fix: ground Work reports in measurements

tests/test_agent_loop.py

@@ -11,6 +11,7 @@ from agent_platform.runtime.loop import (
     AgentLoop,
     deliverable_evidence_failure,
     is_substantive_execution,
+    measurement_tokens,
     non_source_execution_error,
     normalize_write_file_content,
     python_source_error,
@@ -77,7 +78,8 @@ def test_script_and_report_require_distinct_verified_artifacts() -> None:
             written_source_paths={"benchmark.py"},
             written_report_paths=set(),
             reports_written_after_execution=set(),
-            verified_paths={"benchmark.py"},
+            report_grounding_required=True,
+            grounded_report_paths=set(),
         )
         == "Create a separate report file such as `report.md`."
     )
@@ -88,12 +90,18 @@ def test_script_and_report_require_distinct_verified_artifacts() -> None:
             written_source_paths={"benchmark.py"},
             written_report_paths={"report.md"},
             reports_written_after_execution={"report.md"},
-            verified_paths={"benchmark.py", "report.md"},
+            report_grounding_required=True,
+            grounded_report_paths={"report.md"},
         )
         is None
     )
 
 
+def test_measurement_tokens_only_capture_exact_numeric_measurements() -> None:
+    assert measurement_tokens("n=1000 elapsed=0.012345s score=4.7e-06") == {"0.012345", "4.7e-06"}
+    assert measurement_tokens("O(n²), version 3") == set()
+
+
 def test_read_only_shell_commands_do_not_satisfy_execution_evidence() -> None:
     assert not is_substantive_execution("exec", {"command": "cat report.txt"})
     assert not is_substantive_execution("exec", {"command": "sed -n '1,20p' script.py"})
@@ -461,7 +469,7 @@ async def test_script_and_report_completion_requires_both_files(settings) -> Non
             max_tool_output_chars=10_000,
         ).run(
             spec=get_model_spec("work-light"),
-            messages=[{"role": "user", "content": "写脚本对比算法并给一个报告"}],
+            messages=[{"role": "user", "content": "写一个脚本并给一个报告"}],
             identity=UserIdentity("u1", "", "", "user"),
             raw_user_jwt="jwt",
             chat_id="script-report",