feat: 增强任务步骤注入与a11y状态表达，提升树形交互稳定性

- 打通 metadata.steps 传递链路，将任务步骤注入 agent 预测上下文

- 优化 a11y tree 线性化输出：使用中心坐标并新增 states 列（expanded/collapsed/selected 等）

- 放宽可保留节点条件，保留无文本输入类控件（edit/textfield/searchbox 等）

- 强化输出约束：单轮仅允许动作代码或 WAIT/DONE/FAIL，禁止动作与 DONE 同轮返回

- 补充 avogadro 示例步骤：展开 aromatics 并选择 benzene.cjson

mm_agents/accessibility_tree_wrap/heuristic_retrieve.py

@@ -114,6 +114,12 @@ def judge_node(node: ET, platform="ubuntu", check_image=False) -> bool:
             and (
                     node.get("name", "") != "" or node.text is not None and len(node.text) > 0 \
                     or check_image and node.get("image", "false") == "true"
+                    # Keep empty input fields (edit/textfield) - they are important interactive elements
+                    # even without name/text (e.g., search boxes, filter inputs)
+                    or node.tag.endswith("edit") or node.tag.endswith("textfield")
+                    or node.tag.endswith("textarea") or node.tag.endswith("textbox")
+                    or node.tag.endswith("searchbox") or node.tag.endswith("combobox")
+                    or node.tag in {"entry", "combo-box", "check-box", "slider"}
             )
 
     coordinates: Tuple[int, int] = eval(node.get("{{{:}}}screencoord".format(_component_ns), "(-1, -1)"))