fix(swarm): reap orphan agents on terminal run + retry backoff + surface task errors

孤儿 agent 事故修复(2026-06-15 swarm-69e470561bd7-seed 跨 workspace 执行)。
不改去中心化认领逻辑(swarm_dispatch 不动),仅治理 agent pod 生命周期与可观测性:

- B1 终态回收: refresh_swarm_run_status 终态后 best-effort stop_launched 回收 pod+secret
  (此前仅 Manager stop 才回收,自然完成/失败的 run agent 残留 → 孤儿留在共享池抢别的 swarm 任务)
- B2 防驱逐: agent pod 加 karpenter.sh/do-not-disrupt(AGENT_POD_ALLOW_DISRUPTION=1 可关)
- stop_launched: backend=kubernetes 时按标签删,不再被内存集合 _k8s_swarms 门控(跨重启可靠)
- C 重试退避: Task.next_retry_at + fail_task 指数退避 5→30→180s(cap 300, TASK_RETRY_BACKOFF_*),
  is_task_ready 门控;TASK_MAX_RETRIES 可配
- D 错误上浮: task_executor 失败时聚合 subtask error 到顶层 error(根治通用 "Task failed"),
  _execute_subtask 打印 LLM 响应片段
- 配额硬上限 16: _clamp_user_cap

契约测试全过: runtime-contract / merge-smoke / workflow-e2e / contract-freeze /
max-agents-per-user / security-boundary。

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

agent/task_executor.py

@@ -117,7 +117,7 @@ class TaskExecutor:
 
             success = all(r["status"] in ["completed", "handed_off"] for r in results)
             awaiting_handoff = any(r["status"] == "handed_off" for r in results)
-            return {
+            payload = {
                 "success": success,
                 "task_id": task_id,
                 "subtasks": results,
@@ -125,6 +125,19 @@ class TaskExecutor:
                 "agent_id": self.agent_id,
                 "usage": self._usage_payload(time.time() - started_at),
             }
+            if not success:
+                # Surface the real per-subtask failure detail as a top-level `error` so the
+                # orchestrator records WHY (agent/main.py falls back to a generic "Task failed"
+                # when this key is absent — see incident 2026-06-15 swarm-69e470561bd7-seed).
+                failed = [r for r in results if r.get("status") not in ("completed", "handed_off")]
+                detail = "; ".join(
+                    f"{((r.get('subtask') or {}).get('description') or 'subtask')[:80]}: "
+                    f"{r.get('error') or r.get('summary') or 'no detail reported'}"
+                    for r in failed
+                ) or "subtask(s) failed without detail"
+                payload["error"] = detail
+                logger.error(f"Task {task_id} failed: {detail}")
+            return payload
         except Exception as e:
             logger.error(f"Error executing task {task_id}: {e}")
             return {
@@ -166,6 +179,7 @@ Return ONLY the JSON array, no other text."""
             }]
 
     async def _execute_subtask(self, subtask: dict, task_id: str, context: dict, peer_collaboration_callback: Optional[Callable]) -> dict:
+        content = ""
         try:
             description = subtask["description"]
             workspace_files = self._summarize_workspace()
@@ -270,10 +284,26 @@ Return ONLY the JSON, no other text."""
                 if apply_result["errors"]:
                     result["status"] = "failed"
                     result["error"] = "; ".join(apply_result["errors"])
+                    logger.warning(
+                        f"Subtask for task {task_id} failed applying file changes: {result['error']}"
+                    )
+            else:
+                # LLM returned 2xx but declared the subtask not completed: log its reason and a
+                # bounded snippet of the model output so the failure is diagnosable from agent logs.
+                logger.warning(
+                    "Subtask for task %s reported status=%r (error=%s summary=%s); llm_response[:500]=%s",
+                    task_id, result.get("status"), result.get("error"),
+                    result.get("summary"), (content or "").strip()[:500],
+                )
             result["subtask"] = subtask
             return result
         except Exception as e:
-            logger.error(f"Error executing subtask: {e}")
+            # Parse failures / API errors land here; include a bounded model-output snippet
+            # (model-generated text, no secrets) to explain why parsing/execution failed.
+            logger.error(
+                f"Error executing subtask for task {task_id}: {e}; "
+                f"llm_response[:500]={(content or '').strip()[:500]}"
+            )
             return {
                 "subtask": subtask,
                 "status": "failed",