feat: add evaluation label mapping

tests/test_evaluation_runtime.py

@@ -16,6 +16,7 @@ from src.evaluation_runtime import (
     calculate_metrics,
     extract_accessed_refs,
     labels_equivalent,
+    normalize_label_mapping_output,
     normalize_label_master_result,
     normalize_evaluation_output,
     parse_dataset_bytes,
@@ -288,6 +289,22 @@ class EvaluationRuntimeTests(unittest.TestCase):
                 '地图导航',
             )
         )
+        self.assertTrue(labels_equivalent('慢系统', '慢'))
+        self.assertTrue(
+            labels_equivalent(
+                'complex=Ture\n慢系统',
+                '慢',
+                True,
+            )
+        )
+        self.assertTrue(labels_equivalent('complex=Ture', '慢', True))
+        self.assertFalse(
+            labels_equivalent(
+                'complex=Ture\n慢系统',
+                '慢',
+                False,
+            )
+        )
         snapshot_metadata = json.dumps(
             {
                 'label_catalog': {
@@ -340,6 +357,24 @@ class EvaluationRuntimeTests(unittest.TestCase):
             normalized_label['function_output'],
             'Agent(tag="地图导航")',
         )
+        mapping = normalize_label_mapping_output(
+            json.dumps(
+                {
+                    'label_map': {
+                        '慢系统': '慢',
+                        'complex=Ture': 'complex=true',
+                    },
+                    'unmapped': [],
+                    'reason': '统一快慢和复杂度格式',
+                },
+                ensure_ascii=False,
+            ),
+            raw_labels=['慢系统', 'complex=Ture', '未知'],
+            skill_name='label-fast-slow-routing',
+        )
+        self.assertEqual(mapping['label_map']['慢系统'], '慢')
+        self.assertEqual(mapping['label_map']['complex=Ture'], 'complex=true')
+        self.assertEqual(mapping['unmapped'], ['未知'])
 
     def test_fast_slow_prompt_uses_skill_contract_and_relative_paths(self) -> None:
         skill = BundledSkill(
@@ -351,7 +386,15 @@ class EvaluationRuntimeTests(unittest.TestCase):
         prompt, runtime_context = build_evaluation_prompt(
             skill=skill,
             snapshot_root=Path('/tmp/evaluation-snapshot'),
-            canonical_case={'query': '打开空调', 'gold_label': '快'},
+            canonical_case={
+                'query': '打开空调',
+                'gold_label': '快',
+                'source_row': {
+                    'query': '打开空调',
+                    '人工标签': '快',
+                    'secret': 'should-not-leak',
+                },
+            },
         )
         self.assertIn('填写“快”“慢”或“模糊”', prompt)
         self.assertNotIn('填写 fast 或 slow', prompt)
@@ -360,6 +403,64 @@ class EvaluationRuntimeTests(unittest.TestCase):
             '不要添加 `skills/label-fast-slow-routing/` 前缀',
             prompt,
         )
+        self.assertNotIn('should-not-leak', prompt)
+        self.assertNotIn('"人工标签"', prompt)
+
+    def test_label_mapping_suggestion_uses_one_agent_run(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_fast_slow_skill(root)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            try:
+                dataset = runtime.create_dataset(
+                    account_id='alice',
+                    name='fast-slow',
+                    filename='fast-slow.csv',
+                    rows=[
+                        {'query': '打开空调', 'label': '快系统'},
+                        {'query': '帮我规划路线', 'label': '慢系统'},
+                    ],
+                )
+                fake_result = AgentRunResult(
+                    final_output=json.dumps(
+                        {
+                            'label_map': {
+                                '快系统': '快',
+                                '慢系统': '慢',
+                            },
+                            'unmapped': [],
+                            'reason': '统一路由标签',
+                        },
+                        ensure_ascii=False,
+                    ),
+                    turns=1,
+                    tool_calls=1,
+                    transcript=(),
+                    usage=UsageStats(input_tokens=10, output_tokens=5),
+                )
+                with patch(
+                    'src.evaluation_runtime.LocalCodingAgent.run',
+                    return_value=fake_result,
+                ) as run:
+                    suggestion = runtime.suggest_label_mapping(
+                        dataset['id'],
+                        'alice',
+                        mapping=dataset['mapping'],
+                        skill_name='label-fast-slow-routing',
+                    )
+                self.assertEqual(run.call_count, 1)
+                self.assertEqual(
+                    suggestion['label_map'],
+                    {'快系统': '快', '慢系统': '慢'},
+                )
+            finally:
+                runtime.shutdown()
 
     def test_single_analysis_uses_snapshot_without_dataset_records(self) -> None:
         with tempfile.TemporaryDirectory() as directory: