feat(PowerSystems): add SecurityConstrainedDispatch benchmark

benchmarks/PowerSystems/README.md

@@ -9,3 +9,8 @@ Current tasks emphasize realistic operational constraints, economic objectives,
   - Unified benchmark: `task=unified task.benchmark=PowerSystems/EV2GymSmartCharging`
   - Quick run: `python -m frontier_eval task=unified task.benchmark=PowerSystems/EV2GymSmartCharging task.runtime.env_name=frontier-eval-driver algorithm.iterations=0`
   - Description: upstream-aligned EV smart charging with transformer constraints in the real `EV2Gym` simulator
+
+- `SecurityConstrainedDispatch`
+  - Unified benchmark: `task=unified task.benchmark=PowerSystems/SecurityConstrainedDispatch`
+  - Runtime: install the pinned packages in `verification/requirements.txt` and set `task.runtime.python_path` to that environment
+  - Description: cost-and-security dispatch across normal and N-1 PGLib operating scenarios with frozen AC power-flow verification

benchmarks/PowerSystems/README_zh-CN.md

@@ -9,3 +9,8 @@
   - `frontier_eval` 任务：`task=unified task.benchmark=PowerSystems/EV2GymSmartCharging`
   - 快速运行：`python -m frontier_eval task=unified task.benchmark=PowerSystems/EV2GymSmartCharging task.runtime.env_name=frontier-eval-driver algorithm.iterations=0`
   - 简介：在真实上游 `EV2Gym` 模拟器中进行、与上游数据对齐的 EV 智能充电与变压器约束优化
+
+- `SecurityConstrainedDispatch`
+  - `frontier_eval` 任务：`task=unified task.benchmark=PowerSystems/SecurityConstrainedDispatch`
+  - 运行环境：安装 `verification/requirements.txt` 中的固定依赖，并通过 `task.runtime.python_path` 指定解释器
+  - 简介：在正常与 N-1 PGLib 场景中优化成本和安全裕度，并由冻结交流潮流验证器评分

benchmarks/PowerSystems/SecurityConstrainedDispatch/README.md

@@ -0,0 +1,92 @@
+# Security-Constrained Dispatch
+
+This benchmark asks an agent to improve generator active-power and voltage
+setpoints across normal and N-1 operating scenarios derived from PGLib-OPF.
+Every candidate is checked by an independent AC power-flow verifier.
+
+## Benchmark ID
+
+```text
+PowerSystems/SecurityConstrainedDispatch
+```
+
+## Data provenance
+
+The bundled MATPOWER cases come from PGLib-OPF release `v23.07`, commit
+`dc6be4b2f85ca0e776952ec22cbd4c22396ea5a3`:
+
+- `pglib_opf_case24_ieee_rts.m`
+- `pglib_opf_case57_ieee.m`
+- `pglib_opf_case73_ieee_rts.m`
+
+PGLib-OPF is maintained by the IEEE PES Task Force on Benchmarks for Validation
+of Emerging Power System Algorithms. See `references/pglib/LICENSE` and
+`references/PROVENANCE.md`.
+
+## Environment
+
+Use Python 3.12 on Linux. Install the pinned dependencies:
+
+```bash
+python3.12 -m venv .venvs/frontier-pglib-scd
+.venvs/frontier-pglib-scd/bin/python -m pip install \
+  -r benchmarks/PowerSystems/SecurityConstrainedDispatch/verification/requirements.txt
+```
+
+The evaluator is CPU-only and requires Linux seccomp support. Candidate code is
+run in a restricted worker process; Docker and Linux namespaces are not used.
+The worker blocks file access, process creation, and networking from `solve()`.
+Because namespace isolation is unavailable on the target runtime, this is not
+claimed to be container-equivalent isolation. Evaluation fails closed if the
+seccomp filter cannot be installed.
+
+A typical baseline evaluation of all 12 scenarios uses less than 2 GB RAM and
+should complete within seconds on a modern CPU. Candidate import has a separate
+10-second infrastructure timeout. Each `solve()` call retains its full
+contract-defined two-second budget.
+
+## Direct evaluation
+
+From the benchmark directory:
+
+```bash
+../../../.venvs/frontier-pglib-scd/bin/python \
+  verification/evaluator.py scripts/init.py \
+  --metrics-out metrics.json \
+  --artifacts-out artifacts.json
+```
+
+## Frontier Eval
+
+From the repository root, use the repository-owned task runtime:
+
+```bash
+python -m frontier_eval \
+  task=unified \
+  task.benchmark=PowerSystems/SecurityConstrainedDispatch \
+  task.runtime.python_path=uv-env:frontier-pglib-scd \
+  algorithm=openevolve \
+  algorithm.iterations=0
+```
+
+Baseline-only validation does not require a model API key.
+
+## Isolation tests
+
+From the benchmark directory:
+
+```bash
+../../../.venvs/frontier-pglib-scd/bin/python -m pytest -q \
+  verification/test_candidate_isolation.py
+```
+
+## Layout
+
+- `scripts/init.py`: editable dispatch policy.
+- `verification/`: frozen AC power-flow evaluator.
+- `references/pglib/`: version-pinned PGLib source cases and license.
+- `references/scenarios.json`: deterministic scenario baselines and reference costs.
+- `references/build_scenarios.py`: reproducible scenario-generation script.
+- `frontier_eval/`: unified-task metadata.
+
+See `Task.md` for the complete contract and scoring definition.

benchmarks/PowerSystems/SecurityConstrainedDispatch/README_zh-CN.md

@@ -0,0 +1,75 @@
+# 安全约束电力调度
+
+本任务要求 Agent 改进发电机有功功率和电压设定值，在由 PGLib-OPF
+构建的正常及 N-1 故障场景中降低发电成本。冻结验证器使用独立交流潮流
+检查每个候选方案。
+
+## Benchmark ID
+
+```text
+PowerSystems/SecurityConstrainedDispatch
+```
+
+## 数据来源
+
+任务内置的 MATPOWER 案例来自 PGLib-OPF `v23.07`，提交：
+`dc6be4b2f85ca0e776952ec22cbd4c22396ea5a3`。
+
+- `pglib_opf_case24_ieee_rts.m`
+- `pglib_opf_case57_ieee.m`
+- `pglib_opf_case73_ieee_rts.m`
+
+许可证和文件哈希见 `references/pglib/LICENSE` 与
+`references/PROVENANCE.md`。
+
+## 环境
+
+推荐 Linux 与 Python 3.12。本任务仅使用 CPU：
+
+```bash
+python -m venv .venv-pglib-scd
+source .venv-pglib-scd/bin/activate
+python -m pip install -r benchmarks/PowerSystems/SecurityConstrainedDispatch/verification/requirements.txt
+```
+
+评测器仅使用 CPU，并要求 Linux seccomp 支持。候选程序在受限 worker
+进程中运行，不使用 Docker 或 Linux namespace。worker 会禁止 `solve()`
+读取或写入文件、创建进程和访问网络；如果 seccomp 过滤器无法安装，评测将
+直接失败，不会降级为无保护执行。由于运行环境不提供 namespace 隔离，本方案
+不宣称具备与容器完全等价的安全边界。
+
+候选模块加载使用独立的 10 秒基础设施超时；每次 `solve()` 仍拥有任务合同
+规定的完整 2 秒预算，模块加载时间不会占用该预算。
+
+完整12个场景的 baseline 评测通常只需数秒，内存低于2 GB。
+
+## 直接评测
+
+在任务目录执行：
+
+```bash
+python verification/evaluator.py scripts/init.py \
+  --metrics-out metrics.json \
+  --artifacts-out artifacts.json
+```
+
+## Frontier Eval
+
+在仓库根目录执行，并传入任务环境的 Python：
+
+```bash
+python -m frontier_eval \
+  task=unified \
+  task.benchmark=PowerSystems/SecurityConstrainedDispatch \
+  task.runtime.python_path=/absolute/path/to/.venv-pglib-scd/bin/python \
+  algorithm=openevolve \
+  algorithm.iterations=0
+```
+
+若环境位于 `.venvs/frontier-pglib-scd`，可以改用：
+
+```text
+task.runtime.python_path=uv-env:frontier-pglib-scd
+```
+
+完整任务契约和评分方法见 `Task_zh-CN.md`。

benchmarks/PowerSystems/SecurityConstrainedDispatch/Task.md

@@ -0,0 +1,152 @@
+# Task: Security-Constrained Dispatch
+
+## Background
+
+Power-system operators must dispatch generators economically while respecting
+network physics and remaining feasible after credible component outages. A
+dispatch that is cheap in the intact network may overload a transmission line,
+violate voltage limits, or become infeasible after a line outage.
+
+This benchmark uses three PGLib-OPF systems and twelve deterministic operating
+scenarios. Scenarios combine normal operation, load variation, and selected N-1
+line outages that were verified to admit AC-OPF solutions.
+
+## Editable program
+
+Edit only the logic between `EVOLVE-BLOCK-START` and `EVOLVE-BLOCK-END` in
+`scripts/init.py`. Preserve:
+
+```python
+def solve(case: dict) -> dict:
+    ...
+```
+
+The function is called once per scenario. It must return:
+
+```python
+{
+    "pg_mw": [generator active-power setpoints],
+    "vg_pu": [generator voltage-magnitude setpoints],
+}
+```
+
+Both lists must contain exactly one finite value per generator, in the order
+given by `case["generators"]`. All online generators at the same bus must use
+the same voltage setpoint.
+
+## Input
+
+The `case` dictionary has the following exact top-level fields:
+
+- `case_id: str`: PGLib system identifier.
+- `scenario_id: str`: deterministic operating-scenario identifier.
+- `base_mva: float`: system power base in MVA.
+- `load_scale: float`: demand multiplier applied to the source case.
+- `outage_branch: int | None`: zero-based source-case branch index removed in
+  this scenario, or `None` for an intact network.
+- `total_active_load_mw: float`: sum of active bus demand after scaling.
+- `buses: list[dict]`: buses in MATPOWER row order.
+- `generators: list[dict]`: generators in MATPOWER row order. Output arrays
+  must use this same order.
+- `branches: list[dict]`: branches in MATPOWER row order after applying the
+  scenario outage.
+
+Each entry in `case["buses"]` contains:
+
+- `bus_id: int`, `type: int`;
+- `pd_mw: float`, `qd_mvar: float`;
+- `base_kv: float`;
+- `vmin_pu: float`, `vmax_pu: float`.
+
+Each entry in `case["generators"]` contains:
+
+- `index: int`, `bus_id: int`, `online: bool`;
+- `pmin_mw: float`, `pmax_mw: float`;
+- `qmin_mvar: float`, `qmax_mvar: float`;
+- `baseline_pg_mw: float`, `baseline_vg_pu: float`;
+- `cost_model: list[float]`: the unmodified MATPOWER `gencost` row.
+
+`cost_model` uses one of the MATPOWER encodings below. Polynomial rows are
+`[2, startup, shutdown, n, c_(n-1), ..., c_0]`, representing coefficients in
+descending power order. Piecewise-linear rows are
+`[1, startup, shutdown, n, x_1, y_1, ..., x_n, y_n]`.
+
+Each entry in `case["branches"]` contains:
+
+- `index: int`, `from_bus: int`, `to_bus: int`;
+- `resistance_pu: float`, `reactance_pu: float`, `charging_pu: float`;
+- `rate_a_mva: float` (`0` means no active `RATE_A` limit);
+- `tap_ratio: float`, `phase_shift_deg: float`;
+- `in_service: bool`;
+- `angle_min_deg: float`, `angle_max_deg: float`.
+
+The baseline setpoints are feasible but deliberately non-optimal. They are a
+starting point, not reference answers.
+
+## Hard constraints
+
+A candidate is valid only if every scenario satisfies all of the following:
+
+1. AC Newton power flow converges.
+2. Generator active and reactive powers remain within limits.
+3. Bus voltage magnitudes remain within PGLib limits.
+4. Apparent power at both ends of every rated in-service branch stays within
+   `RATE_A`.
+5. Branch angle differences remain within `ANGMIN` and `ANGMAX`.
+6. Output is finite, has the exact required shape, and returns within two seconds
+   per scenario.
+
+Any hard violation makes the overall candidate invalid and sets
+`combined_score` to zero.
+
+## Scenarios
+
+The benchmark evaluates three systems:
+
+- IEEE RTS 24-bus;
+- IEEE 57-bus;
+- IEEE RTS 73-bus.
+
+Each system has four scenarios:
+
+- 98% load, intact network;
+- 100% load, intact network;
+- 100% load, selected line outage A;
+- 102% load, selected line outage B.
+
+The selected outages and fixed reference values are stored in the readonly
+scenario manifest.
+
+## Score
+
+For each feasible scenario, three normalized components are computed:
+
+```text
+cost_efficiency = min(1.05, reference_AC_OPF_cost / candidate_cost)
+thermal_margin  = clip((100 - max_branch_loading_percent) / 20, 0, 1)
+voltage_margin  = 10th percentile normalized distance from voltage limits
+
+scenario_score = 70 * cost_efficiency
+               + 20 * thermal_margin
+               + 10 * voltage_margin
+```
+
+The aggregate score emphasizes both average and worst-case behavior:
+
+```text
+combined_score = 0.75 * mean(scenario_score)
+               + 0.25 * min(scenario_score)
+```
+
+Higher is better. The cost term rewards proximity to the frozen AC-OPF
+references; the two margin terms reward dispatches that do not merely sit on
+thermal or voltage limits. The cost component is capped to limit sensitivity to
+numerical solver differences.
+
+## Prohibited behavior
+
+- Do not modify the evaluator, PGLib cases, scenario manifest, or metadata.
+- Do not read or write files, launch subprocesses, access the network, or use
+  wall-clock-dependent behavior from `solve`.
+- Do not hard-code evaluator outputs or bypass AC power flow.
+- Keep the policy deterministic.