docs: index agent capability evaluation research

[buxibuxi]

Summary

• add a curated external research index for Agent capability evaluation
• record design boundaries so papers do not become provider or release evidence
• seed the index with the lemons Trust Layer paper plus AgentBench, ToolLLM, and MCP-AgentBench

Validation

• npx --yes markdownlint-cli2 'docs/research/*.md'
• npx --yes markdown-link-check docs/research/agent-capability-evaluation.md

Notes

• raw PDFs are intentionally excluded; entries retain stable public links and identifiers.