feat: mode hybride Ollama — gemma3:27b pour CPAM, 12b pour codage

Le pipeline utilise désormais gemma3:12b (rapide) pour le codage CIM-10
et gemma3:27b (meilleur raisonnement) pour la contre-argumentation CPAM.
Configurable via OLLAMA_MODEL_CPAM et OLLAMA_TIMEOUT_CPAM.

Inclut aussi : traçabilité source/page DAS, niveaux CMA ATIH, sévérité,
page tracker PDF, améliorations fusion et filtres DAS.

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

tests/test_cpam_response.py

@@ -19,6 +19,7 @@ from src.control.cpam_response import (
     _build_cpam_prompt,
     _format_response,
     _search_rag_for_control,
+    _validate_references,
     generate_cpam_response,
 )
 
@@ -173,6 +174,31 @@ class TestBuildPrompt:
         assert "contre_arguments_asymetrie" in prompt
         assert "contre_arguments_reglementaires" in prompt
 
+    def test_prompt_contains_cite_exacts(self):
+        """Le prompt renforcé demande des preuves exactes."""
+        dossier = _make_dossier()
+        controle = _make_controle()
+        prompt = _build_cpam_prompt(dossier, controle, [])
+
+        assert "CITE" in prompt
+        assert "EXACTS" in prompt
+
+    def test_prompt_contains_interdiction(self):
+        """Le prompt interdit les références inventées."""
+        dossier = _make_dossier()
+        controle = _make_controle()
+        prompt = _build_cpam_prompt(dossier, controle, [])
+
+        assert "INTERDICTION ABSOLUE" in prompt
+
+    def test_prompt_contains_preuves_dossier_field(self):
+        """Le format JSON demandé inclut preuves_dossier."""
+        dossier = _make_dossier()
+        controle = _make_controle()
+        prompt = _build_cpam_prompt(dossier, controle, [])
+
+        assert "preuves_dossier" in prompt
+
 
 class TestFormatResponse:
     def test_full_response_new_format(self):
@@ -236,11 +262,94 @@ class TestFormatResponse:
         text = _format_response({})
         assert text == ""
 
+    def test_preuves_dossier_formatting(self):
+        """Le nouveau champ preuves_dossier est formaté correctement."""
+        parsed = {
+            "contre_arguments_medicaux": "Arguments...",
+            "preuves_dossier": [
+                {"element": "biologie", "valeur": "CRP 180 mg/L", "signification": "inflammation sévère"},
+                {"element": "imagerie", "valeur": "lithiase cholédocienne", "signification": "confirme le diagnostic"},
+            ],
+            "conclusion": "Conclusion...",
+        }
+        text = _format_response(parsed)
+
+        assert "PREUVES DU DOSSIER" in text
+        assert "CRP 180 mg/L" in text
+        assert "[biologie]" in text
+        assert "[imagerie]" in text
+
+    def test_structured_references_formatting(self):
+        """Les références structurées sont formatées correctement."""
+        parsed = {
+            "contre_arguments_medicaux": "Arguments...",
+            "references": [
+                {"document": "Guide Méthodologique MCO 2026", "page": "64", "citation": "Le DAS doit être..."},
+            ],
+            "conclusion": "Conclusion...",
+        }
+        text = _format_response(parsed)
+
+        assert "REFERENCES" in text
+        assert "Guide Méthodologique MCO 2026" in text
+        assert "p.64" in text
+        assert "Le DAS doit être..." in text
+
+    def test_ref_warnings_appended(self):
+        """Les avertissements de références non vérifiées apparaissent."""
+        parsed = {"conclusion": "Conclusion..."}
+        warnings = ["Référence non vérifiable : Manuel Imaginaire 2025"]
+        text = _format_response(parsed, ref_warnings=warnings)
+
+        assert "AVERTISSEMENT" in text
+        assert "Manuel Imaginaire 2025" in text
+
+
+class TestValidateReferences:
+    def test_valid_reference_no_warning(self):
+        parsed = {
+            "references": [
+                {"document": "Guide Méthodologique MCO 2026", "page": "64", "citation": "..."},
+            ]
+        }
+        sources = [{"document": "guide_methodo", "page": 64, "extrait": "..."}]
+        warnings = _validate_references(parsed, sources)
+        assert len(warnings) == 0
+
+    def test_invented_reference_detected(self):
+        parsed = {
+            "references": [
+                {"document": "Manuel Inventé 2025", "page": "12", "citation": "..."},
+            ]
+        }
+        sources = [{"document": "guide_methodo", "page": 64, "extrait": "..."}]
+        warnings = _validate_references(parsed, sources)
+        assert len(warnings) == 1
+        assert "Manuel Inventé" in warnings[0]
+
+    def test_old_format_string_no_crash(self):
+        """L'ancien format string pour references ne cause pas de crash."""
+        parsed = {"references": "Guide méthodo p.64"}
+        sources = [{"document": "guide_methodo"}]
+        warnings = _validate_references(parsed, sources)
+        assert len(warnings) == 0  # pas de validation sur l'ancien format
+
+    def test_no_sources_no_validation(self):
+        parsed = {
+            "references": [
+                {"document": "Quelque chose", "page": "1", "citation": "..."},
+            ]
+        }
+        warnings = _validate_references(parsed, [])
+        assert len(warnings) == 0
+
 
 class TestGenerateResponse:
     @patch("src.control.cpam_response.call_ollama")
+    @patch("src.control.cpam_response.call_anthropic")
     @patch("src.control.cpam_response._search_rag_for_control")
-    def test_generate_success(self, mock_rag, mock_ollama):
+    def test_generate_success_ollama_cpam(self, mock_rag, mock_anthropic, mock_ollama):
+        """Mode hybride : Ollama CPAM (27b) disponible → utilisé en premier."""
         mock_rag.return_value = [
             {"document": "guide_methodo", "page": 64, "extrait": "Texte guide"},
         ]
@@ -259,12 +368,42 @@ class TestGenerateResponse:
         assert "Contre-arguments médicaux..." in text
         assert len(sources) == 1
         assert sources[0].document == "guide_methodo"
+        # Ollama CPAM appelé en premier (avec model= et timeout=)
         mock_ollama.assert_called_once()
+        mock_anthropic.assert_not_called()
 
     @patch("src.control.cpam_response.call_ollama")
+    @patch("src.control.cpam_response.call_anthropic")
     @patch("src.control.cpam_response._search_rag_for_control")
-    def test_generate_ollama_unavailable(self, mock_rag, mock_ollama):
+    def test_generate_fallback_haiku(self, mock_rag, mock_anthropic, mock_ollama):
+        """Ollama CPAM indisponible → fallback Haiku."""
+        mock_rag.return_value = [
+            {"document": "guide_methodo", "page": 64, "extrait": "Texte guide"},
+        ]
+        mock_ollama.return_value = None
+        mock_anthropic.return_value = {
+            "analyse_contestation": "Analyse Haiku...",
+            "contre_arguments_medicaux": "Contre-args Haiku...",
+            "conclusion": "Conclusion Haiku...",
+        }
+
+        dossier = _make_dossier()
+        controle = _make_controle()
+
+        text, sources = generate_cpam_response(dossier, controle)
+
+        assert "Contre-args Haiku..." in text
+        # Ollama CPAM appelé d'abord (échec), puis Haiku
+        mock_ollama.assert_called_once()
+        mock_anthropic.assert_called_once()
+
+    @patch("src.control.cpam_response.call_ollama")
+    @patch("src.control.cpam_response.call_anthropic")
+    @patch("src.control.cpam_response._search_rag_for_control")
+    def test_generate_all_unavailable(self, mock_rag, mock_anthropic, mock_ollama):
+        """Ollama CPAM, Haiku et Ollama défaut tous indisponibles → texte vide."""
         mock_rag.return_value = []
+        mock_anthropic.return_value = None
         mock_ollama.return_value = None
 
         dossier = _make_dossier()