feat(examples): add reproducible Evaluation + Optimization pipeline

[Adonis-a233]

Summary

Implements the reproducible Evaluation + Optimization closed loop for the issue
"设计并实现可复现的 Evaluation + Optimization pipeline".

Resolves #91

What Changed

Adds examples/optimization/eval_optimize_loop/ — a six-stage pipeline around
AgentEvaluator / AgentOptimizer:

1. Baseline evaluation — train and validation sets scored separately; per-case
   metric sub-scores (final_response / tool_trajectory / rubric), pass/fail,
   failure reasons, and key trace fields.
2. Failure attribution — rule-based clustering over structured trajectories into
   final_response_mismatch / tool_call_error / parameter_error /
   llm_rubric_not_met / knowledge_recall_insufficient / format_error.
   case_meta.json declares an expected category per case and the report carries an
   attribution accuracy self-check (4/4 = 100% on the bundled sample).
3. Optimization — fake mode applies a deterministic scripted candidate; live mode
   runs a real GEPA search via AgentOptimizer.optimize + TargetPrompt.add_path.
4. Candidate validation — full re-run and per-case diff vs baseline
   (new_pass / new_fail / score_up / score_down).
5. Acceptance gate — five independent configurable checks: validation gain
   threshold, no new hard fail, no key-case regression, no train-up/val-down
   overfit, cost within budget (optimizer spend + token-estimated evaluation spend).
6. Audit persistence — append-only runs/<timestamp>_<run_id>/ per run with
   prompt snapshots, JSON/Markdown reports, gate reasons, cost/token split,
   duration, GEPA seed, prompt SHA-256, and a full config snapshot; run_id is
   injected into every log line for cross-artifact tracing.

Robustness and engineering:

• Live agent calls retry with exponential backoff + jitter and a per-call timeout
  (EVAL_OPT_CALL_TIMEOUT / EVAL_OPT_CALL_ATTEMPTS / EVAL_OPT_CALL_BACKOFF).
• optimizer.json is validated at startup (metric weights must sum to 1.0, all
  gate keys present) with readable errors instead of bare KeyErrors.
• 33 IO-free unit tests under tests/ cover attribution, rubric scoring, every
  gate check, case diffing, the self-check, and config validation.
• Generated reports are gitignored (running the example never dirties the tree);
  frozen samples are committed under sample_output/ (.json + .md).
• The 6 sample cases cover all three required situations: optimizable success,
  optimization-ineffective, and post-optimization regression. README includes the
  requested design note (failure attribution, gating, overfit protection, audit).

Validation

Fake mode (no API key, ~1s, deterministic):

python examples/optimization/eval_optimize_loop/run.py --mode fake

[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@8080800). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             main         #99   +/-   ##
==========================================
  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.

[Adonis-a233]

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