fix: back off evaluation rate limits

tests/test_evaluation_runtime.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import base64
 import json
 import tempfile
+import threading
 import time
 import unittest
 from pathlib import Path
@@ -21,6 +22,7 @@ from src.evaluation_runtime import (
     normalize_evaluation_output,
     parse_dataset_bytes,
     prediction_is_allowed,
+    transient_model_error_message,
 )
 
 
@@ -282,6 +284,11 @@ class EvaluationRuntimeTests(unittest.TestCase):
         )
         self.assertEqual(result['prediction'], '地图导航')
         self.assertEqual(result['confidence'], 1.0)
+        duplicated = normalize_evaluation_output(
+            '{"prediction":"慢","complex":false}'
+            '{"prediction":"慢","complex":false}'
+        )
+        self.assertEqual(duplicated['prediction'], '慢')
         self.assertTrue(labels_equivalent('Agent(tag="地图导航")', '地图导航'))
         self.assertTrue(
             labels_equivalent(
@@ -375,6 +382,61 @@ class EvaluationRuntimeTests(unittest.TestCase):
         self.assertEqual(mapping['label_map']['慢系统'], '慢')
         self.assertEqual(mapping['label_map']['complex=Ture'], 'complex=true')
         self.assertEqual(mapping['unmapped'], ['未知'])
+        self.assertIn(
+            'HTTP 429',
+            transient_model_error_message(
+                'HTTP 429 from local model backend: Too many requests'
+            ),
+        )
+
+    def test_transient_model_failure_retries_before_validation(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            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,
+            )
+            transient = AgentRunResult(
+                final_output='HTTP 429: Too many requests',
+                turns=0,
+                tool_calls=0,
+                transcript=(),
+                usage=UsageStats(),
+            )
+            success = AgentRunResult(
+                final_output='{"prediction":"快"}',
+                turns=1,
+                tool_calls=0,
+                transcript=(),
+                usage=UsageStats(),
+            )
+
+            class StubAgent:
+                def __init__(self) -> None:
+                    self.calls = 0
+
+                def run(self, *_args: object, **_kwargs: object) -> AgentRunResult:
+                    self.calls += 1
+                    return transient if self.calls == 1 else success
+
+            agent = StubAgent()
+            try:
+                with patch.object(runtime, '_extend_model_cooldown') as cooldown:
+                    result = runtime._run_agent_with_transient_backoff(
+                        agent,  # type: ignore[arg-type]
+                        'prompt',
+                        session_id='session',
+                        runtime_context='context',
+                        cancel_event=threading.Event(),
+                    )
+                self.assertIs(result, success)
+                self.assertEqual(agent.calls, 2)
+                cooldown.assert_called_once()
+            finally:
+                runtime.shutdown()
 
     def test_fast_slow_prompt_uses_skill_contract_and_relative_paths(self) -> None:
         skill = BundledSkill(