feat: configuration externalisée via .env + audit requirements

- Externalise 13 variables de config via python-dotenv (chemins PDF,
  modèles Ollama/embedding/NER, FINESS, seuils) avec défauts identiques
- Centralise EMBEDDING_MODEL dans config.py (était hardcodé en 3 endroits)
- Ajoute .env.example documenté et .env au .gitignore
- Ajoute openpyxl et pandas manquants au requirements.txt
- Ajoute data/referentiels au mkdir de run.sh

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

src/medical/rag_index.py

@@ -11,7 +11,7 @@ from typing import Optional
 
 import pdfplumber
 
-from ..config import RAG_INDEX_DIR, CIM10_PDF, GUIDE_METHODO_PDF, CCAM_PDF, CCAM_DICT_PATH, REFERENTIELS_DIR
+from ..config import RAG_INDEX_DIR, CIM10_PDF, GUIDE_METHODO_PDF, CCAM_PDF, CCAM_DICT_PATH, REFERENTIELS_DIR, EMBEDDING_MODEL
 
 logger = logging.getLogger(__name__)
 
@@ -426,8 +426,8 @@ def build_index(force: bool = False) -> None:
     # Embeddings — GPU si disponible
     import torch
     _device = "cuda" if torch.cuda.is_available() else "cpu"
-    logger.info("Chargement du modèle d'embedding dangvantuan/sentence-camembert-large (%s)...", _device)
-    model = SentenceTransformer("dangvantuan/sentence-camembert-large", device=_device)
+    logger.info("Chargement du modèle d'embedding %s (%s)...", EMBEDDING_MODEL, _device)
+    model = SentenceTransformer(EMBEDDING_MODEL, device=_device)
     model.max_seq_length = 512  # CamemBERT max position embeddings
 
     texts = [c.text[:2000] for c in all_chunks]  # Tronquer les chunks trop longs

src/medical/rag_search.py

@@ -8,6 +8,7 @@ from concurrent.futures import ThreadPoolExecutor, as_completed
 from ..config import (
     ActeCCAM, Diagnostic, DossierMedical, RAGSource,
     OLLAMA_CACHE_PATH, OLLAMA_MAX_PARALLEL, OLLAMA_MODEL,
+    EMBEDDING_MODEL,
 )
 from .cim10_dict import normalize_code, validate_code as cim10_validate
 from .cim10_extractor import BIO_NORMALS
@@ -36,12 +37,12 @@ def _get_embed_model():
         _device = "cuda" if torch.cuda.is_available() else "cpu"
         try:
             logger.info("Chargement du modèle d'embedding (%s)...", _device)
-            _embed_model = SentenceTransformer("dangvantuan/sentence-camembert-large", device=_device)
+            _embed_model = SentenceTransformer(EMBEDDING_MODEL, device=_device)
         except torch.OutOfMemoryError:
             if _device == "cuda":
                 logger.warning("CUDA OOM pour l'embedding — fallback CPU")
                 torch.cuda.empty_cache()
-                _embed_model = SentenceTransformer("dangvantuan/sentence-camembert-large", device="cpu")
+                _embed_model = SentenceTransformer(EMBEDDING_MODEL, device="cpu")
             else:
                 raise
         _embed_model.max_seq_length = 512