tests: dp_finalizer — 20 tests R1-R5 + pass-through + quality_flags + sérialisation

- TestR1CrhConfirmedOverridesTrackare (2 tests : override + cohérent) - TestR2TrackareCorroborated (2 tests : exact + family3) - TestR3TrackareSymptom (3 tests : override, review prudent, evidence faible) - TestR4Ambiguous (1 test) - TestR5Interdictions (4 tests : Z-code, Z-whitelist, R-code, allow_symptom) - TestPassThrough (3 tests : CRH-only, Trackare-only, aucun DP) - TestFinalizeDp (5 tests : flags merge, alertes append, sources set, sérialisation) 1063 tests passent, 0 régression. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-24 17:50:32 +01:00
parent c7317af447
commit 63354e75bc
2 changed files with 381 additions and 2 deletions
--- a/docs/gold_debug/NUKE3_GOLD_TOP_ERRORS.md
+++ b/docs/gold_debug/NUKE3_GOLD_TOP_ERRORS.md
@@ -1,6 +1,6 @@
 # NUKE-3 — Top erreurs gold CRH

-**Date** : 2026-02-24 14:34  
+**Date** : 2026-02-24 17:44  
 **Cas** : 5  

 | # | Case ID | Choisi | Attendu | Strict | Accept. | Verdict | Conf. | Delta | Reason |
@@ -12,4 +12,4 @@
 | 5 | 73_23139637 | R06.0 | R06.0 | OK | OK | REVIEW | medium | 1.0 | mono_fragile |

 ---
-*Généré le 2026-02-24 14:34*
+*Généré le 2026-02-24 17:44*
--- a/tests/test_dp_finalizer.py
+++ b/tests/test_dp_finalizer.py
@@ -0,0 +1,379 @@
+"""Tests du DP Finalizer — arbitrage Trackare vs CRH-only.
+
+Pas de mocks. Fixtures synthétiques avec objets réels (DossierMedical, DPSelection, DPCandidate).
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from src.config import DossierMedical, DPSelection, DPCandidate, Diagnostic, Sejour
+from src.medical.dp_finalizer import finalize_dp, decide_dp_final
+
+
+# ── Helpers ────────────────────────────────────────────────────────────
+
+
+def _sel(
+    code: str,
+    term: str = "",
+    verdict: str = "REVIEW",
+    confidence: str = "medium",
+    evidence: list[str] | None = None,
+    reason: str = "",
+    candidates: list[DPCandidate] | None = None,
+) -> DPSelection:
+    """Crée un DPSelection minimal."""
+    return DPSelection(
+        chosen_code=code,
+        chosen_term=term or code,
+        verdict=verdict,
+        confidence=confidence,
+        evidence=evidence or [],
+        reason=reason,
+        candidates=candidates or [],
+    )
+
+
+def _cand(code: str, term: str = "", score: float = 0.0) -> DPCandidate:
+    return DPCandidate(index=0, term=term or code, code=code, score=score)
+
+
+def _dossier(
+    doc_type: str = "crh",
+    dp_code: str | None = None,
+    dp_selection: DPSelection | None = None,
+    existing_flags: dict | None = None,
+    existing_alertes: list[str] | None = None,
+) -> DossierMedical:
+    dp = None
+    if dp_code:
+        dp = Diagnostic(texte=dp_code, cim10_suggestion=dp_code)
+    d = DossierMedical(
+        document_type=doc_type,
+        sejour=Sejour(),
+        diagnostic_principal=dp,
+        dp_selection=dp_selection,
+    )
+    if existing_flags:
+        d.quality_flags = existing_flags.copy()
+    if existing_alertes:
+        d.alertes_codage = existing_alertes.copy()
+    return d
+
+
+# ===================================================================
+# R1 — CRH CONFIRMED override Trackare
+# ===================================================================
+
+
+class TestR1CrhConfirmedOverridesTrackare:
+
+    def test_crh_confirmed_with_evidence_wins(self):
+        """R1 : CRH CONFIRMED + evidence forte → override Trackare."""
+        trackare = _sel("I10", verdict="CONFIRMED", evidence=["Source: Trackare (codage établissement)"],
+                        reason="DP Trackare — source d'autorité")
+        crh = _sel("I26.9", term="Embolie pulmonaire", verdict="CONFIRMED",
+                    evidence=["Diagnostic de sortie: «Embolie pulmonaire»"],
+                    reason="CRH CONFIRMED")
+
+        dp_final, flags, alertes = decide_dp_final(trackare, crh)
+
+        assert dp_final.chosen_code == "I26.9"
+        assert dp_final.verdict == "CONFIRMED"
+        assert flags.get("override_trackare_by_crh_confirmed") is True
+        assert any("Trackare" in a and "écarté" in a for a in alertes)
+
+    def test_crh_confirmed_coherent_no_override_flag(self):
+        """R1 : CRH CONFIRMED + même code que Trackare → pas de flag override."""
+        trackare = _sel("I26.9", verdict="CONFIRMED", evidence=["Source: Trackare"])
+        crh = _sel("I26.9", verdict="CONFIRMED",
+                    evidence=["Diagnostic de sortie: «Embolie pulmonaire»"])
+
+        dp_final, flags, _ = decide_dp_final(trackare, crh)
+
+        assert dp_final.chosen_code == "I26.9"
+        assert "override_trackare_by_crh_confirmed" not in flags
+        assert flags.get("crh_confirmed_coherent") is True
+
+
+# ===================================================================
+# R2 — Trackare non-symptôme corroboré par CRH
+# ===================================================================
+
+
+class TestR2TrackareCorroborated:
+
+    def test_trackare_corroborated_exact_match(self):
+        """R2 : Trackare I26.9 + CRH candidates contient I26.9 → CONFIRMED."""
+        trackare = _sel("I26.9", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+        crh = _sel("I26.9", term="EP", verdict="REVIEW",
+                    candidates=[_cand("I26.9", "Embolie pulmonaire", 5.0),
+                                _cand("Q53.9", "Cryptorchidie", 2.0)])
+
+        dp_final, flags, _ = decide_dp_final(trackare, crh)
+
+        assert dp_final.chosen_code == "I26.9"
+        assert dp_final.verdict == "CONFIRMED"
+        assert flags.get("trackare_confirmed_by_crh") is True
+        assert any("corroboré" in e for e in dp_final.evidence)
+
+    def test_trackare_corroborated_family3(self):
+        """R2 : Trackare I26.0 + CRH I26.9 → family3 match → CONFIRMED."""
+        trackare = _sel("I26.0", verdict="CONFIRMED", evidence=["Source: Trackare"])
+        crh = _sel("I26.9", verdict="REVIEW",
+                    candidates=[_cand("I26.9")])
+
+        dp_final, flags, _ = decide_dp_final(trackare, crh)
+
+        assert dp_final.chosen_code == "I26.0"
+        assert dp_final.verdict == "CONFIRMED"
+        assert flags.get("trackare_confirmed_by_crh") is True
+
+
+# ===================================================================
+# R3 — Trackare symptôme (R*) + CRH étiologique
+# ===================================================================
+
+
+class TestR3TrackareSymptom:
+
+    def test_trackare_symptom_overridden_by_crh_confirmed(self):
+        """R3 : Trackare R59.0 + CRH CONFIRMED C83.3 → override."""
+        trackare = _sel("R59.0", term="Adénopathie", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+        crh = _sel("C83.3", term="DLBCL", verdict="CONFIRMED",
+                    evidence=["Conclusion: «DLBCL en progression»"])
+
+        dp_final, flags, alertes = decide_dp_final(trackare, crh)
+
+        assert dp_final.chosen_code == "C83.3"
+        assert dp_final.verdict == "CONFIRMED"
+        assert flags.get("trackare_symptom_overridden") is True
+        assert any("R59.0" in a and "C83.3" in a for a in alertes)
+
+    def test_trackare_symptom_review_when_crh_not_confirmed(self):
+        """R3 : Trackare R59.0 + CRH REVIEW C83.3 → REVIEW prudent."""
+        trackare = _sel("R59.0", term="Adénopathie", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+        crh = _sel("C83.3", term="DLBCL", verdict="REVIEW",
+                    evidence=["Scores proches"])
+
+        dp_final, flags, alertes = decide_dp_final(trackare, crh)
+
+        assert dp_final.chosen_code == "R59.0"  # reste Trackare
+        assert dp_final.verdict == "REVIEW"
+        assert dp_final.confidence == "medium"
+        assert flags.get("trackare_symptom_vs_crh_diagnosis") is True
+        assert any("vérification DIM" in a for a in alertes)
+
+    def test_trackare_symptom_review_when_crh_weak_evidence(self):
+        """R3 : Trackare R06.0 + CRH CONFIRMED mais evidence vide → REVIEW."""
+        trackare = _sel("R06.0", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+        crh = _sel("J18.9", verdict="CONFIRMED", evidence=[])  # pas de preuve forte
+
+        dp_final, flags, _ = decide_dp_final(trackare, crh)
+
+        assert dp_final.verdict == "REVIEW"
+        assert flags.get("trackare_symptom_vs_crh_diagnosis") is True
+
+
+# ===================================================================
+# R4 — Ambigu / preuves faibles
+# ===================================================================
+
+
+class TestR4Ambiguous:
+
+    def test_trackare_non_r_crh_review_different_code(self):
+        """R4 : Trackare K85.1 + CRH REVIEW K85.9 non corroboré → ambigu."""
+        # K85.1 vs K85.9 → family3 match → actually R2 triggers
+        # Use truly different codes
+        trackare = _sel("K85.1", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+        crh = _sel("N17.9", verdict="REVIEW",
+                    candidates=[_cand("N17.9", "IRA", 3.0)])
+
+        dp_final, flags, alertes = decide_dp_final(trackare, crh)
+
+        assert dp_final.verdict == "REVIEW"
+        assert flags.get("review_ambiguous") is True
+        assert any("ambigu" in a.lower() for a in alertes)
+
+
+# ===================================================================
+# R5 — Z-code / R-code interdits auto-confirm
+# ===================================================================
+
+
+class TestR5Interdictions:
+
+    def test_z_code_never_confirmed(self):
+        """R5 : Z95.5 (non whitelisté) → forcer REVIEW."""
+        trackare = _sel("Z95.5", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+
+        dp_final, flags, alertes = decide_dp_final(trackare, None)
+
+        assert dp_final.verdict == "REVIEW"
+        assert flags.get("z_code_dp_review") is True
+
+    def test_z_code_whitelisted_stays_confirmed(self):
+        """R5 : Z51.1 (whitelisté) → CONFIRMED ok."""
+        trackare = _sel("Z51.1", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+
+        dp_final, flags, _ = decide_dp_final(trackare, None)
+
+        assert dp_final.verdict == "CONFIRMED"
+        assert "z_code_dp_review" not in flags
+
+    def test_r_code_not_allowed_if_non_r_candidate(self):
+        """R5 : R06.0 CONFIRMED + candidat J18.9 non-R → REVIEW."""
+        crh = _sel("R06.0", term="Dyspnée", verdict="CONFIRMED",
+                    evidence=["Diagnostic de sortie: «Dyspnée»"],
+                    candidates=[_cand("R06.0", "Dyspnée", 5.0),
+                                _cand("J18.9", "Pneumopathie", 3.0)])
+
+        dp_final, flags, _ = decide_dp_final(None, crh, allow_symptom_dp=False)
+
+        assert dp_final.verdict == "REVIEW"
+        assert flags.get("r_code_dp_with_non_r_candidate") is True
+
+    def test_r_code_allowed_when_flag_true(self):
+        """R5 : R06.0 CONFIRMED + allow_symptom_dp=True → CONFIRMED ok."""
+        crh = _sel("R06.0", term="Dyspnée", verdict="CONFIRMED",
+                    evidence=["Diagnostic de sortie: «Dyspnée»"],
+                    candidates=[_cand("R06.0", "Dyspnée", 5.0),
+                                _cand("J18.9", "Pneumopathie", 3.0)])
+
+        dp_final, flags, _ = decide_dp_final(None, crh, allow_symptom_dp=True)
+
+        assert dp_final.verdict == "CONFIRMED"
+        assert "r_code_dp_with_non_r_candidate" not in flags
+
+
+# ===================================================================
+# Cas dégénérés — pass-through
+# ===================================================================
+
+
+class TestPassThrough:
+
+    def test_no_trackare_crh_only(self):
+        """CRH-only → pass-through."""
+        crh = _sel("A87.0", term="Méningite entérovirus", verdict="CONFIRMED",
+                    evidence=["Diagnostic de sortie: «Méningite à entérovirus»"])
+
+        dp_final, flags, _ = decide_dp_final(None, crh)
+
+        assert dp_final.chosen_code == "A87.0"
+        assert dp_final.verdict == "CONFIRMED"
+        assert flags.get("crh_only_mode") is True
+
+    def test_no_crh_trackare_only(self):
+        """Trackare-only → CONFIRMED (si non Z/R)."""
+        trackare = _sel("K81.0", term="Cholécystite", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"])
+
+        dp_final, flags, _ = decide_dp_final(trackare, None)
+
+        assert dp_final.chosen_code == "K81.0"
+        assert dp_final.verdict == "CONFIRMED"
+        assert flags.get("trackare_only_mode") is True
+
+    def test_no_sources_returns_review(self):
+        """Aucun DP → REVIEW."""
+        dp_final, flags, alertes = decide_dp_final(None, None)
+
+        assert dp_final.verdict == "REVIEW"
+        assert flags.get("no_dp_source") is True
+        assert len(alertes) > 0
+
+
+# ===================================================================
+# Intégration finalize_dp() — dossier complet
+# ===================================================================
+
+
+class TestFinalizeDp:
+
+    def test_quality_flags_merge(self):
+        """Les flags existants ne sont pas écrasés par le finalizer."""
+        crh = _sel("A87.0", verdict="CONFIRMED",
+                    evidence=["Diagnostic de sortie: «Méningite»"])
+        d = _dossier(doc_type="crh", dp_selection=crh,
+                     existing_flags={"my_existing_flag": True})
+
+        finalize_dp(d)
+
+        assert d.quality_flags["my_existing_flag"] is True
+        assert d.quality_flags.get("crh_only_mode") is True
+
+    def test_alertes_codage_appended(self):
+        """Le finalizer ajoute des alertes sans supprimer les existantes."""
+        trackare = _sel("R59.0", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"],
+                        reason="DP Trackare — source d'autorité")
+        crh_candidates = [_cand("C83.3", "DLBCL", 5.0)]
+        crh = _sel("C83.3", term="DLBCL", verdict="REVIEW",
+                    evidence=["Scores proches"],
+                    candidates=crh_candidates)
+
+        # Dossier Trackare avec dp_selection trackare, mais on simule la présence
+        # d'un CRH secondaire via un appel direct decide_dp_final
+        d = _dossier(
+            doc_type="trackare",
+            dp_code="R59.0",
+            dp_selection=trackare,
+            existing_alertes=["Alerte existante"],
+        )
+        finalize_dp(d)
+
+        assert "Alerte existante" in d.alertes_codage
+        assert d.dp_trackare is not None
+        assert d.dp_final is not None
+
+    def test_dp_trackare_and_crh_only_set(self):
+        """Vérifie que dp_trackare et dp_crh_only sont correctement renseignés."""
+        crh = _sel("K85.1", term="Pancréatite biliaire", verdict="CONFIRMED",
+                    evidence=["Conclusion: «Pancréatite aiguë biliaire»"])
+        d = _dossier(doc_type="crh", dp_selection=crh)
+
+        finalize_dp(d)
+
+        assert d.dp_trackare is None  # pas de Trackare
+        assert d.dp_crh_only is not None
+        assert d.dp_crh_only.chosen_code == "K85.1"
+        assert d.dp_final is not None
+        assert d.dp_final.chosen_code == "K85.1"
+
+    def test_trackare_dossier_sets_dp_trackare(self):
+        """Un dossier Trackare voit dp_trackare renseigné."""
+        trackare = _sel("K81.0", verdict="CONFIRMED",
+                        evidence=["Source: Trackare (codage établissement)"],
+                        reason="DP Trackare — source d'autorité")
+        d = _dossier(doc_type="trackare", dp_code="K81.0", dp_selection=trackare)
+
+        finalize_dp(d)
+
+        assert d.dp_trackare is not None
+        assert d.dp_trackare.chosen_code == "K81.0"
+        assert d.dp_crh_only is None
+        assert d.dp_final.chosen_code == "K81.0"
+        assert d.dp_final.verdict == "CONFIRMED"
+
+    def test_serializable(self):
+        """Le dossier reste sérialisable en JSON après finalize_dp."""
+        crh = _sel("I26.9", verdict="CONFIRMED",
+                    evidence=["Diagnostic de sortie: «EP»"],
+                    candidates=[_cand("I26.9", "EP", 6.0)])
+        d = _dossier(doc_type="crh", dp_selection=crh)
+        finalize_dp(d)
+
+        data = d.model_dump(exclude_none=True)
+        assert "dp_final" in data
+        assert data["dp_final"]["chosen_code"] == "I26.9"
+        assert "quality_flags" in data