[REVIEW] agent-security: bind approvals to artifacts and tool provenance

[xianzuyang9-blip]

Summary

Addresses #1595 for the agent-security skill by adding explicit review gates for:

• approval decisions bound to canonical executable artifacts instead of model-written summaries
• MCP/plugin/tool-server provenance, version pinning, manifest review, and per-server secret scoping
• privacy-preserving audit alternatives such as prompt hashes, redacted parameters, policy traces, and correlation IDs

Verification

• git diff --check
• Confirmed the update is scoped to skills/ai-security/agent-security/SKILL.md