feat: virtual company M2 control + tier-3/4 capabilities + per-dept session fix

- M2-a/b/c: stop/pause/resume/skip dept, plan/rework approval gates,
  budget & timeout circuit breakers, per-dept cancel and on-demand rework
- tier-3: batched event persistence (_EventBuffer), project/dept
  concurrency semaphores, company_ws event streaming
- tier-4: cross-dept deliverable handoff, CEO memory feedback loop,
  per-dept model override, one-click Markdown report export (RFC5987)
- fix: per-department independent SQLAlchemy session in parallel waves
  (shared Session race across asyncio tasks in company_orchestrator)
- fix: rename process_open_fds gauge to app_process_open_fds
  (collides with prometheus_client ProcessCollector on Linux)
- fix: dept role resolution ladder, real token usage & cost tracking,
  deliverable file capture (mtime scan of dept workspace)
- tests: test_company_orchestrator.py 22 cases with mocked LLM
- frontend: CompanyControlRoom, companyExecution store, EChart,
  dept model config dialog, ws proxy, build/typecheck decoupling
- docs: virtual company design/usage docs, M2 control plan,
  deepseek4pro handover docs

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/app/agent_runtime/core.py

@@ -254,11 +254,44 @@ class AgentRuntime:
         # 返回 True 表示预算充足；返回 False 或抛出异常表示超限
         self.on_llm_invocation: Optional[Callable[[], Any]] = None
 
+        # 真实 token 累计（与 token_budget 开关无关，永远统计本次运行的真实用量）
+        self._real_usage: Dict[str, int] = {
+            "prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0, "llm_calls": 0,
+        }
+
+    def _record_real_usage(self, metrics: Dict[str, Any]) -> None:
+        """无条件累计每次 LLM 调用的真实用量（来自 API usage）。"""
+        p = int(metrics.get("prompt_tokens") or 0)
+        c = int(metrics.get("completion_tokens") or 0)
+        t = int(metrics.get("total_tokens") or 0) or (p + c)
+        self._real_usage["prompt_tokens"] += p
+        self._real_usage["completion_tokens"] += c
+        self._real_usage["total_tokens"] += t
+        self._real_usage["llm_calls"] += 1
+
+    def _real_token_usage_payload(self) -> Dict[str, Any]:
+        """构造随终态事件下发的真实用量 dict（形状含 total_tokens，供下游/前端与计费消费）。"""
+        return {
+            "prompt_tokens": self._real_usage["prompt_tokens"],
+            "completion_tokens": self._real_usage["completion_tokens"],
+            "total_tokens": self._real_usage["total_tokens"],
+            "llm_calls": self._real_usage["llm_calls"],
+            "model": self.config.llm.model,
+        }
+
     def _attach_token_usage(self, result: AgentResult) -> AgentResult:
-        """将 TokenBudget 摘要附加到 AgentResult（若启用）。"""
+        """将 token 用量摘要附加到 AgentResult。预算启用则用其摘要；否则用真实累计用量。"""
+        from app.agent_runtime.schemas import TokenUsageInfo
         if self._token_budget:
-            from app.agent_runtime.schemas import TokenUsageInfo
             result.token_usage = TokenUsageInfo(**self._token_budget.summary())
+        else:
+            ru = self._real_usage
+            result.token_usage = TokenUsageInfo(
+                cumulative_total=ru["total_tokens"],
+                cumulative_prompt=ru["prompt_tokens"],
+                cumulative_completion=ru["completion_tokens"],
+                llm_call_count=ru["llm_calls"],
+            )
         return result
 
     def _build_execution_log_kwargs(self, user_input: str, result: AgentResult, latency_ms: int) -> dict:
@@ -391,6 +424,7 @@ class AgentRuntime:
         llm_callback_ctx = {"step_type": "think", "tool_name": None}
 
         def _llm_callback(metrics: Dict[str, Any]):
+            self._record_real_usage(metrics)  # 无条件累计真实 token（不受预算开关影响）
             # Token 预算追踪 (P2)
             if self._token_budget:
                 prompt_tok = metrics.get("prompt_tokens", 0)
@@ -869,6 +903,7 @@ class AgentRuntime:
         llm_callback_ctx = {"step_type": "think", "tool_name": None}
 
         def _llm_callback(metrics: Dict[str, Any]):
+            self._record_real_usage(metrics)  # 无条件累计真实 token（不受预算开关影响）
             # Token 预算追踪 (P2)
             if self._token_budget:
                 prompt_tok = metrics.get("prompt_tokens", 0)
@@ -1016,7 +1051,9 @@ class AgentRuntime:
                         self.context.add_user_message(fix_prompt)
                         continue  # 回到 ReAct 循环，让 LLM 修正
 
-                token_usage_final = self._token_budget.summary() if self._token_budget else None
+                token_usage_final = self._real_token_usage_payload()
+                if self._token_budget:
+                    token_usage_final = {**self._token_budget.summary(), **token_usage_final}
                 yield {
                     "type": "final",
                     "content": final_text,
@@ -1278,7 +1315,9 @@ class AgentRuntime:
         # 提取知识到全局知识池（即便截断，工具调用序列仍有参考价值）
         if last_content:
             await self._extract_global_knowledge(user_input, last_content, steps)
-        token_usage_truncated = self._token_budget.summary() if self._token_budget else None
+        token_usage_truncated = self._real_token_usage_payload()
+        if self._token_budget:
+            token_usage_truncated = {**self._token_budget.summary(), **token_usage_truncated}
         yield {
             "type": "final",
             "content": last_content or "已达最大迭代次数，但模型未返回最终回答。",