feat(examples): add evaluation optimization closed-loop example

[YAO-001]

English

This PR implements issue #91 with a reproducible Evaluation + Optimization closed-loop example.

Highlights:

• Adds examples/optimization/eval_optimize_loop with train/validation evalsets, baseline prompt, optimizer config, README, DESIGN, and example reports.
• Provides a deterministic fake mode that runs without API keys and covers baseline train/validation evaluation, failure attribution, candidate optimization, validation regression, per-case delta, configurable gate, and audit artifacts.
• Rejects overfit candidates that improve train results but regress validation/protected cases.
• Adds SDK mode using real AgentOptimizer and TargetPrompt integration. SDK mode maps OptimizeResult aggregate metrics, cost, duration, token usage, best prompts, and round summaries into the same report.
• Keeps SDK wrapper gate config separate from SDK OptimizeConfigFile via --gate-config.
• Supports multiple TargetPrompt fields without requiring system_prompt.
• Writes strict JSON and append-only audit artifacts with input hashes, prompt hashes, per-field prompt snapshots, diffs, case results, gate reasons, and reproducibility command.
• Source prompt write-back is disabled by default; --update-source is required.
• Adds tests for fake hidden-sample generalization, gate rejection paths, failure attribution, report schema, SDK adapter wiring, SDK aggregate gate, audit safety, run-id validation, strict JSON, and non-finite numeric rejection.

Validation:

• python -m compileall examples/optimization/eval_optimize_loop
• python -m pytest examples/optimization/eval_optimize_loop/tests
• python examples/optimization/eval_optimize_loop/run_pipeline.py --mode fake --trace --output-dir /tmp/eval-optimize-loop-fake
• python examples/optimization/eval_optimize_loop/run_pipeline.py --fake-model --fake-judge --trace --output-dir /tmp/eval-optimize-loop-legacy

中文

本 PR 实现 issue #91，新增一个可复现的 Evaluation + Optimization 闭环示例。

重点：

• 新增 examples/optimization/eval_optimize_loop，包含 train/validation evalsets、baseline prompt、optimizer config、README、DESIGN 和示例报告。
• 提供确定性的 fake mode，无需 API key 即可运行完整流程，覆盖 baseline train/validation 评测、失败归因、候选优化、验证集退化、per-case delta、可配置 gate 和审计产物。
• gate 会拒绝 train 提升但 validation/protected case 退化的过拟合候选。
• 新增 SDK mode，接入真实 AgentOptimizer 和 TargetPrompt。SDK mode 将 OptimizeResult 的聚合指标、成本、耗时、token usage、best prompts 和 round summary 映射到同一份审计报告中。
• SDK wrapper gate config 通过 --gate-config 与 SDK OptimizeConfigFile 解耦。
• 支持多个 TargetPrompt 字段，不强制要求 system_prompt。
• 输出严格 JSON，并写入 append-only 审计产物，包括输入哈希、prompt 哈希、分字段 prompt 快照、diff、case results、gate reasons 和 reproducibility command。
• 默认不回写源 prompt；必须显式传入 --update-source。
• 测试覆盖 fake hidden-sample 泛化、gate 拒绝路径、失败归因、报告 schema、SDK adapter wiring、SDK aggregate gate、审计安全、run-id 校验、严格 JSON 和非有限数值拒绝。

验证：

• python -m compileall examples/optimization/eval_optimize_loop
• python -m pytest examples/optimization/eval_optimize_loop/tests
• python examples/optimization/eval_optimize_loop/run_pipeline.py --mode fake --trace --output-dir /tmp/eval-optimize-loop-fake
• python examples/optimization/eval_optimize_loop/run_pipeline.py --fake-model --fake-judge --trace --output-dir /tmp/eval-optimize-loop-legacy

[github-actions]

CLA Assistant Lite bot All contributors have signed the CLA ✍️ ✅

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@73655ab). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             main        #119   +/-   ##
==========================================
  Coverage        ?   87.51506%           
==========================================
  Files           ?         467           
  Lines           ?       44005           
  Branches        ?           0           
==========================================
  Hits            ?       38511           
  Misses          ?        5494           
  Partials        ?           0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[YAO-001]

I have read the CLA Document and I hereby sign the CLA