms_whitelist.yml"
     if config_path.exists() and yaml:
         try:
             with open(config_path, 'r', encoding='utf-8') as f:
@@ -345,48 +336,20 @@ def load_medical_whitelists():
     # 2. Charger la whitelist des médicaments (edsnlp + BDPM + manuels)
     _MEDICATION_WHITELIST = _load_edsnlp_drug_names()
     _MEDICATION_WHITELIST.update(_load_bdpm_medication_names())
-    # Ajouter médicaments manquants
-    additional_meds = {
-        "idacio", "salazopyrine", "infliximab", "apranax",
-        "ketoprofene", "prevenar", "pneumovax", "bétadine"
-    }
-    _MEDICATION_WHITELIST.update(additional_meds)
+    _MEDICATION_WHITELIST.update(
+        _load_wordlist_file(
+            Path(__file__).parent / "data" / "bdpm" / "medication_whitelist_manual.txt",
+            transform=str.lower,
+            label="Whitelist médicaments manuelle",
+            min_len=3,
+        )
+    )
     log.info(f"Whitelist médicaments chargée: {len(_MEDICATION_WHITELIST)} médicaments (edsnlp+BDPM)")
 
 # Charger les whitelists au démarrage du module
 load_medical_whitelists()
 
 
-# ----------------- Defaults & Config -----------------
-DEFAULTS_CFG = {
-    "version": 1,
-    "encoding": "utf-8",
-    "normalization": "NFKC",
-    "whitelist": {
-        "sections_titres": ["DIM", "GHM", "GHS", "RUM", "COMPTE", "RENDU", "DIAGNOSTIC"],
-        "noms_maj_excepts": ["Médecin DIM", "Praticien conseil"],
-        "org_gpe_keep": False,
-    },
-    "blacklist": {
-        "force_mask_terms": [],
-        "force_mask_regex": [],
-    },
-    "kv_labels_preserve": ["FINESS", "IPP", "N° OGC", "Etablissement"],
-    "regex_overrides": [
-        {
-            "name": "OGC_court",
-            "pattern": r"\b(?:N°\s*)?OGC\s*[:\-]?\s*([A-Za-z0-9\-]{1,3})\b",
-            "placeholder": "[OGC]",
-            "flags": ["IGNORECASE"],
-        }
-    ],
-    "flags": {
-        "case_insensitive": True,
-        "unicode_word_boundaries": True,
-        "regex_engine": "python",
-    },
-}
-
 PLACEHOLDERS = {
     "EMAIL": "[EMAIL]",
     "TEL": "[TEL]",
@@ -445,408 +408,49 @@ def validate_nir(nir_raw: str) -> bool:
         return False
     return key_int == (97 - (body_int % 97))
 
-# Mots médicaux/techniques/courants qui ne sont pas des noms de personnes
-_MEDICAL_STOP_WORDS_SET = {
-    # Mots français courants (déterminants, prépositions, adverbes, etc.)
-    "pas", "mon", "bien", "ancien", "ancienne", "bon", "bonne", "tout", "tous",
-    "mais", "donc", "car", "que", "qui", "avec", "dans", "pour", "sur", "par",
-    "les", "des", "une", "est", "son", "ses", "nos", "aux", "cette", "ces",
-    "cher", "chez", "entre", "sans", "sous", "vers", "selon", "après", "avant",
-    "puis", "aussi", "très", "plus", "moins", "peu", "non", "oui", "quelques",
-    "mise", "début", "fin", "suite", "fait", "lieu", "cas", "jour", "jours",
-    "semaine", "semaines", "mois", "temps", "place", "nouvelle", "nouveau",
-    "franche", "légère", "quelque", "depuis", "comme", "encore", "votre",
-    "date", "note", "notes", "nom", "heure", "matin", "soir", "midi",
-    "signé", "réalisé", "courrier", "cabinet", "rue",
-    # Verbes / participes courants
-    "remontée", "associée", "réalisée", "débuté", "prolongé", "prolongée",
-    "prescrit", "prescrite", "présente", "présent", "absente", "absent",
-    "reprise", "introduction", "arrêt", "relais",
-    # Titres / rôles hospitaliers
-    "chef", "assistant", "assistante", "praticien", "praticienne",
-    "docteur", "professeur", "hospitalier", "hospitalière", "hospitaliers",
-    "spécialiste", "contractuel", "contractuelle", "titulaire",
-    "confrère", "consoeur", "coordonnateur", "coordonnatrice",
-    "médecin", "médical", "infirmier", "infirmière",
-    "praticiens", "patient", "patiente",
-    # Structure hospitalière
-    "service", "pôle", "clinique", "consultation", "secrétariat",
-    "hôpital", "hôpitaux", "centre", "établissement", "polyclinique",
-    # Villes / géographie (pas des noms de personnes)
-    "bordeaux", "bayonne", "paris", "lyon", "lille", "marseille",
-    "toulouse", "nantes", "montpellier", "pessac", "biarritz", "soustons",
-    "basque", "basques", "sud", "côte",
-    # Médicaments génériques et spécialités (DCI + noms commerciaux)
-    "colchicine", "aspirine", "cortancyl", "bisoprolol", "entresto",
-    "methotrexate", "eplerenone", "speciafoldine", "prednisone",
-    "corticoïdes", "cortisone",
-    "paracetamol", "metformine", "solupred", "novorapid", "abasaglar",
-    "lovenox", "methylprednisolone", "potassium", "humalog", "furosemide",
-    "insuline", "trulicity", "forxiga", "atorvastatine", "amlodipine",
-    "ondansetron", "eliquis", "nebivolol", "gaviscon", "loxen",
-    "morphine", "oxycodone", "kardegic", "tercian", "zopiclone",
-    "seresta", "tramadol", "alprazolam", "forlax", "levothyrox",
-    "bromazepam", "gliclazide", "zymad", "pravastatine", "spiriva",
-    "quetiapine", "sertraline", "crestor", "lercanidipine", "amoxicilline",
-    "opocalcium", "ferinject", "candesartan", "ceftriaxone", "calcidose",
-    "laroxyl", "brintellix", "ketoprofene", "adrenaline", "exacyl",
-    "terbutaline", "ipratropium", "actiskenan", "vialebex", "oxynormoro",
-    "lansoprazole", "perindopril", "sodium", "velmetia",
-    "doliprane", "dafalgan", "efferalgan", "spasfon", "vogalene",
-    "augmentin", "inexium", "omeprazole", "pantoprazole", "esomeprazole",
-    "ramipril", "lisinopril", "enalapril", "losartan", "valsartan",
-    "irbesartan", "olmesartan", "telmisartan", "hydrochlorothiazide",
-    "spironolactone", "furosemide", "lasilix", "aldactone",
-    "tahor", "crestor", "rosuvastatine", "simvastatine", "fluvastatine",
-    "xarelto", "pradaxa", "apixaban", "rivaroxaban", "dabigatran",
-    "plavix", "clopidogrel", "ticagrelor", "brilique",
-    "ventoline", "seretide", "symbicort", "salmeterol", "fluticasone",
-    "salbutamol", "tiotropium", "budesonide", "beclometasone",
-    "oxycodone", "oxynorm", "skenan", "actiskenan", "fentanyl",
-    "nubain", "nalbuphine", "nefopam", "acupan", "profenid",
-    "ibuprofene", "diclofenac", "naproxene", "celecoxib",
-    "gabapentine", "pregabaline", "lyrica", "neurontin",
-    "amitriptyline", "duloxetine", "venlafaxine", "fluoxetine",
-    "paroxetine", "escitalopram", "citalopram", "mirtazapine",
-    "olanzapine", "risperidone", "aripiprazole", "haloperidol",
-    "loxapine", "cyamemazine", "diazepam", "oxazepam", "lorazepam",
-    "clonazepam", "midazolam", "hydroxyzine", "atarax", "melatonine",
-    "stilnox", "zolpidem", "imovane",
-    "levothyroxine", "metformine", "glimepiride", "sitagliptine",
-    "januvia", "jardiance", "empagliflozine", "dapagliflozine",
-    "ozempic", "semaglutide", "dulaglutide", "liraglutide", "victoza",
-    "heparine", "enoxaparine", "tinzaparine", "innohep",
-    "warfarine", "coumadine", "fluindione", "previscan",
-    "ciprofloxacine", "levofloxacine", "ofloxacine", "metronidazole",
-    "vancomycine", "gentamicine", "tazocilline", "piperacilline",
-    "meropenem", "imipenem", "clindamycine", "doxycycline",
-    "azithromycine", "clarithromycine", "cotrimoxazole", "bactrim",
-    "polyionique", "propranolol", "apidra", "solostar",
-    # Noms et suffixes laboratoires pharmaceutiques
-    "arw", "myl", "myp", "arg", "teva", "bga", "agt",
-    "mylan", "biogaran", "arrow", "sandoz", "zentiva", "cristers",
-    "accord", "viatris", "ranbaxy", "ratiopharm", "almus", "qualimed",
-    "evolugen", "alter", "zydus", "medisol", "substipharm",
-    "sdz", "bgr", "egt", "rnb",
-    # Formes galéniques / voies d'administration
-    "cpr", "sachet", "orale", "oral", "sol", "buv", "stylo", "flexpen",
-    "flestouch", "kwikpen", "inj", "susp", "gelule", "comprime",
-    "unidose", "perf", "inh", "seringue", "aerosol", "sach", "pdr",
-    "orodisp", "capsule", "patch", "suppositoire", "gouttes",
-    # Termes de prescription / pharmacie
-    "prescription", "prescriptions", "dose", "fréquence", "statut",
-    "technique", "capteur", "bandelettes", "glycemiques", "glycemique",
-    "lancettes", "aiguilles", "fines", "micro", "pompe", "réserve",
-    "glycemie", "capillaire", "hgt",
-    # Termes médicaux / cliniques
-    "myocardite", "myosite", "corticothérapie", "biopsie", "pathologie",
-    "dysimmunitaire", "récidive", "récidivante", "traitement", "diagnostic",
-    "antécédents", "examen", "bilan", "résultats", "analyse",
-    "interne", "externe", "médecine", "chirurgie", "rhumatologie",
-    "dermatologie", "immunologie", "cardiologie", "pneumologie",
-    "neurologie", "gynécologie", "radiologie", "sénologie",
-    "douleur", "douleurs", "douloureux", "musculaire", "musculaires",
-    "thoracique", "thoraciques", "membres", "supérieurs", "inférieurs",
-    "normale", "normaux", "habituelle", "habituelles",
-    "synthèse", "hospitalisation", "syndrome", "vaccination", "ophtalmo",
-    "pelvien", "diabétique", "sommeil", "régime", "diet",
-    "desinfection", "environnement", "identification", "bracelet",
-    "toilettes", "accompagner", "installer", "transfusion",
-    "signes", "vitaux", "alimentaire", "avis", "zone",
-    "calcémie",
-    # Abréviations médicales
-    "irm", "ett", "ecg", "mtx", "fevg", "bdc", "crp", "sfu", "hdj",
-    "bnp", "asat", "alat", "cpk", "ctc", "hba", "hba1c",
-    "saos", "tsh", "inr", "vgm", "pnn", "plq", "hb",
-    "poc", "bax", "act", "bic", "cfx", "acc", "ado", "acf", "vfo",
-    "qvl", "cci", "pse", "pca", "chl", "crt", "bbm", "pds", "ren",
-    "vit", "zen",
-    "scanner", "radio", "écho", "échographie",
-    # Spécialités médicales (éviter faux positifs NOM)
-    "hépato-gastro-entérologue", "gastro-entérologue", "gastro-entérologie",
-    "proctologue", "oncologue", "anesthésiste", "pneumologue", "gérontologue",
-    "cardiologue", "néphrologue", "urologue", "gériatre",
-    "hépatologue", "endocrinologue", "stomatologue",
-    # Termes médicaux / titres fréquemment détectés comme NOM par le NER
-    "supplémentation", "supplementation", "endocrinologie", "monsieur", "madame",
-    "suivi", "sortie", "emog", "ophtalmo",
-    # Médicaments détectés comme NOM/PRENOM par EDS-Pseudo
-    "eliquis", "trulicity", "saos", "wind", "taxotere", "eupantol", "ezetimibe",
-    "lansoyl", "xatral", "xenetix", "trimbow", "buspirone", "cetirizine",
-    "depakote", "versatis", "durogesic", "montelukast", "metformine", "viatris",
-    "rosuvastatine", "gliclazide", "amlodipine", "perindopril", "nebivolol",
-    "pravastatine", "bisoprolol", "amoxicilline", "kardegic", "lovenox",
-    # Termes médicaux / soins / actes détectés comme NOM
-    "partielle", "cutanee", "cutané", "cutanée", "osseuse", "diabetique",
-    "diabétique", "transdermique", "transderm", "diarrhees", "diarrhées",
-    "ionogramme", "scintigraphie", "thoraco", "thorax", "négative", "negative",
-    "diététicienne", "pressurise", "pressuriser", "inhalee", "inhalée", "inhal",
-    # Mots courants français détectés comme NOM dans les trackare
-    "toilette", "repas", "poche", "installation", "education", "éducation",
-    "refection", "réfection", "complete", "complète", "regime", "régime",
-    "normal", "traité", "traite", "arrêté", "arrete", "volume",
-    "commentaires", "france", "covid", "framboise", "epoux", "époux",
-    # Abréviations médicales courtes (3-4 chars) détectées comme NOM
-    "ide", "ipp", "pcr", "tap", "gel", "ahl", "ssr", "hds", "tca", "etp",
-    "mcg", "sdz", "iao", "ser", "orod", "clav", "disp", "cart", "atcd", "mdrd",
-    "amox", "endoc", "microg", "item", "pyélo", "néphro",
-    # En-têtes de colonnes / mots structurels trackare
-    "observations", "observation", "commentaires", "commentaire",
-    "surveillance", "température", "temperature", "glycémie", "glycemie",
-    "diurèse", "diurese", "balance", "pouls", "systolique", "diastolique",
-    "saturation", "fréquence", "frequence", "respiratoire", "douleur",
-    "alertes", "alerte", "antécédents", "antecedents", "habitus",
-    "allergies", "prescriptions", "prescription", "administration",
-    "catégorie", "categorie", "expiration", "message",
-    "destination", "diagnostique", "diagnostiques",
-    "date", "note", "nom", "heure", "type", "code", "etat",
-    "comprime", "comprimé", "gelule", "gélule", "solution", "injectable",
-    # Médicaments supplémentaires détectés dans les trackare
-    "depakote", "versatis", "humalog", "forxiga", "durogesic",
-    "montelukast", "rosuvastatine",
-    # Abréviations pharma courtes
-    "cpr", "sol", "bic", "agt", "poche", "inhal",
-    # Termes chirurgicaux/cliniques FP
-    "cure", "endocrine", "operatoire", "opératoire", "realisee", "réalisée",
-    "gauche", "droit", "droite", "face", "profil",
-    # Faux positifs EDS supplémentaires
-    "psy", "inhales", "inhalés", "kwikpen", "lansoprazole", "tiorfan", "smecta",
-    "axa", "ttt", "anionique", "abdomino", "cod", "omi", "urg", "med",
-    "10mg", "20mg", "40mg", "100mg", "300ui", "500ml", "innohep", "coaprovel",
-    "actiskenan", "simvastatine", "forlax",
-    # Mots temporels / contextuels détectés comme EDS_HOPITAL
-    "semaine", "jour", "matin", "soir", "nuit", "midi",
-    # Mots clés de contexte document
-    "compétences", "maladies", "inflammatoires", "systémiques", "rares",
-    "fret", "fax", "contexte", "résultat", "resultat", "résultats", "resultats",
-    "haute", "maison", "aide", "rpps", "poste", "fonct",
-    "sante", "santé", "etxe", "ttipi", "gastro", "concha",
-    "endoscopie", "endoscopique", "fibroscopie",
-    "indication", "conclusion", "technique", "anesthésie",
-    "digestif", "digestive", "digestives", "nutritive",
-    # Abréviations soins trackare détectées comme NOM (batch 20 OGC)
-    "soins", "lit", "jeun", "lever", "pose", "surv", "ggt", "vvp",
-    # Verbes d'instructions soins (aussi des patronymes INSEE → FP)
-    "coucher", "manger", "marcher", "sortir",
-    "verif", "crop", "evs", "maco", "pan", "cet", "trou", "nit", "nfs",
-    # Mots narratifs CRH capturés par fusion sidebar 2-colonnes
-    "evolution", "évolution", "explorations", "fermeture", "allergie", "allergies",
-    "lotissement", "cholangiographie", "cholecystectomie", "cholécystectomie",
-    "paracetamol", "paracétamol", "unité", "unite",
-    # FP résiduels batch 10 OGC (termes médicaux/instructions soins)
-    "glyc", "glycosurie", "vider", "forte",
-    # FP audit batch 59 OGC (mots courants/médicaux flagués comme NOM)
-    "oncologie", "confrères", "confrere", "doubles", "chers", "motif",
-    "responsable", "autre", "autres", "autonome", "autonomes",
-    "préparations", "preparations", "prévenir", "prevenir",
-    "acétylsalicylique", "acetylsalicylique", "angio",
-    "desc", "diu", "barreau",
-    "haitz", "alde",
-    # FP audit OGC 21 — termes médicaux/courants flagués NOM_GLOBAL
-    "alimentation", "augmentation", "amelioration", "amélioration",
-    "biliaire", "biliaires", "bili", "voies", "voie",
-    "apyrexie", "apyréxie", "apyrétique", "apyretique",
-    "clavulanique", "mecillinam", "sulfamides", "sulfamide",
-    "tazobactam", "temocilline", "ecoflac", "furanes", "furane",
-    "exilar", "lipruzet", "mopral",
-    "sensible", "sensibles", "dossier", "dossiers",
-    "entero", "entéro", "medecine", "bio",
-    "aviation", "contention", "isolement",
-    "elimination", "élimination", "infectieux",
-    "hémodynamique", "hemodynamique", "pancréatite", "pancreatite",
-    "cholecystite", "cholécystite", "cholécystectomie", "cholecystectomie",
-    "appendicectomie", "néoplasie", "neoplasie",
-    "ovarienne", "prandial", "fébrile", "febrile",
-    "eupnéique", "eupneique", "normocarde", "normotendue",
-    "variable", "dosage", "posologie",
-    # Abréviations diététiques/soins trackare
-    "bcy", "po2", "po1", "po3", "bha", "atg", "ras", "cat", "ass",
-    # FP audit OGC 17 CRH
-    "mode", "retraitee", "retraité", "retraitée", "régression", "regression", "tel",
-    "strasbourg", "bordeaux", "toulouse", "paris", "lyon", "marseille", "bayonne", "anglet",
-    "saint-palais", "tarnos", "hendaye", "dax", "orthez", "oloron", "pau", "cambo",
-    # Spécialités/services récurrents comme FP NOM
-    "cancérologie", "cancerologie", "réanimation", "reanimation",
-    "urologie", "néphrologie", "nephrologie", "hématologie", "hematologie",
-    "gériatrie", "geriatrie", "pédiatrie", "pediatrie",
-    "ophtalmologie", "stomatologie", "allergologie",
-    "kinésithérapie", "kinesitherapie", "ergothérapie", "ergotherapie",
-    "orthopédie", "orthopedie", "traumatologie",
-    "palliatifs", "palliative", "palliatif",
-    "addictologie", "alcoologie", "tabacologie",
-    # FP soignants trackare (mots courants capturés par patterns Note d'évolution / Signé / Flacon)
-    "discussion", "echelle", "échelle", "scope", "tdm", "bouteille",
-    "evendol", "relais", "repas", "poursuite", "indication",
-    # FP pattern timestamp (termes ALL-CAPS capturés par "HH:MM NOM")
-    "eliminatin", "elimination", "élimination", "preremplie", "pré-remplie",
-    "thermie", "alim", "alimentation", "admin",
-    # Médicaments/tests labo capturés par patterns soignants
-    "biprofenid", "bi-profenid", "phosphatase", "phosphatases",
-    "ecbu", "suppo", "suppositoire", "mucite", "microlax", "normacol",
-    "ciprofloxacine", "lavement", "desinfection", "désinfection",
-    "avaler", "rachis", "lombaire", "thoraco-lombaire",
-    "cérébrale", "cerebrale", "cérébral", "cerebral", "hépatique", "hepatique",
-    "thoracique", "abdominale", "abdominal", "pelvienne", "pelvien", "médullaire",
-    # Dosages et labos pharma (FP fréquents dans prescriptions Trackare)
-    "faible", "fort", "forte",
-    "myl", "mylan", "sandoz", "teva", "arrow", "biogaran", "zentiva", "eg",
-    "arw", "pan", "mso", "bgn", "ratiopharm", "accord", "vts", "viatris",
-    "abdomino-pelvien", "abdomino", "pelvien", "thoraco-abdominal",
-    "entree", "entrée", "continu", "continue",
-    "morphine", "claforan", "skenan", "actiskenan",
-    # Fragments de noms de médicaments (pdfplumber split)
-    "sium", "pegic", "fenid", "profenid",
-    # Catégories cliniques Trackare (en-têtes de section masqués à tort)
-    "respi", "respiratoire", "nephro", "cardio", "neuro", "onco", "pulmo",
-    "hemato", "hémato", "infectieux", "thermie", "diurese", "diurèse",
-    "transit", "anemie", "anémie", "constantes", "examen",
-    "post-op", "postop", "pré-op", "preop", "chimio", "elim",
-    "toilette", "sommeil", "hypota", "hypotension", "spo2",
-    "urine", "urines", "sng",
-    "rénale", "renale", "rénal", "renal", "cardiaque",
-    # Termes structurels trackare
-    "transmissions", "transmission", "releve", "relevé",
-    "objectif", "objectifs", "evaluation", "évaluation",
-    "planification", "planifié", "planifiee",
-    # ── FP détectés automatiquement par audit_fp_detector.py ──
-    # Lot 2 : tokens basse confiance (DICT_FR seul) clairement non-noms
-    "acide", "adulte", "ambulatoire", "avenue", "bandelette", "bassin",
-    "bijoux", "bouche", "bouchon", "changes", "court", "demande", "dessert",
-    "devenir", "diffusé", "douche", "entrée", "escarre", "espace",
-    "explications", "fauteuil", "feuillet", "fixateur", "fois", "gamma",
-    "germes", "glace", "habillage", "liste", "maquillage", "matelas",
-    "mettre", "obésité", "ongles", "palais", "perlant", "pertes",
-    "pièce", "plaie", "risque", "saint", "sang", "signe", "sonde",
-    "tenue", "texte", "transaminases", "transit", "transmis", "urinal",
-    "vernis", "vessie", "vrac",
-    # Lot 2 : termes médicaux (préfixes/suffixes)
-    "anatomo-pathologique", "anemie", "anémie", "angioscanner",
-    "cétonurie", "cetonurie", "depilation", "dépilation",
-    "folique", "gastroentérologue", "gastroenterologue",
-    "microgrammes", "nalidixique", "naso-gastrique",
-    "angio-irm", "neuro", "neuro-chirurgie", "endoplasmique",
-    "cyto", "plaie-colle", "bionolyte",
-    # Lot 1 (103 tokens, confiance >= 0.5) ──
-    # Anatomie / clinique
-    "abdomen", "aortique", "bilirubine", "drain", "gastrique", "histologique",
-    "intraveineuse", "lithiasique", "macroscopie", "miction", "molles", "pelvienne",
-    "plaquette", "plaquettes", "rectale", "sanguine", "sciatique", "selle", "urinaire",
-    # Pathologies / symptômes
-    "algie", "angoisse", "antécédent", "douloureuse", "dyslipidémie",
-    "hemodialyse", "hemorragique", "hyperthermie", "hématologue",
-    # Médicaments / matériel médical
-    "ampoule", "antalgique", "antiseptique", "compresse", "flacon",
-    "oxygène", "pansement", "vitamine",
-    # Biologie / examens
-    "biochimie", "biologie", "fer",
-    # Actions / états cliniques
-    "ablation", "absence", "admission", "bloc", "changement", "cliniquement",
-    "cognitif", "couchage", "dispositif", "dynamique", "entretien", "histoire",
-    "intervention", "position", "rappel", "relation", "retour", "réalisation",
-    "résistant", "réévaluation", "situation", "temporaire", "urgence", "urgences",
-    "urgent", "validation",
-    # Mots courants / contextuels
-    "angle", "bille", "boisson", "bureau", "cases", "circuit",
-    "concubin", "confortable", "demain", "densité", "dernière",
-    "distant", "domaine", "elle", "fils", "frère", "grand", "horizon",
-    "hui", "identifiant", "minuit", "murent", "neuf", "original", "pages",
-    "personne", "premier", "quartier", "retraite", "route", "rés",
-    "trouve", "verrouillé", "villa", "étage",
-    # Termes médicaux courants faussement détectés comme NOM (Phase 2 audit mars 2026)
-    "ains", "ponction", "hanche", "burkitt", "orl", "gds", "oap", "tvp", "epp",
-    "bronchite", "accueil", "cadre", "transfert", "relecture", "examens",
-    "traitements", "traitement", "infectiologie", "cancérologie", "cancerologie",
-    "maternité", "orale", "sachet", "absence",
-    # FP audit 30 fichiers Phase 2 (mars 2026)
-    "bouffee", "bouffée", "discontinue", "respimat", "lyoc",
-    "probnp", "pro-bnp", "nt-probnp",
-    "bpco", "colle", "gsc", "masse",
-    "selle", "selles",
-    # Acronymes médicaux courts (3 lettres) souvent FP comme NOM
-    "epo", "irc", "sib", "inr", "iec", "ira", "ait", "avc",
-    "imc", "ipp", "ivo", "amp", "ivg", "img", "had", "ssr",
-    "hta", "ecg", "irm", "tep", "crp", "nfs", "bhc", "vgm",
-    "vni", "aeg", "bas", "snv", "hba", "ide", "dci",
-    # Termes pharmaceutiques FP comme NOM (audit 30 fichiers mars 2026)
-    "buvable", "buvables", "nominal", "nominaux",
-    "acide", "principale", "principal", "principaux",
-    "hyaluronique", "valproique", "valproïque", "tranexamique", "tranéxamique",
-    "clavulanique", "nalidixique",
-    "grancher",  # Centre de réadaptation (nom d'établissement homonyme)
-    "experf",  # Prestataire HAD (nom commercial homonyme)
-    # Noms de services hospitaliers (FP comme [NOM])
-    "ortho", "mobile", "polyvalente", "polyvalent",
-    "geriatrie", "gériatrie", "ambulatoire", "provisoire",
-    "intraveineuse", "intraveineux", "sous-cutanee", "sous-cutané",
-    # Noms de services hospitaliers (aussi patronymes INSEE → FP récurrents)
-    "viscerale", "viscérale", "vasculaire", "vasculaires",
-    "conventionnelle", "conventionnel",
-    "polyvalente", "polyvalent",
-    "infectieuse", "infectieuses",
-    # Termes soins infirmiers / activités de la vie quotidienne (FP trackare doc 216)
-    "aide", "partielle", "partiel", "complete", "complète", "complet",
-    "contention", "lavabo", "blader", "scan", "post", "lunettes",
-    "deshabillage", "déshabillage", "habillage",
-    "surveillance", "surv", "refection", "réfection",
-    "miction", "toilette", "douche", "changes",
-    "installation", "transfert", "mobilisation",
-    "alimentation", "hydratation", "collation",
-    "stimulation", "prevention", "prévention",
-    # Termes pharmaceutiques/matériel médical FP (retour relecteur 2026-03-16)
-    "chlorure",
-    # Dispositifs médicaux (FP "OXYGENE LUNETTES" → [NOM])
-    "canule", "canules", "masque", "sonde", "sondes",
-    # Termes chirurgicaux FP comme [NOM] (retour relecteur 2026-03-17)
-    "totale", "total", "partielle", "partiel",
-    "prothese", "prothèse", "protheses", "prothèses", "unicompartimentale",
-    # Antiseptiques / produits de soins (FP trackare prescriptions)
-    "betascrub", "hibiscrub", "betadine", "biseptine", "chlorhexidine",
-    # Nutrition entérale / compléments
-    "fresubin", "nutrison", "sondalis", "isosource", "novasource",
-    # Termes médicaux FP dans bactério / texte libre
-    "nombreuses", "nombreux", "plusieurs", "quelques",
-    "internationale", "international",
-    "resorbable", "résorbable", "resorbables", "résorbables",
-    "alfa", "capsule", "capsules",
+# Mots médicaux/techniques/courants qui ne sont pas des noms de personnes.
+# Source de vérité externalisée dans data/stopwords_manuels.txt + BDPM/edsnlp.
+_MEDICAL_STOP_WORDS_FALLBACK = {
+    "date",
+    "note",
+    "heure",
+    "type",
+    "traitement",
+    "traitements",
+    "soins",
+    "surveillance",
+    "consultation",
+    "hospitalisation",
 }
-# Enrichissement automatique avec les ~4000 noms de médicaments d'edsnlp
-_MEDICAL_STOP_WORDS_SET.update(_load_edsnlp_drug_names())
-
-# Enrichissement depuis fichier externe (modifiable sans toucher au code)
-_stopwords_file = Path(__file__).parent / "data" / "stopwords_manuels.txt"
-if _stopwords_file.exists():
-    try:
-        _sw_count = 0
-        for _line in _stopwords_file.read_text(encoding="utf-8").splitlines():
-            _w = _line.strip()
-            if _w and not _w.startswith("#"):
-                _MEDICAL_STOP_WORDS_SET.add(_w)
-                _sw_count += 1
-        log.info("Stop-words manuels chargés : %d mots depuis %s", _sw_count, _stopwords_file.name)
-    except Exception as _exc:
-        log.error("Stop-words manuels : erreur de lecture %s — %s", _stopwords_file, _exc)
-else:
-    log.warning("Stop-words manuels : fichier introuvable %s — qualité dégradée", _stopwords_file)
-
-# Enrichissement BDPM : ~7300 noms commerciaux + DCI/substances actives
-_bdpm_path = Path(__file__).parent / "data" / "bdpm" / "medicaments_stopwords.txt"
-if _bdpm_path.exists():
-    try:
-        _bdpm_count = 0
-        for _line in _bdpm_path.read_text(encoding="utf-8").splitlines():
-            _w = _line.strip()
-            if _w and not _w.startswith("#"):
-                _MEDICAL_STOP_WORDS_SET.add(_w)
-                _bdpm_count += 1
-        log.info("BDPM stop-words chargés : %d mots", _bdpm_count)
-    except Exception as _exc:
-        log.error("BDPM stop-words : erreur de lecture %s — %s", _bdpm_path, _exc)
-else:
-    log.warning("BDPM stop-words : fichier introuvable %s — qualité dégradée", _bdpm_path)
-
-_MEDICAL_STOP_WORDS = (
-    r"(?:" + "|".join(re.escape(w) for w in _MEDICAL_STOP_WORDS_SET) + r")"
+_MEDICAL_STOP_WORDS_SET = _load_wordlist_file(
+    Path(__file__).parent / "data" / "stopwords_manuels.txt",
+    transform=str.lower,
+    label="Stop-words manuels",
 )
+_MEDICAL_STOP_WORDS_SET.update(_load_edsnlp_drug_names())
+_MEDICAL_STOP_WORDS_SET.update(
+    _load_wordlist_file(
+        Path(__file__).parent / "data" / "bdpm" / "medicaments_stopwords.txt",
+        transform=str.lower,
+        label="BDPM stop-words",
+    )
+)
+if not _MEDICAL_STOP_WORDS_SET:
+    _MEDICAL_STOP_WORDS_SET = set(_MEDICAL_STOP_WORDS_FALLBACK)
+_BASE_MEDICAL_STOP_WORDS_SET = set(_MEDICAL_STOP_WORDS_SET)
+
+
+def _refresh_medical_stopwords_pattern() -> None:
+    global _MEDICAL_STOP_WORDS
+    if not _MEDICAL_STOP_WORDS_SET:
+        _MEDICAL_STOP_WORDS = r"(?!)"
+        return
+    _MEDICAL_STOP_WORDS = (
+        r"(?:" + "|".join(re.escape(w) for w in sorted(_MEDICAL_STOP_WORDS_SET)) + r")"
+    )
+
+
+_refresh_medical_stopwords_pattern()
 # Un token de nom : commence par majuscule, lettres/tirets/apostrophes (PAS d'espace ni de point)
 _PERSON_TOKEN = r"[A-ZÉÈÀÙÂÊÎÔÛÄËÏÖÜÇ][A-ZÉÈÀÙÂÊÎÔÛÄËÏÖÜÇa-zéèàùâêîôûäëïöüç\-\']+"
 RE_PERSON_CONTEXT = re.compile(
@@ -985,7 +589,17 @@ RE_CIVILITE_INITIALE = re.compile(
 
 # --- N° examen / N° patient imagerie (radiologie) ---
 RE_NUM_EXAMEN_PATIENT = re.compile(
-    r"N[°o]?\s*(?:d['''\u2019]\s*)?(?:examen|patient|accession|passage)\s*[:\-]?\s*([A-Za-z]{0,4}\d{5,12})",
+    r"N[°o]?\s*(?:d['''\u2019]\s*)?(?:examen|patient(?:\s+imagerie)?|accession|passage)\s*[:\-]?\s*"
+    r"((?=[A-Za-z0-9\-]{6,20}\b)(?=[A-Za-z0-9\-]*\d)[A-Za-z0-9\-]+)",
+    re.IGNORECASE,
+)
+# --- N° bare d'entête labo / imagerie ---
+# Exemple:
+#   N° 23L35781
+#   Prélevé le 26/07/2023     Enregistré le 27/07/2023
+RE_NUM_ACCESSION_HEADER = re.compile(
+    r"(?:^|\n)\s*N[°o]\s*[:\-]?\s*([A-Za-z0-9\-]{6,20})\s*\n"
+    r"(?:[^\n]*\n){0,2}\s*(?:Pr[ée]lev[ée]\s+le|Enregistr[ée]\s+le)",
     re.IGNORECASE,
 )
 
@@ -1177,6 +791,7 @@ _DPI_LABELS_SET: set = _load_txt_set(
 )
 if not _DPI_LABELS_SET:
     _DPI_LABELS_SET = set(_DPI_LABELS_FALLBACK)
+_BASE_DPI_LABELS_SET = set(_DPI_LABELS_SET)
 
 # Companion blacklist : termes EN MAJUSCULES qui ne sont JAMAIS des noms
 # (spécialités, labos pharma, mots courants ambigus).
@@ -1189,6 +804,7 @@ _COMPANION_BLACKLIST_SET: set = _load_txt_set(
 )
 if not _COMPANION_BLACKLIST_SET:
     _COMPANION_BLACKLIST_SET = set(_COMPANION_BLACKLIST_FALLBACK)
+_BASE_COMPANION_BLACKLIST_SET = set(_COMPANION_BLACKLIST_SET)
 
 
 _WHITELIST_FUNCTION_WORDS = {
@@ -1223,14 +839,15 @@ def _load_whitelist_phrases(phrases) -> int:
 
 
 def load_dictionaries(config_path: Optional[Path]) -> Dict[str, Any]:
-    cfg = DEFAULTS_CFG.copy()
-    if config_path and config_path.exists() and yaml is not None:
-        try:
-            user = yaml.safe_load(config_path.read_text(encoding="utf-8")) or {}
-            for k, v in user.items():
-                cfg[k] = v
-        except Exception:
-            pass
+    global _MEDICAL_STOP_WORDS_SET, _VILLE_BLACKLIST, _DPI_LABELS_SET, _COMPANION_BLACKLIST_SET
+    cfg = load_default_dictionaries_dict() if config_path is None else load_effective_dictionaries_dict(config_path)
+
+    _MEDICAL_STOP_WORDS_SET = set(_BASE_MEDICAL_STOP_WORDS_SET)
+    _VILLE_BLACKLIST = set(_BASE_VILLE_BLACKLIST)
+    _DPI_LABELS_SET = set(_BASE_DPI_LABELS_SET)
+    _COMPANION_BLACKLIST_SET = set(_BASE_COMPANION_BLACKLIST_SET)
+    _WHITELIST_NEVER_MASK_TOKENS.clear()
+    _WHITELIST_NEVER_MASK_PHRASES.clear()
 
     # Charger les stop-words et villes supplémentaires depuis le YAML
     extra_sw = cfg.get("additional_stopwords", [])
@@ -1239,6 +856,7 @@ def load_dictionaries(config_path: Optional[Path]) -> Dict[str, Any]:
             if w and str(w).strip():
                 _MEDICAL_STOP_WORDS_SET.add(str(w).strip().lower())
         log.info("Stop-words YAML supplémentaires : %d", len(extra_sw))
+    _refresh_medical_stopwords_pattern()
 
     extra_villes = cfg.get("additional_villes_blacklist", [])
     if extra_villes:
@@ -1871,8 +1489,49 @@ def _mask_critical_in_key(key: str, audit: List[PiiHit], page_idx: int) -> str:
     return key
 
 
+def _replace_captured_value(full_match: str, captured_value: str, placeholder: str) -> str:
+    start = full_match.find(captured_value)
+    if start < 0:
+        return placeholder
+    end = start + len(captured_value)
+    return full_match[:start] + placeholder + full_match[end:]
+
+
+def _mask_structured_line(line: str, audit: List[PiiHit], page_idx: int) -> str:
+    """Masque les champs structurés dont la détection dépend du libellé de la ligne."""
+
+    def _repl_code_postal(m: re.Match) -> str:
+        original = m.group(1) or m.group(2) or m.group(0)
+        audit.append(PiiHit(page_idx, "CODE_POSTAL", original, PLACEHOLDERS["CODE_POSTAL"]))
+        if m.group(1):
+            return _replace_captured_value(m.group(0), m.group(1), PLACEHOLDERS["CODE_POSTAL"])
+        return PLACEHOLDERS["CODE_POSTAL"]
+
+    def _repl_num_examen(m: re.Match) -> str:
+        audit.append(PiiHit(page_idx, "DOSSIER", m.group(1), PLACEHOLDERS["DOSSIER"]))
+        return _replace_captured_value(m.group(0), m.group(1), PLACEHOLDERS["DOSSIER"])
+
+    def _repl_dossier(m: re.Match) -> str:
+        original = m.group(1) or m.group(2) or m.group(0)
+        audit.append(PiiHit(page_idx, "DOSSIER", original, PLACEHOLDERS["DOSSIER"]))
+        return _replace_captured_value(m.group(0), original, PLACEHOLDERS["DOSSIER"])
+
+    def _repl_venue(m: re.Match) -> str:
+        audit.append(PiiHit(page_idx, "NDA", m.group(1), PLACEHOLDERS["NDA"]))
+        return _replace_captured_value(m.group(0), m.group(1), PLACEHOLDERS["NDA"])
+
+    masked = RE_CODE_POSTAL.sub(_repl_code_postal, line)
+    masked = RE_NUM_EXAMEN_PATIENT.sub(_repl_num_examen, masked)
+    masked = RE_NUMERO_DOSSIER.sub(_repl_dossier, masked)
+    masked = RE_VENUE_SEJOUR.sub(_repl_venue, masked)
+    return masked
+
+
 def _kv_value_only_mask(line: str, audit: List[PiiHit], page_idx: int, cfg: Dict[str, Any]) -> str:
     line = _mask_admin_label(line, audit, page_idx)
+    structured_line = _mask_structured_line(line, audit, page_idx)
+    if structured_line != line:
+        return structured_line
     parts = SPLITTER.split(line, maxsplit=1)
     if len(parts) == 2:
         key, value = parts
@@ -2413,6 +2072,35 @@ def _extract_document_names(full_text: str, cfg: Dict[str, Any]) -> Tuple[set, s
     for m in _RE_EMAIL_HEADER.finditer(full_text):
         _add_tokens_force_all(m.group(1), "EMAIL_HEADER", "medium")
 
+    # En-têtes patient en capitales, sans libellé explicite.
+    # Exemple:
+    #   ETCHEVERRY JEAN CLAUDE
+    # On reste conservateur: 2-4 tokens uppercase, avec au moins un prénom
+    # INSEE et un nom de famille INSEE. Les tokens proposés viennent
+    # exclusivement des dictionnaires INSEE, sans blacklist codée en dur ici.
+    _UPPER_NAME_LINE_RE = re.compile(
+        r"^[ \t]*([A-ZÉÈÀÙÂÊÎÔÛÄËÏÖÜÇ][A-ZÉÈÀÙÂÊÎÔÛÄËÏÖÜÇ\-' ]+"
+        r"[A-ZÉÈÀÙÂÊÎÔÛÄËÏÖÜÇ])[ \t]*$",
+        re.MULTILINE,
+    )
+    for m in _UPPER_NAME_LINE_RE.finditer(full_text):
+        raw_line = re.sub(r"\s+", " ", m.group(1)).strip()
+        tokens = [tok.strip(" .-'") for tok in raw_line.split() if tok.strip(" .-'")]
+        if len(tokens) < 2 or len(tokens) > 4:
+            continue
+        if any(len(tok) < 3 for tok in tokens):
+            continue
+
+        norm_tokens = [_normalize_nfkd_upper(tok) for tok in tokens]
+        has_prenom = any(tok in _INSEE_PRENOMS_SET for tok in norm_tokens)
+        has_nom = any(tok in _INSEE_NOMS_FAMILLE for tok in norm_tokens)
+        if not (has_prenom and has_nom):
+            continue
+
+        for tok, norm_tok in zip(tokens, norm_tokens):
+            if norm_tok in _INSEE_PRENOMS_SET or norm_tok in _INSEE_NOMS_FAMILLE:
+                _add_candidate(tok, "UPPER_NAME_LINE", "low", False)
+
     # Pour les noms composés avec tiret (ex: "LACLAU-LACROUTS"),
     # ajouter aussi les parties individuelles pour capturer les occurrences standalone.
     # _apply_extracted_names traite le composé en premier (plus long) puis les parties.
@@ -2582,10 +2270,10 @@ def _apply_extracted_names(text: str, names: set, audit: List[PiiHit], force_nam
 
 
 def _apply_trackare_hits_to_text(text: str, audit: List[PiiHit]) -> str:
-    """Applique les PiiHit non-NOM dans le texte (NDA footers, EPISODE, RPPS, FINESS, etc.).
+    """Applique les PiiHit non-NOM dans le texte (NDA, DOSSIER, EPISODE, RPPS, FINESS, etc.).
     Ces hits sont détectés par _extract_trackare_identity ou la phase 0c
     mais n'étaient appliqués qu'au PDF raster, pas au fichier .pseudonymise.txt."""
-    _APPLY_KINDS = {"EPISODE", "RPPS", "FINESS"}
+    _APPLY_KINDS = {"DOSSIER", "EPISODE", "FINESS", "NDA", "RPPS"}
     # Collecter les valeurs à remplacer, groupées par placeholder
     replacements: Dict[str, str] = {}  # original → placeholder
     for h in audit:
@@ -2698,7 +2386,14 @@ def anonymise_document_regex(pages_text: List[str], tables_lines: List[List[str]
     for m in _RE_IPP_MULTILINE.finditer(full_raw):
         audit.append(PiiHit(-1, "IPP", m.group(1), PLACEHOLDERS["IPP"]))
 
-    # Phase 0f : DEMANDE N° multiline (DEMANDE N°\n2300261164)
+    # Phase 0f : numéro d'accession / d'examen en en-tête de labo ou imagerie
+    # Ex:
+    #   N° 23L35781
+    #   Prélevé le 26/07/2023
+    for m in RE_NUM_ACCESSION_HEADER.finditer(full_raw):
+        audit.append(PiiHit(-1, "DOSSIER", m.group(1), PLACEHOLDERS["DOSSIER"]))
+
+    # Phase 0g : DEMANDE N° multiline (DEMANDE N°\n2300261164)
     _RE_DEMANDE_MULTILINE = re.compile(
         r"DEMANDE\s+N[°o]?\s*\n\s*(\d{8,})",
         re.IGNORECASE,
@@ -2706,14 +2401,14 @@ def anonymise_document_regex(pages_text: List[str], tables_lines: List[List[str]
     for m in _RE_DEMANDE_MULTILINE.finditer(full_raw):
         audit.append(PiiHit(-1, "DOSSIER", m.group(1), PLACEHOLDERS["DOSSIER"]))
 
-    # Phase 0g : N° venue multiline (tableaux BACTERIO : label et valeur séparés)
+    # Phase 0h : N° venue multiline (tableaux BACTERIO : label et valeur séparés)
     _RE_VENUE_MULTILINE = re.compile(
         r"N[°o]?\s*venue\s*[:\-]?\s*\n(?:[^\n]*\n){0,2}\s*(\d{6,})",
         re.IGNORECASE,
     )
     for m in _RE_VENUE_MULTILINE.finditer(full_raw):
         audit.append(PiiHit(-1, "NDA", m.group(1), PLACEHOLDERS["NDA"]))
-    # Phase 0g-bis : N° venue inversé (layout-aware réordonne : numéro AVANT label)
+    # Phase 0h-bis : N° venue inversé (layout-aware réordonne : numéro AVANT label)
     _RE_VENUE_REVERSE = re.compile(
         r"(?<!\d)(\d{7,10})(?!\d)\s*\n(?:[^\n]*\n){0,4}N[°o]?\s*venue\s*[:\-]?\s*$",
         re.IGNORECASE | re.MULTILINE,
@@ -3092,55 +2787,17 @@ def _build_finess_ac():
         return
 
     # Mots génériques qui ne doivent jamais être matchés seuls
-    _ac_generic_blacklist = {
-        # Types d'établissements
-        "clinique", "pharmacie", "hopital", "centre", "foyer",
-        "residence", "maison", "cabinet", "service", "laboratoire",
-        "institut", "association", "fondation", "mutuelle", "polyclinique",
-        "dispensaire", "hospice", "annexe", "antenne", "site",
-        # Mots français courants qui sont aussi des noms d'établissements
-        "collegiale", "collegial", "cathedral", "cathedrale",
-        "providence", "esperance", "renaissance", "liberation",
-        "republique", "fraternite", "solidarite", "independance",
-        "beauregard", "bellevue", "belvedere",
-        "promenade", "esplanade", "corniche", "prefecture",
-        "croissant", "confluence", "bienvenue",
-        "chartreuse", "commanderie", "chapelle", "basilique",
-        "departement", "departementale", "communautaire",
-        # Spécialités médicales / termes cliniques courants
-        "chirurgicale", "radiologie", "addictologie", "prevention",
-        "psychotherapique", "ambulatoire", "hospitalisation",
-        "consultation", "surveillance", "therapeutique",
-        "readaptation", "reeducation", "reanimation",
-        "specialisee", "conventionnelle", "professionnelle",
-        "informatique", "administrative", "regionale",
-        # Mots communs
-        "generation", "revolution", "assomption", "visitation",
-        "consolation", "atlantique", "manutention", "prefiguration",
-        "intervalle", "pharmaciens", "pharmacien", "transfert",
-        "comprimee", "comprimees", "injectable", "injectables",
-        "maintenant", "actuellement", "auparavant", "prochainement",
-        "rapidement", "correctement", "directement", "simplement",
-        "internationale", "international", "intercommunal", "intercommunale",
-        # Termes médicaux homonymes d'établissements FINESS (retour relecteur 2026-03-17)
-        "resistance", "radiotherapie", "chimiotherapie", "curietherapie",
-        "hormonotherapie", "immunotherapie", "kinesitherapie",
-        "ergotherapie", "orthophonie", "psychomotricite",
-        "reeducation", "readaptation", "convalescence",
-        "dependance", "autonomie", "gerontologie",
-    }
+    _ac_generic_blacklist = _load_wordlist_file(
+        data_dir / "generic_name_blacklist.txt",
+        transform=str.lower,
+        label="FINESS noms génériques blacklist",
+    )
     # Expressions multi-mots trop génériques
-    _ac_generic_phrases = {
-        "a domicile", "au domicile", "menage a domicile",
-        "du nord", "du sud", "de l est", "de l ouest",
-        "la maison", "la residence", "les jardins",
-        "le village", "le parc", "la colline",
-        "au soleil", "en france",
-        # Expressions médicales homonymes d'établissements FINESS (FP relecteur 2026-03-16)
-        "long cours", "au long cours",
-        "le bourg", "le val", "le clos", "le mas",
-        "les pins", "les chenes", "les oliviers",
-    }
+    _ac_generic_phrases = _load_wordlist_file(
+        data_dir / "generic_phrase_blacklist.txt",
+        transform=str.lower,
+        label="FINESS expressions génériques blacklist",
+    )
     # Whitelist explicite de mono-mots < 10 chars considérés comme distinctifs
     # (sinon rejetés par le filtre général). Exemple : EMBRUNS (7 chars).
     # Alimentée depuis data/finess/mono_mots_distinctifs.txt — curation manuelle.
@@ -3365,8 +3022,11 @@ def _build_finess_addr_ac():
                        "sentier", "rond-point", "traverse", "esplanade",
                        "promenade", "montee", "voie", "carrefour", "faubourg"}
         # Patterns non-adresse à exclure
-        _addr_blacklist = {"cabinet medical", "cabinet dentaire", "cabinet infirmier",
-                          "cabinet paramedical", "cabinet sage-femme"}
+        _addr_blacklist = _load_wordlist_file(
+            data_dir / "address_blacklist.txt",
+            transform=str.lower,
+            label="FINESS adresses blacklist",
+        )
         for line in addr_path.read_text(encoding="utf-8").splitlines():
             name = line.strip()
             if not name or len(name) < 10:
@@ -3804,11 +3464,19 @@ def selective_rescan(text: str, cfg: Dict[str, Any] | None = None) -> str:
     protected = RE_ADRESSE.sub(PLACEHOLDERS["ADRESSE"], protected)
     protected = RE_ADRESSE_LIEU_DIT.sub(PLACEHOLDERS["ADRESSE"], protected)
     protected = RE_BP.sub(PLACEHOLDERS["ADRESSE"], protected)
-    protected = RE_CODE_POSTAL.sub(PLACEHOLDERS["CODE_POSTAL"], protected)
+    def _rescan_code_postal(m: re.Match) -> str:
+        if m.group(1):
+            return _replace_captured_value(m.group(0), m.group(1), PLACEHOLDERS["CODE_POSTAL"])
+        return PLACEHOLDERS["CODE_POSTAL"]
+
+    protected = RE_CODE_POSTAL.sub(_rescan_code_postal, protected)
     # N° Episode
     protected = RE_EPISODE.sub(PLACEHOLDERS["EPISODE"], protected)
     # N° venue / séjour
-    protected = RE_VENUE_SEJOUR.sub(PLACEHOLDERS["NDA"], protected)
+    protected = RE_VENUE_SEJOUR.sub(
+        lambda m: _replace_captured_value(m.group(0), m.group(1), PLACEHOLDERS["NDA"]),
+        protected,
+    )
     # N° RPPS
     protected = RE_RPPS.sub(PLACEHOLDERS["RPPS"], protected)
     # FINESS par gazetteer (nombres 9 chiffres matchant un vrai numéro FINESS)
@@ -4825,7 +4493,7 @@ if __name__ == "__main__":
     ap.add_argument("--out", type=str, default="out")
     ap.add_argument("--no-vector", action="store_true")
     ap.add_argument("--raster", action="store_true")
-    ap.add_argument("--config", type=str, default=str(Path("config/dictionnaires.yml")))
+    ap.add_argument("--config", type=str, default=str(RUNTIME_DICTIONARIES_CONFIG_PATH))
     ap.add_argument("--hf", action="store_true", help="Activer NER ONNX sur narratif (nécessite ner_manager_onnx)")
     ap.add_argument("--model", type=str, default="cmarkea/distilcamembert-base-ner")
     args = ap.parse_args()