[codex] add agent eval failure packet

[yangfei222666-9]

Summary

Adds a concrete, machine-checkable agent eval data packet to self-improving-loop:

• docs/ANNOTATION_GUIDELINE.md defines failure labels, hard vs soft signals, and routing policy.
• docs/AI_CODING_TOOL_FAILURE_NOTES.md maps Claude Code / Cursor / OpenClaw / Hermes failure modes into labels and guard actions.
• examples/agent_eval_cases.jsonl adds 30 non-authorizing eval cases.
• examples/verify_agent_eval_cases.py verifies schema, labels, unique case IDs, and hard no-execution flags.
• tests/test_agent_eval_cases.py locks the packet boundary in CI.

Why

This turns the project from only a regression-rollback runtime into a clearer reliability data asset: traces can now be mapped into failure labels, eval cases, and conservative routing without granting execution authority.

Safety boundary

The packet is eval data only. The verifier enforces:

judgment_allowed=false
paper_buy_allowed=false
trade_allowed=false
promote_allowed=false

Validation

python3 examples/verify_agent_eval_cases.py examples/agent_eval_cases.jsonl
python3 -m py_compile examples/verify_agent_eval_cases.py
python3 -m pytest -q tests/test_agent_eval_cases.py tests/test_examples.py
python3 -m pytest -q

Results:

verdict=ok
case_count=30
10 passed
66 passed