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

[YAO-001]

English

Closes #91.

This PR adds a reproducible Evaluation + Optimization closed loop under examples/optimization/eval_optimize_loop.

What changed

• Runs baseline evaluation on the complete train and validation EvalSets, preserving per-case metrics, pass/fail, failure evidence, and trace availability.
• Classifies failures across response, format, tool, parameter, rubric, and knowledge-recall categories.
• Supports deterministic offline fake mode and a real SDK path using AgentOptimizer, TargetPrompt, and post-optimization AgentEvaluator runs.
• Deduplicates every round candidate plus best_prompts, then fully reevaluates each candidate on train and validation. SDK optimizer aggregates remain audit context, not gate evidence.
• Applies one complete gate in both modes: validation improvement, overfit rejection, new hard failures, protected-case regressions, per-case score drops, and fail-closed cost budgets.
• Keeps AgentOptimizer(update_source=False) internally. --update-source commits only the selected candidate after full gate acceptance, input-integrity checks, and audit preparation, with transactional rollback/conflict reporting.
• Publishes each run once to an immutable run-specific directory using atomic no-replace semantics, with SHA-256/size manifest, input hashes, rounds, case results, deltas, gate decisions, costs, durations, and reproducibility command.
• Recursively redacts credential-shaped optimizer keys, including separator/camelCase token variants, from all public audit snapshots while retaining safe fields such as max_tokens and token_budget.
• Includes 3 train + 3 validation public cases, stable example JSON/Markdown reports, a 300–500 Chinese-character design note, independent hidden-style gate/attribution fixtures, and real-SDK integration coverage without live API calls.

Validation

• python -m pytest examples/optimization/eval_optimize_loop/tests --tb=short — 349 passed, 2 Windows symlink-capability skips
• python -m pytest examples/optimization/eval_optimize_loop/tests/test_acceptance_thresholds.py examples/optimization/eval_optimize_loop/tests/test_sdk_integration.py -q — 7 passed
• python -m yapf --diff -r examples/optimization/eval_optimize_loop
• python -m flake8 examples/optimization/eval_optimize_loop
• changed trpc_agent_sdk/ YAPF + flake8 checks
• python -m build — built trpc_agent_py-1.1.12 sdist and wheel; isolated wheel import succeeded
• Python 3.12 repository run reached 87.35% coverage. The local Windows-only failures are the repository's existing POSIX/Bash/permission/symlink cases; the PR's Linux CI is the merge authority.

中文

本 PR 完成 Issue #91：新增可复现的 Evaluation + Optimization 自动闭环。fake 与 SDK 模式统一执行 baseline、失败归因、候选 train/validation 完整复评、逐 case delta、完整 gate 和审计落盘；SDK 内部始终禁止提前回写，只有候选通过全部门禁与完整性检查后，流水线才可事务性更新源 prompt。运行目录采用不可覆盖的原子发布，记录哈希清单、成本、耗时、轮次、评测结果和复现命令；配置快照会递归脱敏常见及 camelCase 密钥。独立验收测试覆盖 ≥80% gate 决策准确率、≥75% 归因准确率、过拟合拒绝和三分钟性能上限。

[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@e113610). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             main        #119   +/-   ##
==========================================
  Coverage        ?   87.56402%           
==========================================
  Files           ?         467           
  Lines           ?       44130           
  Branches        ?           0           
==========================================
  Hits            ?       38642           
  Misses          ?        5488           
  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