feat: timings appels LLM CPAM (génération, validation, correction)

Ajoute des mesures time.time() autour de chaque appel Ollama dans le flux CPAM :
- [CPAM-EXTRACT] : extraction structurée (passe 1, role=cpam)
- [CPAM-GEN] : génération argumentation (passe 2, role=cpam)
- [CPAM-VALID] : validation adversariale (role=validation)
- [CPAM-CORR] : correction post-validation (role=cpam)

Permet de mesurer le temps réel de chaque phase et d'identifier
les coûts de swap de modèle VRAM entre les rôles cpam/validation.

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

src/control/cpam_response.py

@@ -11,6 +11,7 @@ from __future__ import annotations
 import json
 import logging
 import os
+import time
 from datetime import datetime
 from pathlib import Path
 
@@ -149,14 +150,18 @@ def _extraction_pass(
     )
 
     logger.debug("  Passe 1 — extraction structurée")
+    t0 = time.time()
     result = call_ollama(prompt, temperature=0.0, max_tokens=3000, role="cpam")
     if result is None:
         result = call_anthropic(prompt, temperature=0.0, max_tokens=3000)
+    elapsed = time.time() - t0
     if result is not None:
-        logger.info("  Passe 1 OK : %d éléments cliniques extraits",
+        logger.info("  [CPAM-EXTRACT] %.1fs — OGC %s — %d éléments cliniques extraits",
+                     elapsed, controle.numero_ogc,
                      len(result.get("elements_cliniques_pertinents", [])))
     else:
-        logger.warning("  Passe 1 échouée — fallback single-pass")
+        logger.warning("  [CPAM-EXTRACT] %.1fs — OGC %s — passe 1 échouée",
+                        elapsed, controle.numero_ogc)
     return result
 
 
@@ -195,14 +200,17 @@ def generate_cpam_response(
     prompt, tag_map = _build_cpam_prompt(dossier, controle, sources, extraction)
 
     # 4. Appel LLM — Ollama (rôle cpam) > Haiku fallback
+    t_gen = time.time()
     result = call_ollama(prompt, temperature=0.1, max_tokens=8000, role="cpam")
     if result is not None:
-        logger.info("  Contre-argumentation via Ollama")
+        logger.info("  [CPAM-GEN] %.1fs — OGC %s — contre-argumentation via Ollama",
+                     time.time() - t_gen, controle.numero_ogc)
     else:
         logger.info("  Ollama indisponible → fallback Anthropic Haiku")
         result = call_anthropic(prompt, temperature=0.1, max_tokens=8000)
         if result is not None:
-            logger.info("  Contre-argumentation via Anthropic Haiku")
+            logger.info("  [CPAM-GEN] %.1fs — OGC %s — contre-argumentation via Anthropic",
+                         time.time() - t_gen, controle.numero_ogc)
 
     # 5. Conversion des sources RAG
     rag_sources = [
@@ -285,9 +293,12 @@ def generate_cpam_response(
                         validation.get("score_confiance"), attempt + 1, max_corrections, len(erreurs_v))
 
         correction_prompt = _build_correction_prompt(prompt, result, validation)
+        t_corr = time.time()
         corrected = call_ollama(correction_prompt, temperature=0.0, max_tokens=16000, role="cpam")
         if corrected is None:
             corrected = call_anthropic(correction_prompt, temperature=0.0, max_tokens=16000)
+        logger.info("  [CPAM-CORR] %.1fs — OGC %s — correction %d/%d",
+                     time.time() - t_corr, controle.numero_ogc, attempt + 1, max_corrections)
 
         if not corrected:
             break

src/control/cpam_validation.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import logging
 import re
+import time
 
 from ..config import ControleCPAM, DossierMedical
 from ..medical.bio_normals import BIO_NORMALS
@@ -477,11 +478,14 @@ def _validate_adversarial(
     )
 
     logger.debug("  Validation adversariale")
+    t_val = time.time()
     result = call_ollama(prompt, temperature=0.0, max_tokens=6000, role="validation")
     if result is None:
         result = call_anthropic(prompt, temperature=0.0, max_tokens=6000)
+    elapsed = time.time() - t_val
     if result is None:
-        logger.warning("  Validation adversariale échouée — LLM indisponible")
+        logger.warning("  [CPAM-VALID] %.1fs — OGC %s — validation adversariale échouée",
+                        elapsed, controle.numero_ogc)
         return None
 
     coherent = result.get("coherent", True)
@@ -489,12 +493,13 @@ def _validate_adversarial(
     score = result.get("score_confiance", -1)
 
     if not coherent and erreurs:
-        logger.warning("  Validation adversariale : %d incohérence(s) détectée(s) (score %s/10)",
-                        len(erreurs), score)
+        logger.warning("  [CPAM-VALID] %.1fs — OGC %s — %d incohérence(s) (score %s/10)",
+                        elapsed, controle.numero_ogc, len(erreurs), score)
         for e in erreurs:
             logger.warning("    - %s", e)
     else:
-        logger.info("  Validation adversariale OK (score %s/10)", score)
+        logger.info("  [CPAM-VALID] %.1fs — OGC %s — OK (score %s/10)",
+                     elapsed, controle.numero_ogc, score)
 
     return result