feat: add model benchmark system for multi-LLM stock prediction validation

  New module: src/services/model_benchmark.py (1062 lines)

  Provides automated comparison of LLM models for US stock prediction:
  - DISCOVER: reads all configured models from llm_model_list (LLM_CHANNELS
    + legacy env vars), no manual enumeration needed
  - ANALYZE: runs identical stock analysis through each model, stores
    results tagged with model_id in analysis_history.context_snapshot
  - EVALUATE: reuses existing BacktestEngine to backtest predictions
    against actual forward prices after configurable eval window
  - REPORT: ranks models by composite score (direction accuracy 35%,
    win rate 30%, excess return 20%, consistency 10%, conviction 5%)

  CLI: python -m src.services.model_benchmark [--analyze|--evaluate|--report|--full]

  Design:
  - Zero schema changes — model_id stored in existing context_snapshot JSON
  - _ModelOverride context manager temporarily sets LITELLM_MODEL per run
  - Deduplication via get_configured_llm_models() from authoritative config
  - Per-stock breakdown in report for granular accuracy analysis

  Docs:
  - docs/design-model-benchmark.html — full architecture, scoring formula,
    data model, CLI reference, edge cases, limitations & future work
  - docs/LINUX_DEPLOYMENT.html — added benchmark CLI section to quick reference

  Governance:
  - AGENTS.md: new hard rule — all new features require "三件套" (design doc,
    deploy doc update, commit comments) to be considered complete

  Verified: import chain, CLI --help, --list-models, text + JSON report
  formatting with synthetic data.

Author: aibot-hub

AGENTS.md

@@ -24,6 +24,10 @@
 - **新增用户自维护的技术文档**（部署指南、配置指南、专题说明书等）优先使用 HTML 格式：语义化 HTML5 + 内嵌 `<style>`，不使用外部 CSS/JS 依赖。HTML 提供一致的跨平台渲染、原生 `<details>` 折叠、CSS 提示框组件、更明确的 DOM 结构便于 AI 工具解析与 diff 审核。上游 fork 同步的 `.md` 文档保持原格式不动，不转换。
 - 变更中英双语文档之一时，需评估另一份是否需要同步；若未同步，交付说明里要写明原因。
 - 注释、docstring、日志文案以清晰准确为准，不强制要求英文，但应与文件语境保持一致。
+- **新增功能/模块/子系统时，必须同步产出"三件套"方可视为交付完成**：
+  1. **设计文档** — `docs/design-<name>.html`，说明架构、数据流、算法/打分公式、模块交互、边界条件、配置项语义；
+  2. **部署/运维文档更新** — 如涉及新 CLI 入口、新配置项、新环境变量、新数据表或定时任务，必须同步更新对应的部署指南和配置指南；
+  3. **英文 commit comments** — 供 maintainer 参考，包含 what/why/验证情况/风险点，遵循 `AGENTS.md` 交付结构。
 
 ## 2. AI 协作资产治理
 

docs/LINUX_DEPLOYMENT.html

@@ -632,11 +632,30 @@ <h3>常用命令快速参考</h3>
 systemctl restart stock-analyzer                # Systemd</code></pre>
 
 <!-- ═══════════════════════════════════════════════════════════════ -->
+<h2>模型基准测试（Model Benchmark）</h2>
+
+<p>系统支持多模型对比验证，用于评估不同 LLM 对美股预测的准确度：</p>
+
+<pre><code># 发现所有已配置的模型
+python -m src.services.model_benchmark --list-models
+
+# 对指定美股运行多模型分析
+python -m src.services.model_benchmark --analyze --stocks AAPL,NVDA,MSFT
+
+# 等待评估窗口后运行回测
+python -m src.services.model_benchmark --evaluate --days 5
+
+# 生成对比排行榜
+python -m src.services.model_benchmark --report</code></pre>
+
+<p>详见 <a href="design-model-benchmark.html">Model Benchmark 设计文档</a>。</p>
+
 <h2>相关文档</h2>
 <ul>
   <li><a href="DEPLOY.md">部署指南</a> - 详细的部署说明</li>
   <li><a href="full-guide.md">完整配置指南</a> - 高级配置选项</li>
   <li><a href="desktop-package.md">桌面端打包说明</a> - 桌面应用部署</li>
+  <li><a href="design-model-benchmark.html">Model Benchmark 设计文档</a> - 多模型对比验证</li>
   <li><a href="FAQ.md">FAQ</a> - 常见问题解答</li>
 </ul>