refactor: réorganisation référentiels, nouveaux modules extraction, nettoyage code obsolète

- Réorganisation data/referentiels/ : pdfs/, dicts/, user/ (structure unifiée)
- Fix badges "Source absente" sur page admin référentiels
- Ré-indexation COCOA 2025 (555 → 1451 chunks, couverture 94%)
- Fix VRAM OOM : embeddings forcés CPU via T2A_EMBED_CPU
- Nouveaux modules : document_router, docx_extractor, image_extractor, ocr_engine
- Module complétude (quality/completude.py + config YAML)
- Template DIM (synthèse dimensionnelle)
- Gunicorn config + systemd service t2a-viewer
- Suppression t2a_install_rag_cleanup/ (copie obsolète)
- Suppression scripts/ et scripts_t2a_v2/ (anciens benchmarks)
- Suppression 81 fichiers _doc.txt de test
- Cache Ollama : TTL configurable, corrections loader YAML
- Dashboard : améliorations templates (base, index, detail, cpam, validation)

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

src/control/cpam_response.py

@@ -8,9 +8,13 @@ Orchestrateur principal — délègue aux sous-modules :
 
 from __future__ import annotations
 
+import json
 import logging
+import os
+from datetime import datetime
+from pathlib import Path
 
-from ..config import ControleCPAM, DossierMedical, RAGSource, rule_enabled
+from ..config import ControleCPAM, DossierMedical, RAGSource, STRUCTURED_DIR, rule_enabled
 from ..medical.ollama_client import call_anthropic, call_ollama
 from ..prompts import CPAM_EXTRACTION
 
@@ -50,6 +54,70 @@ from .cpam_validation import _CIM10_CODE_RE, _validate_adversarial as _validate_
 logger = logging.getLogger(__name__)
 
 
+def _save_version(
+    dossier: DossierMedical,
+    controle: ControleCPAM,
+) -> None:
+    """Sauvegarde la version actuelle de l'argumentaire avant régénération.
+
+    Stocke dans output/structured/{dossier}/_cpam_versions/{ogc}_{timestamp}.json
+    """
+    if not controle.contre_argumentation and not controle.response_data:
+        return  # rien à versionner
+
+    # Trouver le dossier structuré (depuis source_files ou source_file)
+    dossier_dir = None
+    if not STRUCTURED_DIR.is_dir():
+        logger.debug("Versioning : STRUCTURED_DIR inexistant, skip")
+        return
+
+    structured_dirs = [d for d in STRUCTURED_DIR.iterdir() if d.is_dir()]
+
+    # Tentative 1 : matcher un source_file contre les noms de sous-dossiers
+    candidates = list(dossier.source_files or [])
+    if dossier.source_file and dossier.source_file not in candidates:
+        candidates.append(dossier.source_file)
+
+    for src in candidates:
+        src_stem = Path(src).stem.replace(" ", "_")
+        for d in structured_dirs:
+            if src_stem in d.name:
+                dossier_dir = d
+                break
+        if dossier_dir:
+            break
+
+    if not dossier_dir:
+        logger.debug("Versioning : pas de dossier structuré trouvé, skip")
+        return
+
+    versions_dir = dossier_dir / "_cpam_versions"
+    versions_dir.mkdir(exist_ok=True)
+
+    # Compter les versions existantes pour cet OGC
+    existing = sorted(versions_dir.glob(f"{controle.numero_ogc}_*.json"))
+    version_num = len(existing) + 1
+
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    filename = f"{controle.numero_ogc}_{timestamp}_v{version_num}.json"
+
+    version_data = {
+        "numero_ogc": controle.numero_ogc,
+        "version": version_num,
+        "timestamp": timestamp,
+        "contre_argumentation": controle.contre_argumentation,
+        "response_data": controle.response_data,
+        "quality_tier": controle.quality_tier,
+        "validation_dim": controle.validation_dim,
+    }
+
+    (versions_dir / filename).write_text(
+        json.dumps(version_data, ensure_ascii=False, indent=2),
+        encoding="utf-8",
+    )
+    logger.info("  Version %d sauvegardée : %s", version_num, filename)
+
+
 def _extraction_pass(
     dossier: DossierMedical,
     controle: ControleCPAM,
@@ -121,6 +189,9 @@ def generate_cpam_response(
     logger.info("CPAM : génération contre-argumentation pour OGC %d — %s",
                 controle.numero_ogc, controle.titre)
 
+    # 0. Versioning — sauvegarder la version précédente avant d'écraser
+    _save_version(dossier, controle)
+
     # 1. Passe 1 — Extraction structurée (compréhension avant argumentation)
     extraction = _extraction_pass(dossier, controle)
     degraded_pass1 = extraction is None
@@ -137,12 +208,12 @@ def generate_cpam_response(
     prompt, tag_map = _build_cpam_prompt(dossier, controle, sources, extraction)
 
     # 4. Appel LLM — Ollama (rôle cpam) > Haiku fallback
-    result = call_ollama(prompt, temperature=0.1, max_tokens=16000, role="cpam")
+    result = call_ollama(prompt, temperature=0.1, max_tokens=8000, role="cpam")
     if result is not None:
         logger.info("  Contre-argumentation via Ollama")
     else:
         logger.info("  Ollama indisponible → fallback Anthropic Haiku")
-        result = call_anthropic(prompt, temperature=0.1, max_tokens=16000)
+        result = call_anthropic(prompt, temperature=0.1, max_tokens=8000)
         if result is not None:
             logger.info("  Contre-argumentation via Anthropic Haiku")
 
@@ -213,8 +284,8 @@ def generate_cpam_response(
         if adversarial_warnings:
             adversarial_warnings.append(f"Score de confiance : {score}/10")
 
-    # 8b. Boucle de correction (max 2 retries)
-    max_corrections = 2
+    # 8b. Boucle de correction (configurable via T2A_CPAM_MAX_CORRECTIONS, défaut 2)
+    max_corrections = int(os.environ.get("T2A_CPAM_MAX_CORRECTIONS", "2"))
     for attempt in range(max_corrections):
         if not (validation
                 and not validation.get("coherent", True)