Add read-only production runtime audit

apps/ai-service/rag/answer.py

@@ -839,6 +839,7 @@ class GroundedAnswerService:
         evidence_drug_ids: tuple[str | None, ...] = (),
         budget: RequestBudget | None = None,
         plan: AnswerPlan | None = None,
+        retry_unsupported_patient_list: bool = True,
     ) -> "_GenOutcome":
         """A verified generation, a clarifying question, or empty to fall back."""
         if self._generator is None or not evidence_texts:
@@ -935,6 +936,26 @@ class GroundedAnswerService:
             )
             return _GenOutcome(reject_reason=verification.reason)
         if not verification.supported:
+            # Patient-specific candidate comparisons occasionally receive a
+            # noisy negative entailment verdict even though the same evidence
+            # and a fresh answer clear both fail-closed checks immediately
+            # afterwards (observed in the C03 contextual renal-safety turn).
+            # Retry only this known conversational lane, once. Ordinary AI
+            # answers and monograph browsing are intentionally unchanged.
+            if patient_specific and list_mode and retry_unsupported_patient_list:
+                return self._generate(
+                    query,
+                    evidence_texts,
+                    prompt_evidence_texts,
+                    intro=intro,
+                    list_mode=list_mode,
+                    patient_specific=patient_specific,
+                    candidate_drug_ids=candidate_drug_ids,
+                    evidence_drug_ids=evidence_drug_ids,
+                    budget=budget,
+                    plan=plan,
+                    retry_unsupported_patient_list=False,
+                )
             self._metrics.increment(
                 metric_names.GENERATION_REJECTED, reason="unsupported_claim"
             )