feat: add InlineScorer service with LLM client caching

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

tests/webapp/test_score_api.py

@@ -126,3 +126,62 @@ class TestScoreResponse:
         assert resp.scores["faithfulness"] == 0.85
         assert resp.scores["answer_relevancy"] is None
         assert resp.latency_ms == 1200
+
+
+class TestInlineScorer:
+    def test_score_returns_dict_with_requested_metrics(self):
+        """InlineScorer.score returns a dict keyed by the requested metrics."""
+        from unittest.mock import AsyncMock, MagicMock, patch
+        from webapp.services.inline_scorer import InlineScorer
+        from rag_eval.settings import EvaluationSettings
+
+        mock_score = MagicMock()
+        mock_score.metrics = {"faithfulness": 0.9, "answer_relevancy": 0.8}
+        mock_score.error = ""
+
+        mock_pipeline = MagicMock()
+        mock_pipeline.score_sample = AsyncMock(return_value=mock_score)
+
+        with patch("webapp.services.inline_scorer.build_models", return_value=(MagicMock(), MagicMock())):
+            with patch("webapp.services.inline_scorer.MetricPipeline", return_value=mock_pipeline):
+                with patch("webapp.services.inline_scorer._build_metric_instances", return_value={}):
+                    scorer = InlineScorer()
+                    result = scorer.score(
+                        question="q", answer="a",
+                        contexts=["ctx1"],
+                        ground_truth=None,
+                        metrics=["faithfulness", "answer_relevancy"],
+                        judge_model="test-model",
+                        embedding_model="test-embed",
+                        settings=EvaluationSettings(_env_file=None),
+                    )
+        assert "faithfulness" in result
+        assert "answer_relevancy" in result
+        assert result["faithfulness"] == pytest.approx(0.9)
+
+    def test_score_converts_nan_to_none(self):
+        """NaN scores are converted to None in the returned dict."""
+        import math
+        from unittest.mock import AsyncMock, MagicMock, patch
+        from webapp.services.inline_scorer import InlineScorer
+        from rag_eval.settings import EvaluationSettings
+
+        mock_score = MagicMock()
+        mock_score.metrics = {"faithfulness": float("nan")}
+        mock_score.error = ""
+
+        mock_pipeline = MagicMock()
+        mock_pipeline.score_sample = AsyncMock(return_value=mock_score)
+
+        with patch("webapp.services.inline_scorer.build_models", return_value=(MagicMock(), MagicMock())):
+            with patch("webapp.services.inline_scorer.MetricPipeline", return_value=mock_pipeline):
+                with patch("webapp.services.inline_scorer._build_metric_instances", return_value={}):
+                    scorer = InlineScorer()
+                    result = scorer.score(
+                        question="q", answer="a", contexts=["c"],
+                        ground_truth=None,
+                        metrics=["faithfulness"],
+                        judge_model="m", embedding_model="e",
+                        settings=EvaluationSettings(_env_file=None),
+                    )
+        assert result["faithfulness"] is None