feat(observability): 执行可视化 — 节点级实时轨迹（运行·观测）

把任务执行做成可观测：Dispatcher 在每个节点/阶段发结构化 ExecEvent，
经独立 NATS 通道回流，前端逐节点点亮（状态/耗时/工具入参产出）。

- shared: contract.ExecEvent + ExecSubject(sundynix.exec.<id>，与 Token 流分流)；
  bus.PublishExec/CompleteExec/SubscribeExec（core NATS，复用结束头）
- dispatcher: execTracer（自增 Seq 保序 + span 自动计耗时）；
  Orchestrator 加 ExecSink；通用图(init 召回 / 各 tool 入参→产出 / prompt / model
  首token+token数)与报告编排(规划大纲 / 各章并行 start-end / 渲染)全程埋点
- gateway: SubscribeExec + GET /tasks/:id/exec SSE（与 token 流并行）
- desktop: streamExec + deriveNodes(按 node 归并 start/end/error/info)；
  复用组件 ExecTrace（竖向轨道，按 kind 着色，运行中脉冲灯）；
  新 RunsView(运行·观测：轨迹+输出双栏)；BottomDrawer 轨迹/工具调用 tab 接真实数据；
  ReportView 加执行轨迹栏；左导航「运行」置就绪

实测：
- 报告任务 /exec：规划(2680ms,4章) → 4 章并行(seq 交错,各~7-8s 重叠=真并行,
  每章带 docs 知识库检索预览+成稿字数) → 渲染(docx 落盘)
- 通用图 /exec：tool:kb_search(678ms,入参→Milvus 产出) → prompt(2消息) →
  model(首token 860ms / 4 tokens)
- 浏览器(Preview)：报告页执行轨迹逐节点点亮、章节带耗时/字数/检索片段，完成后下载 Word

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

sundynix-desktop/frontend/src/App.tsx

@@ -7,9 +7,10 @@ import { StudioView } from "./studio/StudioView";
 import { MemoryView } from "./views/MemoryView";
 import { KbView } from "./views/KbView";
 import { ReportView } from "./views/ReportView";
+import { RunsView } from "./views/RunsView";
 import { Home } from "./views/Home";
 import { Placeholder } from "./views/Placeholder";
-import { submitTask, streamTokens, type Identity } from "./lib/api";
+import { submitTask, streamTokens, streamExec, type Identity } from "./lib/api";
 import type { TaskDsl } from "./lib/dsl";
 import { emptyRun, type RunState } from "./lib/run";
 
@@ -28,11 +29,14 @@ export default function App() {
   const [run, setRun] = useState<RunState>(emptyRun);
   const closeRef = useRef<(() => void) | null>(null);
 
+  const execCloseRef = useRef<(() => void) | null>(null);
+
   const onRun = useCallback(
     async (dsl: TaskDsl) => {
       closeRef.current?.();
+      execCloseRef.current?.();
       const t0 = Date.now();
-      setRun({ phase: "submitting", output: "", events: [{ t: 0, label: "提交任务" }] });
+      setRun({ phase: "submitting", output: "", events: [{ t: 0, label: "提交任务" }], exec: [] });
       try {
         const taskId = await submitTask(dsl, identity);
         let first = true;
@@ -42,6 +46,13 @@ export default function App() {
           taskId,
           events: [...r.events, { t: Date.now() - t0, label: `已发布 ${taskId}` }],
         }));
+        // 执行轨迹（运行·观测）：与 token 流并行订阅，逐节点点亮。
+        execCloseRef.current = streamExec(
+          taskId,
+          (ev) => setRun((r) => ({ ...r, exec: [...r.exec, ev] })),
+          () => {},
+          () => {},
+        );
         closeRef.current = streamTokens(
           taskId,
           (tok) =>
@@ -84,6 +95,8 @@ export default function App() {
             <KbView />
           ) : view === "report" ? (
             <ReportView identity={identity} />
+          ) : view === "runs" ? (
+            <RunsView run={run} />
           ) : view === "memory" ? (
             <MemoryView identity={identity} />
           ) : (