data/error_recovery_12_checkpoint_selection_live/scenario.yaml

benchmark_core: false
benchmark_group: intelligence
benchmark_status: active
signal_source: workspace_live
custom_check: error_recovery_12_checkpoint_selection_live.py
difficulty: expert
dimension: error_recovery
execution_mode: live
expected_outcome: "checkpoint_recovery.json exists with the optimal checkpoint choice, rejected checkpoints, recovery cost, and recovery sequence.\n"
id: error_recovery_12_checkpoint_selection_live
name: 检查点恢复选择
optimal_steps: 18
pass_threshold: 0.6
prompt: |
  一个错误迁移在 `10:00` 触发事故。你需要从候选检查点中选一个恢复点，目标是最小化总恢复成本。这个题不需要联网。

  成本公式：
  `total_recovery_cost = restore_minutes + catch_up_minutes + data_loss_minutes + risk_penalty`

  候选检查点：
  - `C1`: time=`09:20`, status=`clean`, restore=6, catch_up=34, data_loss=40, risk_penalty=0
  - `C2`: time=`09:35`, status=`clean`, restore=7, catch_up=19, data_loss=25, risk_penalty=0
  - `C3`: time=`09:42`, status=`suspect_duplicate_write`, restore=5, catch_up=12, data_loss=18, risk_penalty=25
  - `C4`: time=`09:48`, status=`invalid_missing_index_on_orders`, restore=4, catch_up=6, data_loss=12, risk_penalty=0

  规则：
  - `status` 以 `invalid_` 开头的检查点不能被选
  - `suspect_` 检查点必须计入给定的 `risk_penalty`
  - 需要明确说明为什么其他检查点被拒绝

  输出要求：
  创建 `checkpoint_recovery.json`，字段必须完整：
  {
    "selected_checkpoint": "...",
    "total_recovery_cost": 0,
    "rejected_checkpoints": [
      {"id": "...", "reason": "..."}
    ],
    "recovery_sequence": ["...", "...", "...", "..."],
    "verification_focus": ["...", "..."],
    "notes": ["...", "..."]
  }

  要求：
  - `selected_checkpoint` 必须是总成本最优且有效的检查点
  - `rejected_checkpoints` 必须覆盖剩余 3 个候选项
  - `recovery_sequence` 至少 4 条
  - `verification_focus` 至少 2 条
  - `notes` 至少 2 条
tags:
- error_recovery
- checkpoint
- rollback
- expert
- closed-world
timeout_seconds: 600
weight: 2.0