给 2.0 安全叙事补个审计闸门（开源）

AgentScope 2.0 的细粒度权限和 bypass_immune 栅栏在开源框架里少见。我们的审计层与之同路：对 multi-agent 报告输出做覆盖率算术、分簇加总验算、引用核验，MCP 即接：github.com/chenhz01/adversarial-research-audit。想看回一句；不相关请忽略。

The issue only requests adding an audit gate and links MCP/`github.com/chenhz01/adversarial-research-audit`, with no code file, test, or exact acceptance criteria given. First inspect the repository for where multi-agent report output is produced and any existing permission or security/audit hook points. Then compare that flow with the referenced audit approach (coverage arithmetic, cluster-sum checking, citation validation) to identify a concrete insertion point for the gate. Done means the issue defines a concrete flow and validation points for accepted/rejected reports.

由索引模型根據 Issue 內容生成。