docs(eval): add reproducible memory evaluation guide

[RerankerGuo]

Description | 描述

Add a lightweight reproducible memory evaluation guide for benchmark runs and future eval integrations.

The guide is framework-neutral and documents:

• required run metadata such as commit, host, model, dataset split, session layout, and resolved config
• short-term, long-term, layer stress, and temporal drift evaluation designs
• L0/L1/L2/L3 layer audit checkpoints
• recommended metrics for accuracy, token use, recall quality, budget use, and latency
• a result template that contributors can reuse when reporting benchmark reproductions

This intentionally avoids adding a full benchmark runner or new runtime dependency.

Related Issue | 关联 Issue

Related to #106

Change Type | 修改类型

• Bug fix | Bug 修复
• New feature | 新功能
• Documentation update | 文档更新
• Code optimization | 代码优化

Self-test Checklist | 自测清单

• Verified locally | 本地验证通过
• No existing features affected | 无影响现有功能

Additional Notes | 其他说明

Verified with npm test and npm run build using Node v24.15.0.

[Maxwell-Code07]

A lightweight evaluation guide is a welcome addition for benchmark reproducibility. The framework-neutral design makes it easy for contributors to report results consistently. We'll take a look.