fix(llm): resolve score runtime config from saved profiles

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

webapp/services/inline_scorer.py

@@ -54,13 +54,22 @@ class InlineScorer:
         self._model_cache: dict[tuple[str, str], tuple[Any, Any]] = {}
         self._lock = threading.Lock()
 
+    def invalidate_cache(self) -> None:
+        """Clear the model cache so the next call rebuilds clients from current profiles."""
+        with self._lock:
+            self._model_cache.clear()
+
     def _get_models(
         self,
         judge_model: str,
         embedding_model: str,
         settings: EvaluationSettings,
     ) -> tuple[Any, Any]:
-        """Return cached LLM/embedding clients, building them on first use."""
+        """Return cached LLM/embedding clients, building them on first use.
+
+        Cache is keyed by (judge_model, embedding_model). Call invalidate_cache()
+        after updating an LLM Profile to force a fresh client on the next request.
+        """
         cache_key = (judge_model, embedding_model)
         with self._lock:
             if cache_key not in self._model_cache: