Add qwen235b prefill-only tuning support

tests/test_core_flow.py

@@ -666,6 +666,44 @@ class CoreFlowTests(unittest.TestCase):
             self.assertFalse(evaluations[1].passed)
             self.assertEqual(summary["slo_pass_rate"], 0.5)
 
+    def test_trace_completion_tokens_override_forces_min_and_max_tokens(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            study_path = _write_study_assets(
+                Path(tmp),
+                trace_overrides={"completion_tokens_override": 1},
+            )
+            study = load_study_spec(study_path)
+            _, requests = load_trace_requests(study, study_spec_path=study_path)
+            self.assertEqual(len(requests), 3)
+            self.assertEqual(requests[0].completion_tokens_hint, 1)
+            self.assertEqual(requests[1].completion_tokens_hint, 1)
+            self.assertEqual(requests[2].completion_tokens_hint, 1)
+            self.assertEqual(requests[0].body["min_tokens"], 1)
+            self.assertEqual(requests[0].body["max_tokens"], 1)
+            self.assertEqual(requests[2].body["min_tokens"], 1)
+            self.assertEqual(requests[2].body["max_tokens"], 1)
+
+    def test_build_prompt_mentions_completion_tokens_override(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            study_path = _write_study_assets(
+                Path(tmp),
+                trace_overrides={"completion_tokens_override": 1},
+                slo_overrides={"tpot_rule": None},
+            )
+            study = load_study_spec(study_path)
+            store = StudyStore(Path(tmp) / ".aituner")
+            store.init_study(spec_path=study_path, study=study)
+            state = store.load_state(study.study_id)
+            window, requests = load_trace_requests(study, study_spec_path=study_path)
+            prompt = build_prompt(
+                study=study,
+                window_summary=summarize_window(requests, window),
+                state=state,
+                capability_profile=None,
+            )
+            self.assertIn('"completion_tokens_override": 1', prompt)
+            self.assertIn("min_tokens=max_tokens=1", prompt)
+
     def test_slo_evaluation_supports_tpot_only_95_percent_target(self) -> None:
         with tempfile.TemporaryDirectory() as tmp:
             study = load_study_spec(