feat: 圆桌第3步大屏多页改用专职智能体 + 健壮 report-json 抽取

修复「大屏多页抽取 json 又失败」：旧实现复用画 HTML 的 roundtable-report
智能体 + rfind 围栏 + 单次 json.loads 的简单粗暴抽取，弱模型跑偏/格式偏差即失败。

- 新内置专职单例 roundtable-dashboard「大屏多页数据智能体」：SOUL 专讲逐席位
  分析 → report-json 契约，禁工具/禁文件/禁 HTML；新 run policy dashboard 禁用一切工具
- harness 新增 report_json.py：多策略抽取（report-json/json/裸 JSON/截断）+ schema
  归一化（移植前端 extractReportData/coerceReportData），返回可直接渲染的归一化 JSON
- 后台 run_dashboard_data 改用专职 agent + dashboard policy + 健壮抽取
- 新增 /api/multi-agent/dashboard/init；前端 initDashboardThread + useStep3Dashboard 改用之

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

frontend-web/src/roundtable-planning/api/multi-agent.ts

@@ -835,6 +835,32 @@ export async function initReportThread(
   return { agentName: json.agent_name, threadId: json.thread_id };
 }
 
+/**
+ * Create a thread for the Step 3「大屏多页」dashboard-data agent
+ * (`roundtable-dashboard`, a built-in singleton — distinct from `roundtable-report`;
+ * its run policy disables ALL tools so it only emits a ```report-json block). Run it
+ * with {@link streamMultiAgent} (`agentType: "special"`, `agentName: "roundtable-dashboard"`,
+ * `dAgentThreadId: { [agentName]: threadId }`); the frontend then `extractReportData`s
+ * the streamed text. Follow-up Q&A / refinements reuse the SAME thread (do not re-init)
+ * so the agent keeps the prior dashboard context.
+ */
+export async function initDashboardThread(
+  options: { signal?: AbortSignal } = {},
+): Promise<ReportThreadResponse> {
+  const res = await apiFetch(`${getBackendBaseURL()}/api/multi-agent/dashboard/init`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: "{}",
+    signal: options.signal,
+  });
+  if (!res.ok) {
+    const text = await res.text().catch(() => "");
+    throw new Error(`init dashboard thread failed (${res.status}): ${text}`);
+  }
+  const json = (await res.json()) as { agent_name: string; thread_id: string };
+  return { agentName: json.agent_name, threadId: json.thread_id };
+}
+
 /**
  * Create a thread for the Step 3「总结报告」summary agent (`roundtable-summary`,
  * a built-in singleton — distinct from `roundtable-report`; this only spins up a

frontend-web/src/roundtable-planning/hooks/useStep3Dashboard.ts

@@ -15,7 +15,7 @@ import { useCallback, useEffect, useRef, useState } from "react";
 import { useArtifacts } from "@/components/workspace/artifacts";
 import type { DraftStep3Snapshot } from "@/roundtable-planning/api/drafts";
 import {
-  initReportThread,
+  initDashboardThread,
   streamMultiAgent,
   type StreamPhase,
 } from "@/roundtable-planning/api/multi-agent";
@@ -35,7 +35,7 @@ import {
 } from "@/roundtable-planning/lib/report-template";
 import { nowTime } from "@/roundtable-planning/lib/roundtable-constants";
 
-const DASH_AGENT_FALLBACK = "roundtable-report";
+const DASH_AGENT_FALLBACK = "roundtable-dashboard";
 
 export type DashboardPhase = "idle" | "creating" | "streaming" | "done" | "error";
 
@@ -269,7 +269,7 @@ export function useStep3Dashboard(options: UseStep3DashboardOptions): UseStep3Da
       const { bubbleId, message, aborter } = args;
       let threadId = liveThreadRef.current;
       if (!threadId) {
-        const init = await initReportThread({ signal: aborter.signal });
+        const init = await initDashboardThread({ signal: aborter.signal });
         if (aborter.signal.aborted) return;
         threadId = init.threadId;
         agentNameRef.current = init.agentName || DASH_AGENT_FALLBACK;