Summary

This PR makes OpenViking auto-capture non-blocking in the OpenClaw context-engine path.

Previously, afterTurn awaited the full auto-capture pipeline (createSession -> addSessionMessage -> extractSessionMemories -> deleteSession), which could delay final message delivery in chat
channels when extraction was slow.

This change moves auto-capture to an in-memory background queue keyed by session, so user-facing reply completion is no longer blocked by capture latency.

What changed

• examples/openclaw-plugin/context-engine.ts
  • Added sessionKey?: string to afterTurn params typing.
  • Extracted existing capture logic into runAutoCapture(...) (behavior/logging kept intact).
  • Added enqueueAutoCapture(...) with per-session sequential queue:
    • Same session: serialized capture jobs
    • Different sessions: can proceed independently
  • Updated afterTurn(...) to enqueue and return immediately (no synchronous wait on extraction).

Why

In real traffic, slow capture/extraction could create a visible gap between model completion and final message/card completion.
By decoupling capture from the synchronous afterTurn return path, we preserve memory capture while removing user-visible blocking.

Behavior impact

• User-facing latency: improved (final reply is no longer gated by auto-capture extraction latency)
• Capture semantics: preserved
• Logging: preserved (capture-check, auto-captured, capture-detail, error/warn logs)
• Safety: per-session queue avoids concurrent capture contention for the same session

Validation

Manual log-based validation in Feishu/OpenClaw integration:

• Before: agent end could be followed by several seconds of openviking auto-capture activity before final deliver called
• After: deliver called / finalization proceeds immediately; auto-capture runs in background and logs later if needed

Risks / trade-offs

• In-flight queued capture jobs are in-memory and may be lost on abrupt process crash/restart
• This PR intentionally prioritizes user-facing response latency and does not introduce persistent job infrastructure

中文说明（简版）

1) 问题背景

当前 afterTurn 会同步等待 auto-capture 全流程（创建会话、写入消息、提取记忆、删除会话）。当提取较慢时，会拖慢用户可见的最终回包/卡片完成时机，出现“模型结束后还要等几秒”的体验问题。

2) 本次改动

将 auto-capture 从同步链路改为后台队列执行：afterTurn 只负责入队并立即返回；实际提取逻辑在后台运行。队列按 sessionKey/sessionId 串行，确保同会话内不并发冲突；原有日志与捕获逻辑保持不变。

3) 效果与取舍

效果是显著降低用户侧感知延迟，最终回包不再被 auto-capture 阻塞。取舍是队列为内存态，进程异常退出时未完成任务可能丢失；本 PR 优先保障在线响应时延，不引入更重的持久化任务系统。