perf(webview2): avoid queuing fire-and-forget scripts

- feat(config): add atomic configuration persistence
- fix(ui): repair chat and model settings rendering
- fix(capture): map selections across high-DPI displays
- fix(webview2): scale native bounds for fractional DPI
- chore: remove test scripts from distribution

core/webview2.py

@@ -214,7 +214,7 @@ class Wv2Session:
         self._js_pump = QTimer()
         self._js_pump.setInterval(25)
         self._js_pump.timeout.connect(self._js_pump_tick)
-        self._js_pump.start()
+        # 无回调脚本不进入队列；没有待处理任务时不需要常驻唤醒 UI 线程。
         # 🆕 预热：立即导航 about:blank，让 msedgewebview2 进程/GPU 在 UI 构建期间冷启动
         # （实测本机首次真实页面导航需 12-15s，预热后降到 ~1s）
         try:
@@ -339,8 +339,27 @@ class Wv2Session:
             print("[WV2] navigate error:", ex)
 
     def execute_js(self, script: str):
-        """fire-and-forget（ChatBridge.run_js 的替换，JS 文本完全同构）"""
-        self._js_run(script, None)
+        """执行无需返回值的脚本，不进入结果轮询队列。
+
+        流式正文每个 token 都会走这里。旧实现统一使用
+        ``ExecuteScriptWithResultAsync`` 并把任务放入 ``_js_pending``，高频
+        输出时会在渲染器和 Python 侧同时堆积大量无用结果；WebView2 原生
+        ``ExecuteScriptAsync`` 已经提供了真正的 fire-and-forget 路径。
+        """
+        try:
+            execute_async = getattr(self.core, "ExecuteScriptAsync")
+        except AttributeError:
+            # 兼容旧版/测试替身未暴露 ExecuteScriptAsync 的情况。即使只能
+            # 使用带结果 API，也直接丢弃 Task，不把无用结果放进轮询队列。
+            try:
+                getattr(self.core, "ExecuteScriptWithResultAsync")(script)
+            except BaseException as ex:
+                print("[WV2] execute_js fallback error:", ex)
+            return
+        try:
+            execute_async(script)
+        except BaseException as ex:
+            print("[WV2] execute_js error:", ex)
 
     def execute_js_async(self, script: str, cb):
         """带回调执行：真异步，回调在主线程定时器 tick 中发出（绝不阻塞）"""
@@ -358,13 +377,25 @@ class Wv2Session:
                     pass
             return
         self._js_pending.append([task, cb, time.time()])
+        try:
+            if not self._js_pump.isActive():
+                self._js_pump.start()
+        except Exception:
+            pass
 
     def _js_pump_tick(self):
         if not self._js_pending:
+            try:
+                self._js_pump.stop()
+            except Exception:
+                pass
             return
         import json
-        remaining = []
-        for task, cb, t0 in self._js_pending:
+        # 先摘下本批次；回调中再次入队的任务写入新的 _js_pending，不能被
+        # 本轮收尾赋值覆盖。
+        pending = self._js_pending
+        self._js_pending = []
+        for task, cb, t0 in pending:
             done = False
             try:
                 done = task.IsCompleted
@@ -373,7 +404,7 @@ class Wv2Session:
             if not done and time.time() - t0 > 10:
                 done = True  # 10s 安全超时（渲染器死亡时不永久卡队列）
             if not done:
-                remaining.append([task, cb, t0])
+                self._js_pending.append([task, cb, t0])
                 continue
             result = None
             try:
@@ -395,7 +426,11 @@ class Wv2Session:
                     cb(result)
                 except Exception as ex:
                     print("[WV2] js callback error:", ex)
-        self._js_pending = remaining
+        if not self._js_pending:
+            try:
+                self._js_pump.stop()
+            except Exception:
+                pass
 
     def close(self):
         try:

ui/views/main_window.py

@@ -9,12 +9,23 @@ import json
 from ui.views.custom_web_page import CustomWebPage
 from PyQt6.QtCore import QUrl # 🌟 新增：用于加载本地 HTML
 
-# ========== 流式诊断日志（排查显示问题的根本手段，写 stream_diag.log） ==========
+# ========== 流式诊断日志（仅显式开启时写 stream_diag.log） ==========
 import time as _diag_time
+
+
+def _stream_diagnostics_enabled() -> bool:
+    """仅在显式要求时开启高开销流式诊断。"""
+    return (os.environ.get("HAOCODE_STREAM_DIAG") == "1"
+            or os.environ.get("HAOCODE_SHOT") == "1")
+
+
 DIAG_LOG_PATH = os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))),
                              "stream_diag.log")
 
+
 def diag_log(line: str):
+    if not _stream_diagnostics_enabled():
+        return
     try:
         t = _diag_time.time()
         with open(DIAG_LOG_PATH, "a", encoding="utf-8") as f:
@@ -1577,7 +1588,7 @@ class SessionContextPopup(QtWidgets.QWidget):
 
         self._anim_group.addAnimation(opacity_anim)
         self._anim_group.addAnimation(pos_anim)
-        self._anim_group.start()
+        self._anim_group.start(QtCore.QAbstractAnimation.DeletionPolicy.DeleteWhenStopped)
 
 class RenameOverlay(QtWidgets.QWidget):
     """无边框重命名 — 独立顶层透明窗口（P2-03 修复）
@@ -2077,7 +2088,7 @@ class ModelSelectPopup(QtWidgets.QWidget):
 
     def reload_config(self, config_data):
         """按最新配置重建列表；配置缺失时仍保留可理解、可绘制的空状态。"""
-        self._finish_drawer()
+        self._finish_drawer(preserve_top=True)
         self.config_data = config_data if isinstance(config_data, dict) else {}
         self.list_widget.clear()
         self._groups.clear()
@@ -2256,7 +2267,9 @@ class ModelSelectPopup(QtWidgets.QWidget):
         if g is None:
             return
         # 上一个抽屉动画进行中 → 先立即收敛到终态（允许快速连点切换）
-        self._finish_drawer()
+        # 折叠/展开只改变高度时锁定当前顶部；否则按底边贴按钮会把
+        # 用户正在看的内容整体向下挤。
+        self._finish_drawer(preserve_top=True)
         g["expanded"] = not g["expanded"]
         # SVG 箭头切换（展开=朝下 / 收起=朝右）
         g["chevron"].setPixmap(
@@ -2264,7 +2277,7 @@ class ModelSelectPopup(QtWidgets.QWidget):
         self._clear_fx()
         self._apply_group_state(g)
         self.list_widget.doItemsLayout()
-        self._apply_height_and_position()
+        self._apply_height_and_position(preserve_top=True)
         self.list_widget.viewport().update()
 
     def _apply_group_state(self, g):
@@ -2404,12 +2417,16 @@ class ModelSelectPopup(QtWidgets.QWidget):
             if d["t"] >= 1.0:
                 self._finish_drawer()
 
-    def _finish_drawer(self):
+    def _finish_drawer(self, preserve_top=False):
         """抽屉动画收敛到终态（动画完成 / 连点打断共用）"""
         d = self._drawer
         if d is None:
             return
+        top_anchor = self.y() if preserve_top and self.isVisible() else None
         d["timer"].stop()
+        # 每次动画都会创建一个 timer；停用后立即释放，避免反复折叠后
+        # 大量已停止的 QObject 留在弹窗树中。
+        d["timer"].deleteLater()
         self._drawer = None
         # 🆕 恢复滑移供应商头的透明背景（_start_drawer 里铺的白底）
         for it in d["after_items"]:
@@ -2425,7 +2442,9 @@ class ModelSelectPopup(QtWidgets.QWidget):
             return
         parent = self.parentWidget()
         btn = getattr(parent, "model_selector", None) if parent else None
-        if btn is not None:
+        if top_anchor is not None:
+            self.move(self.x(), top_anchor)
+        elif btn is not None:
             # 真实 App：回到“按钮正上方右对齐”锚点
             btn_pos = btn.mapToGlobal(QtCore.QPoint(0, 0))
             self.move(btn_pos.x() + btn.width() - self.width(),
@@ -2434,8 +2453,9 @@ class ModelSelectPopup(QtWidgets.QWidget):
             # 无锚定按钮（如调参工具宿主）：保持动画起点底缘不漂
             self.move(self.x(), d["bottom0"] - self.minimumHeight())
 
-    def _apply_height_and_position(self):
-        """折叠/展开后：按可见项重算高度；若弹窗正显示中，保持“按钮正上方右对齐”锚点"""
+    def _apply_height_and_position(self, preserve_top=False):
+        """重算高度；折叠时可锁定顶部，避免当前列表位置被挤走。"""
+        top_anchor = self.y() if preserve_top and self.isVisible() else None
         self.adjust_popup_height()
         if not self.isVisible():
             return
@@ -2445,7 +2465,7 @@ class ModelSelectPopup(QtWidgets.QWidget):
             return
         btn_pos = btn.mapToGlobal(QtCore.QPoint(0, 0))
         x = btn_pos.x() + btn.width() - self.width()
-        y = btn_pos.y() - self.height() - 5
+        y = top_anchor if top_anchor is not None else btn_pos.y() - self.height() - 5
         self.move(x, y)
 
     def adjust_popup_height(self):
@@ -2494,7 +2514,7 @@ class ModelSelectPopup(QtWidgets.QWidget):
 
         self.anim_group.addAnimation(self.opacity_anim)
         self.anim_group.addAnimation(self.pos_anim)
-        self.anim_group.start()
+        self.anim_group.start(QtCore.QAbstractAnimation.DeletionPolicy.DeleteWhenStopped)
 
     def on_item_clicked(self, item):
         data = item.data(QtCore.Qt.ItemDataRole.UserRole)
@@ -2680,7 +2700,7 @@ class SessionModePopup(QtWidgets.QWidget):
         self.pos_anim.setEasingCurve(QtCore.QEasingCurve.Type.OutQuad)
         self.anim_group.addAnimation(self.opacity_anim)
         self.anim_group.addAnimation(self.pos_anim)
-        self.anim_group.start()
+        self.anim_group.start(QtCore.QAbstractAnimation.DeletionPolicy.DeleteWhenStopped)
 
 
 # ==================== 🌟 PDF 图片提取子线程工作者 ====================
@@ -3064,7 +3084,7 @@ class PdfModePopup(QtWidgets.QWidget):
         ps.setEasingCurve(QtCore.QEasingCurve.Type.OutQuad)
         self.anim_group.addAnimation(op)
         self.anim_group.addAnimation(ps)
-        self.anim_group.start()
+        self.anim_group.start(QtCore.QAbstractAnimation.DeletionPolicy.DeleteWhenStopped)
 
     # 关闭方式：依赖 Qt.Popup 原生行为——点击弹窗以外任意区域即关闭（与 ModelSelectPopup 一致），
     # 鼠标移出不再关闭。
@@ -3292,24 +3312,29 @@ class MainWindow(QtWidgets.QMainWindow):
 
         # 🌟 提前初始化（init_model_popup 内部会触发 update_context_display 用到）
         self._active_streams = {}
+        self._stream_diagnostics = _stream_diagnostics_enabled()
         self._diag_chunk_n = 0
         self._diag_think_n = 0
-        try:
-            open(DIAG_LOG_PATH, "w").close()
-            diag_log("APP_START")
-        except Exception:
-            pass
+        if self._stream_diagnostics:
+            try:
+                open(DIAG_LOG_PATH, "w").close()
+                diag_log("APP_START")
+            except Exception:
+                pass
         # 🌟 渲染看门狗：若前端 rAF/定时器被浏览器节流（窗口隐藏/GPU 问题），
         #    由 Qt 侧定时器每 200ms 主动 forceRenderNow，保证流式正文一定上屏
         self._render_watchdog = QtCore.QTimer(self)
         self._render_watchdog.setInterval(200)
         self._render_watchdog.timeout.connect(self._render_watchdog_tick)
         self._render_watchdog.start()
-        # 人肉 debug：JS console 桥抽取器（JS console.log → Python 控制台）
-        self._jslog_timer = QtCore.QTimer(self)
-        self._jslog_timer.setInterval(500)
-        self._jslog_timer.timeout.connect(self._jslog_drain_tick)
-        self._jslog_timer.start()
+        # 人肉 debug：JS console 桥抽取器（JS console.log → Python 控制台）。
+        # 这是诊断通道，不应在普通运行中每 500ms 跨进程执行一次 JS。
+        self._jslog_timer = None
+        if self._stream_diagnostics:
+            self._jslog_timer = QtCore.QTimer(self)
+            self._jslog_timer.setInterval(500)
+            self._jslog_timer.timeout.connect(self._jslog_drain_tick)
+            self._jslog_timer.start()
         # 🆕 Fix E: 上下文标签防抖定时器（工具/思考/正文任一显现 → 400ms 内合并刷新一次）
         self._ctx_refresh_timer = QtCore.QTimer(self)
         self._ctx_refresh_timer.setSingleShot(True)
@@ -4977,7 +5002,8 @@ class MainWindow(QtWidgets.QMainWindow):
         
         if session_id == self.current_session_id:
             self.chat_bridge.finish_message(msg_id)
-            self.browser.page().runJavaScript("dumpDiag()", self._on_diag_dumped)
+            if self._stream_diagnostics:
+                self.browser.page().runJavaScript("dumpDiag()", self._on_diag_dumped)
             print("\n[系统]:✅ 回复完毕。")
         else:
             print(f"\n[系统]: ✅ 会话 {session_id[:8]} 回复完毕（后台）")
@@ -4993,21 +5019,27 @@ class MainWindow(QtWidgets.QMainWindow):
                       f"usage_output={_du.get('output')} timeline={len(_tl0)}条")
         except Exception:
             pass
-        # 完成时刻屏幕快照 + DOM 体检
-        try:
-            self._shot_n = getattr(self, "_shot_n", 0) + 1
-            import os as _os
-            _shot_path = _os.path.join(
-                _os.path.dirname(_os.path.dirname(_os.path.dirname(_os.path.abspath(__file__)))),
-                f"diag_shot_{self._shot_n:02d}_FINISH.png")
-            self.browser.grab().save(_shot_path)
-            print(f"[画面] 完成快照 -> {_os.path.basename(_shot_path)}", flush=True)
-        except Exception as _e:
-            print(f"[画面] 完成截图失败: {_e}", flush=True)
-        def _finish_dom_probe(_res):
-            diag_log(f"FINISH_DOM {_res}")
-            print(f"[画面] 完成时刻 DOM 体检: {_res}", flush=True)
-        self.browser.page().runJavaScript(f"probeStream('{msg_id}')", _finish_dom_probe)
+        # 完成时刻屏幕快照 + DOM 体检均为诊断操作；尤其是 WebView2 grab()
+        # 会同步读回原生窗口，可能阻塞渲染器数百毫秒，普通运行绝不执行。
+        if self._stream_diagnostics:
+            if os.environ.get("HAOCODE_SHOT") == "1":
+                try:
+                    self._shot_n = getattr(self, "_shot_n", 0) + 1
+                    import os as _os
+                    _shot_path = _os.path.join(
+                        _os.path.dirname(_os.path.dirname(_os.path.dirname(_os.path.abspath(__file__)))),
+                        f"diag_shot_{self._shot_n:02d}_FINISH.png")
+                    self.browser.grab().save(_shot_path)
+                    print(f"[画面] 完成快照 -> {_os.path.basename(_shot_path)}", flush=True)
+                except Exception as _e:
+                    print(f"[画面] 完成截图失败: {_e}", flush=True)
+
+            def _finish_dom_probe(_res):
+                diag_log(f"FINISH_DOM {_res}")
+                print(f"[画面] 完成时刻 DOM 体检: {_res}", flush=True)
+
+            self.browser.page().runJavaScript(
+                f"probeStream('{msg_id}')", _finish_dom_probe)
         # 🚀 入库，自动成为时间线新叶子！
         if (stream_state["content"] or stream_state["reasoning"]
                 or stream_state.get("timeline")):
@@ -5389,7 +5421,7 @@ class MainWindow(QtWidgets.QMainWindow):
             now = _t.time()
             if not hasattr(self, "_last_probe_at"):
                 self._last_probe_at = 0
-            if now - self._last_probe_at >= 2.0:
+            if self._stream_diagnostics and now - self._last_probe_at >= 2.0:
                 self._last_probe_at = now
                 self.browser.page().runJavaScript(
                     f"probeStream('{st['msg_id']}')", self._on_probe_result)
@@ -5438,7 +5470,8 @@ class MainWindow(QtWidgets.QMainWindow):
         """接收到 Token"""
         self._diag_chunk_n += 1
         _match = session_id == self.current_session_id
-        if self._diag_chunk_n <= 3 or self._diag_chunk_n % 50 == 0 or not _match:
+        if self._stream_diagnostics and (
+                self._diag_chunk_n <= 3 or self._diag_chunk_n % 50 == 0 or not _match):
             _st0 = self._active_streams.get(session_id)
             diag_log(f"CHUNK n={self._diag_chunk_n} match={_match} "
                      f"mid={(_st0 or {}).get('msg_id')} +{len(chunk)}c")
@@ -5458,13 +5491,14 @@ class MainWindow(QtWidgets.QMainWindow):
         else:
             st["timeline"][-1]["text"] += chunk
         
-        # 人肉 debug：每个正文 token 控制台打印
-        try:
-            _mid = self._active_streams[session_id]["msg_id"]
-            _flat = " ".join(chunk.split())
-            print(f"[正文] mid={_mid[:8]} 匹配={_match} +{len(chunk)}c {_flat[:100]}", flush=True)
-        except Exception:
-            pass
+        # 人肉 debug：每个正文 token 控制台打印（仅显式诊断时开启）。
+        if self._stream_diagnostics:
+            try:
+                _mid = self._active_streams[session_id]["msg_id"]
+                _flat = " ".join(chunk.split())
+                print(f"[正文] mid={_mid[:8]} 匹配={_match} +{len(chunk)}c {_flat[:100]}", flush=True)
+            except Exception:
+                pass
 
         # 只有当前激活的会话才推送到前端显示
         if session_id == self.current_session_id:
@@ -5825,7 +5859,8 @@ class MainWindow(QtWidgets.QMainWindow):
         """接收到思考过程"""
         self._diag_think_n += 1
         _match = session_id == self.current_session_id
-        if self._diag_think_n <= 3 or self._diag_think_n % 50 == 0 or not _match:
+        if self._stream_diagnostics and (
+                self._diag_think_n <= 3 or self._diag_think_n % 50 == 0 or not _match):
             _st0 = self._active_streams.get(session_id)
             diag_log(f"THINK n={self._diag_think_n} match={_match} "
                      f"mid={(_st0 or {}).get('msg_id')} +{len(chunk)}c")
@@ -5845,13 +5880,14 @@ class MainWindow(QtWidgets.QMainWindow):
         else:
             st["timeline"][-1]["text"] += chunk
         
-        # 人肉 debug：每个思考 token 控制台打印
-        try:
-            _mid = self._active_streams[session_id]["msg_id"]
-            _flat = " ".join(chunk.split())
-            print(f"[思考] mid={_mid[:8]} 匹配={_match} +{len(chunk)}c {_flat[:60]}", flush=True)
-        except Exception:
-            pass
+        # 人肉 debug：每个思考 token 控制台打印（仅显式诊断时开启）。
+        if self._stream_diagnostics:
+            try:
+                _mid = self._active_streams[session_id]["msg_id"]
+                _flat = " ".join(chunk.split())
+                print(f"[思考] mid={_mid[:8]} 匹配={_match} +{len(chunk)}c {_flat[:60]}", flush=True)
+            except Exception:
+                pass
 
         # 只有当前激活的会话才推送到前端显示
         if session_id == self.current_session_id:
@@ -5898,6 +5934,9 @@ class MainWindow(QtWidgets.QMainWindow):
         """JS 就绪检测的回调"""
         if is_ready:
             print("[System]:✅ JS 引擎已就绪，正在加载历史记录...")
+            if self._stream_diagnostics:
+                self.chat_bridge.run_js(
+                    "if (window.setStreamDiagnostics) window.setStreamDiagnostics(true);")
             # 🆕 P1-01：先注入渲染窗口配置，再开始窗口化加载
             if not self._rw_config_pushed:
                 self._rw_config_pushed = True

ui/web/app.js

@@ -31,23 +31,40 @@
     } catch (e) {}
 })();
 window.__APP_VER = '20260721-v7';
+// 流式逐 token 诊断默认关闭。逐 token 读取 DOM 高度/文本并写 console 会
+// 触发强制布局和跨进程日志传输，长时间运行会明显拖慢渲染器；需要排查时
+// 由 Python 显式调用 setStreamDiagnostics(true) 打开。
+window.__STREAM_DIAG = false;
+window.setStreamDiagnostics = function(enabled) {
+    window.__STREAM_DIAG = !!enabled;
+    if (window.__STREAM_DIAG && window.__startStreamHeartbeat) {
+        window.__startStreamHeartbeat();
+    }
+};
 console.log('[JS] ===== app.js 加载 ver=20260721-v7 =====');
 // 公式渲染依赖本地 KaTeX（离线）；此处确认资源加载结果，缺失时打印告警便于定位
 console.log('[JS] KaTeX ' + (typeof katex !== 'undefined' ? katex.version + ' 就绪' : '缺失（公式将退化为纯文本）'));
 // 渲染器主线程心跳：dt 异常大 = 渲染器被阻塞（截图/重绘/GPU 等）
 (function() {
     var _hbLast = Date.now();
-    setInterval(function() {
+    var _started = false;
+    function tick() {
         var _now = Date.now();
         var _dt = _now - _hbLast;
         _hbLast = _now;
         if (_dt >= 1500) {
             console.log('[JS] 心跳 dt=' + _dt + 'ms （渲染器主线程曾卡顿）');
         }
-    }, 1000);
+    }
+    window.__startStreamHeartbeat = function() {
+        if (_started) return;
+        _started = true;
+        setInterval(tick, 1000);
+    };
 })();
 window.__diag = { events: [], cap: 300, tokenN: 0, thinkN: 0 };
 function diagEvent(name, extra) {
+    if (!window.__STREAM_DIAG) return;
     try {
         var d = window.__diag;
         d.events.push({ t: Date.now() % 1000000, e: name, x: extra });
@@ -749,7 +766,8 @@ function createMessage(msgId, role, initialText, senderName, branchInfo) {
         if (role === 'user') {
             replyDiv.innerText = initialText;
         } else {
-            // 🆕 助手纯文本消息（无时间线历史路径）也包进 md-segment → 与时间线路径同样有浅灰气泡背景
+            // 🆕 助手纯文本消息（无时间线历史路径）也包进 md-segment，
+            // 与时间线路径保持相同的透明正文布局
             var _tHtml = typeof marked !== 'undefined' ? safeHtml(marked.parse(initialText)) : safeHtml(initialText);
             replyDiv.innerHTML = '<div class="md-segment markdown-body">' + _tHtml + '</div>';
         }
@@ -1208,7 +1226,8 @@ function syncRenderThrottled(msgId, minGapMs) {
 // Qt 看门狗入口：强制渲染（幂等，尾部无变化时开销极小）
 function forceRenderNow(msgId) {
     var _frNow = Date.now();
-    if (!window.__frLast || _frNow - window.__frLast > 1000) {
+    if (window.__STREAM_DIAG &&
+        (!window.__frLast || _frNow - window.__frLast > 1000)) {
         window.__frLast = _frNow;
         console.log('[JS] forceRenderNow id=' + msgId + ' (看门狗)');
     }
@@ -1236,7 +1255,7 @@ function _fullRenderSegment(el, c, isThink) {
 }
 function doStreamingRender(msgId) {
     window.__renderN = (window.__renderN || 0) + 1;
-    if (window.__renderN % 10 === 1) {
+    if (window.__STREAM_DIAG && window.__renderN % 10 === 1) {
         console.log('[JS] render#' + window.__renderN + ' 开始 id=' + msgId);
     }
     var buf = messageBuffer[msgId];
@@ -1288,9 +1307,11 @@ function appendReasoning(msgId, token) {
     buf.reasoning += token;
     var wrapper = document.getElementById(msgId);
     if (!wrapper) { diagEvent('appendReasoning', 'NO_WRAPPER:' + msgId); return; }
-    window.__diag.thinkN++;
-    if (window.__diag.thinkN === 1 || window.__diag.thinkN % 25 === 0) {
-        diagEvent('appendReasoning', 'n=' + window.__diag.thinkN);
+    if (window.__STREAM_DIAG) {
+        window.__diag.thinkN++;
+        if (window.__diag.thinkN === 1 || window.__diag.thinkN % 25 === 0) {
+            diagEvent('appendReasoning', 'n=' + window.__diag.thinkN);
+        }
     }
     wrapper.classList.add('streaming');
     var tl = ensureTimeline(wrapper, buf);
@@ -1308,7 +1329,11 @@ function appendReasoning(msgId, token) {
         buf.thinkSegs.push(tc);
     }
     tc.__buf = (tc.__buf || '') + token;
-    console.log('[JS] think#' + window.__diag.thinkN + ' 段buf=' + tc.__buf.length + 'c dom=' + (tc.textContent || '').length + 'c h=' + tc.offsetHeight);
+    if (window.__STREAM_DIAG) {
+        console.log('[JS] think#' + window.__diag.thinkN +
+            ' 段buf=' + tc.__buf.length + 'c dom=' +
+            (tc.textContent || '').length + 'c h=' + tc.offsetHeight);
+    }
     syncRenderThrottled(msgId);   // ★ 同步通道
     scheduleStreamingRender(msgId);
 }
@@ -1319,9 +1344,11 @@ function appendToken(msgId, token) {
     buf.content += token;
     var wrapper = document.getElementById(msgId);
     if (!wrapper) { diagEvent('appendToken', 'NO_WRAPPER:' + msgId); return; }
-    window.__diag.tokenN++;
-    if (window.__diag.tokenN === 1 || window.__diag.tokenN % 25 === 0) {
-        diagEvent('appendToken', 'n=' + window.__diag.tokenN);
+    if (window.__STREAM_DIAG) {
+        window.__diag.tokenN++;
+        if (window.__diag.tokenN === 1 || window.__diag.tokenN % 25 === 0) {
+            diagEvent('appendToken', 'n=' + window.__diag.tokenN);
+        }
     }
     wrapper.classList.add('streaming');
     var tl = ensureTimeline(wrapper, buf);
@@ -1345,9 +1372,14 @@ function appendToken(msgId, token) {
     } else {
         seg.__buf = (seg.__buf || '') + token;
     }
-    console.log('[JS] token#' + window.__diag.tokenN + ' 段buf=' + seg.__buf.length + 'c dom=' + (seg.textContent || '').length + 'c 段数=' + buf.textSegs.length + ' h=' + seg.offsetHeight);
-    if (window.__diag.tokenN === 1 || window.__diag.tokenN % 25 === 0) {
-        diagEvent('tokenDOM', { buf: seg.__buf.length, dom: (seg.textContent || '').length });
+    if (window.__STREAM_DIAG) {
+        console.log('[JS] token#' + window.__diag.tokenN +
+            ' 段buf=' + seg.__buf.length + 'c dom=' +
+            (seg.textContent || '').length + 'c 段数=' +
+            buf.textSegs.length + ' h=' + seg.offsetHeight);
+        if (window.__diag.tokenN === 1 || window.__diag.tokenN % 25 === 0) {
+            diagEvent('tokenDOM', { buf: seg.__buf.length, dom: (seg.textContent || '').length });
+        }
     }
     syncRenderThrottled(msgId);   // ★ 同步通道：token 到 → 内容必现
     scheduleStreamingRender(msgId);  // rAF 通道：更平滑（环境允许时）
@@ -1357,7 +1389,12 @@ function appendToken(msgId, token) {
 function finishMessage(msgId) {
     console.log('[JS] finishMessage id=' + msgId);
     var wrapper = document.getElementById(msgId);
-    if (!wrapper) { delete messageBuffer[msgId]; diagEvent('finish', 'NO_WRAPPER:' + msgId); return; }
+    if (!wrapper) {
+        delete messageBuffer[msgId];
+        delete __lastSyncRender[msgId];
+        diagEvent('finish', 'NO_WRAPPER:' + msgId);
+        return;
+    }
     diagEvent('finish', { id: msgId });
     var buf = messageBuffer[msgId];
     cancelStreamingRender(msgId);
@@ -1440,6 +1477,7 @@ function finishMessage(msgId) {
         if (buf.raf) { cancelAnimationFrame(buf.raf); buf.raf = 0; }
     }
     delete messageBuffer[msgId];
+    delete __lastSyncRender[msgId];
 
     // --- H. 刷新自定义滚动条 ---
     reportWebScroll();
@@ -1940,6 +1978,7 @@ function renderTimelineHistory(msgId, timelineJson) {
             // 删除 buffer → finishMessage 不再全量重渲（避免破坏时间线 DOM），
             // 但仍会执行代码高亮/滚动等收尾
             delete messageBuffer[msgId];
+            delete __lastSyncRender[msgId];
         }
     } catch (e) {
         console.error('[JS]: renderTimelineHistory 解析失败', e);
@@ -2063,6 +2102,23 @@ function hideLoadingOverlay() {
     setTimeout(function() { el.style.display = 'none'; }, 480);
 }
 
+// 消息从当前窗口移除时一并释放附件缓存。附件使用 msgId-att-N 作为键，
+// 只删除父消息键会让长文本和元数据跨会话持续留在内存中。
+function clearMessageStores(msgId) {
+    if (!msgId) return;
+    delete messageBuffer[msgId];
+    delete __lastSyncRender[msgId];
+    delete longTextStore[msgId];
+    delete finalContentStore[msgId];
+    delete attachmentMetaStore[msgId];
+    var prefix = msgId + '-att-';
+    [longTextStore, attachmentMetaStore].forEach(function(store) {
+        Object.keys(store).forEach(function(key) {
+            if (key.indexOf(prefix) === 0) delete store[key];
+        });
+    });
+}
+
 // ==================== 历史记录与视图控制 ====================
 function clearChat() {
     var bubbles = chatContainer.querySelectorAll('.message-wrapper, .system-note');
@@ -2073,6 +2129,7 @@ function clearChat() {
     longTextStore = {};
     finalContentStore = {};
     attachmentMetaStore = {};
+    __lastSyncRender = {};
 
     // 🆕 P1-01：清渲染窗口状态机（游标/缓存/未决请求/代次；配置模式与大小保留）
     if (typeof rwState !== 'undefined' && rwState && typeof RenderWindowState !== 'undefined') {
@@ -2093,9 +2150,7 @@ function deleteMessage(msgId) {
         wrapper.style.transform = "translateY(-10px)";
         setTimeout(function() { wrapper.remove(); }, 300);
     }
-    delete messageBuffer[msgId];
-    delete longTextStore[msgId];
-    delete finalContentStore[msgId];
+    clearMessageStores(msgId);
     setTimeout(reportWebScroll, 350);
 }
 
@@ -2267,10 +2322,7 @@ function rwCaptureAnchor() {
 function rwRemoveMessageDom(msgId) {
     var wrapper = document.getElementById(msgId);
     if (wrapper) wrapper.remove();
-    delete messageBuffer[msgId];
-    delete longTextStore[msgId];
-    delete finalContentStore[msgId];
-    delete attachmentMetaStore[msgId];
+    clearMessageStores(msgId);
 }
 
 /* 加载入口：固定在 chat-container 首/尾；无更多消息时隐藏。 */

ui/web/style.css

@@ -157,7 +157,7 @@ body, html {
     width: 85%;
 }
 /* 🆕 内层同步撑满：.reply-content 基类是 fit-content（贴内容缩），
-   可见气泡（md-segment 背景/思考卡/工具卡）都在它里面 → 必须 100% 才真正恒定 */
+   可见内容块（正文段/思考卡/工具卡）都在它里面 → 必须 100% 才真正恒定 */
 .assistant .reply-content { width: 100%; }
 .user .message-content { align-items: flex-end; }
 
@@ -893,15 +893,15 @@ body, html {
 }
 
 
-/* ========== 助手正文气泡（流式中/完成后一致） ========== */
+/* ========== 助手正文（流式中/完成后一致，保持透明） ========== */
 .message-wrapper.assistant .reply-content .md-segment {
     width: 100%;
     box-sizing: border-box;
-    padding: 10px 14px;
+    padding: 0;
     color: #30343b;
-    background-color: #f6f7f9;
-    border: 1px solid #eceff3;
-    border-radius: 8px;
+    background: transparent;
+    border: none;
+    border-radius: 0;
 }
 .message-wrapper.assistant .reply-content .md-segment + .md-segment {
     margin-top: 6px;