Document decode harness one-shot mechanism

tests/test_core_flow.py

@@ -480,6 +480,7 @@ class CoreFlowTests(unittest.TestCase):
             )
             study = load_study_spec(study_path)
             self.assertEqual(study.trace.request_mode, "decode_only")
+            self.assertTrue(study.trace.restart_engine_after_early_stop)
             window, requests = load_trace_requests(study, study_spec_path=study_path)
             prompt = build_prompt(
                 study=study,
@@ -488,9 +489,31 @@ class CoreFlowTests(unittest.TestCase):
                 capability_profile=None,
             )
             self.assertIn('"request_mode": "decode_only"', prompt)
+            self.assertIn('"restart_engine_after_early_stop": true', prompt)
             self.assertIn("There is no TTFT SLO for this study.", prompt)
             self.assertIn("decode-only", prompt)
 
+    def test_decode_only_restart_after_early_stop_can_be_disabled(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            tmp_path = Path(tmp)
+            study_path = _write_study_assets(
+                tmp_path,
+                trace_overrides={
+                    "request_mode": "decode_only",
+                    "restart_engine_after_early_stop": False,
+                },
+            )
+            study = load_study_spec(study_path)
+            self.assertFalse(study.trace.restart_engine_after_early_stop)
+
+    def test_chat_mode_does_not_restart_after_early_stop_by_default(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            tmp_path = Path(tmp)
+            study_path = _write_study_assets(tmp_path)
+            study = load_study_spec(study_path)
+            self.assertEqual(study.trace.request_mode, "chat")
+            self.assertFalse(study.trace.restart_engine_after_early_stop)
+
     def test_decode_only_harness_defaults_to_decode_tpot(self) -> None:
         with tempfile.TemporaryDirectory() as tmp:
             tmp_path = Path(tmp)