feat: post-execution audit trail pattern to complement AgentGovernancePlugin

Now that #141 (AgentGovernancePlugin) is merged, the pre-execution governance loop is closed: policy load → tool-call decision → allow/review/block.

The natural next question is post-execution: once the tool call is allowed and executed, is there a standard pattern for recording what actually happened in a tamper-evident way?

The gap: AgentGovernancePlugin captures the decision. Nothing currently captures the outcome — tool name, args hash, output hash, timestamp — in a form that an external auditor can verify independently.

Proposed pattern: an optional audit_backend parameter on AgentGovernancePlugin (or a companion hook) that receives a structured post-execution record after each allowed tool call. Implementations could range from local JSONL to on-chain anchoring (e.g. Mycelium Trails).

This would complete the loop: pre-execution policy check → execution → post-execution evidence. Both sides independently verifiable.

Happy to contribute a spec or draft implementation if there's interest.

首先阅读 #141 中描述的 AgentGovernancePlugin，并确定其对允许的工具调用所作的决定与执行后处理之间的关系。将提议的 audit_backend 参数或配套 hook 与所需的记录字段进行比较：tool name、args hash、output hash 和 timestamp。当形成了关于可独立验证的执行后证据的共识规范或实现草案时，即视为完成。