feat(WindEnergy): add WakeSteeringControl benchmark

benchmarks/WindEnergy/README.md

@@ -0,0 +1,9 @@
+# Wind Energy
+
+Wind-energy benchmarks evaluate control and design decisions with executable,
+physics-based wind-plant models.
+
+## Tasks
+
+- [`WakeSteeringControl`](WakeSteeringControl/README.md): optimize turbine yaw
+  control to increase weighted farm energy while limiting control effort.

benchmarks/WindEnergy/README_zh-CN.md

@@ -0,0 +1,8 @@
+# 风能工程
+
+风能工程任务使用可执行的风电场物理模型，评测控制与设计策略的优化能力。
+
+## 任务
+
+- [`WakeSteeringControl`](WakeSteeringControl/README_zh-CN.md)：优化风机偏航控制，
+  在限制控制代价的同时提升风电场加权发电量。

benchmarks/WindEnergy/WakeSteeringControl/README.md

@@ -0,0 +1,69 @@
+# Wake Steering Control
+
+This benchmark asks an agent to improve a deterministic yaw-control policy for
+a nine-turbine wind farm. A frozen FLORIS 4.6.6 verifier evaluates the policy
+over multiple wind directions and speeds.
+
+## Structure
+
+```text
+scripts/init.py                 editable yaw policy
+references/farm_config.json     public farm and score contract
+verification/evaluator.py       frozen FLORIS evaluator
+verification/run_candidate.py   isolated candidate runner
+frontier_eval/                  unified-task metadata
+baseline/result_log.json        measured zero-yaw baseline
+```
+
+## Environment
+
+From the Frontier-Engineering repository root, create the task runtime. The task
+is tested with Python 3.12 on Linux.
+
+```bash
+python3.12 -m venv .venvs/frontier-wake-steering
+.venvs/frontier-wake-steering/bin/python -m pip install \
+  -r benchmarks/WindEnergy/WakeSteeringControl/verification/requirements.txt
+```
+
+The task is CPU-only, and FLORIS itself does not require a GPU. A baseline
+evaluation typically completes within seconds. No external dataset, Docker, or
+runtime network access is required.
+
+## Direct Evaluation
+
+From this benchmark directory:
+
+```bash
+../../../.venvs/frontier-wake-steering/bin/python \
+  verification/evaluator.py scripts/init.py
+```
+
+The evaluator writes `metrics.json` and `artifacts.json`. The shipped zero-yaw
+policy is feasible and has a combined score of approximately zero by design.
+
+## Unified Evaluation
+
+The unified benchmark ID is `WindEnergy/WakeSteeringControl`:
+
+```bash
+python -m frontier_eval \
+  task=unified \
+  task.benchmark=WindEnergy/WakeSteeringControl \
+  task.runtime.python_path=uv-env:frontier-wake-steering \
+  algorithm=openevolve \
+  algorithm.iterations=0
+```
+
+Optimization runs require the normal model API configuration used by
+Frontier Eval. Baseline-only validation does not require an API key.
+
+## Reproducibility and Security
+
+- FLORIS is pinned to version 4.6.6.
+- The verifier uses the built-in NREL 5 MW turbine model and a fixed farm.
+- Candidate code runs in a separate subprocess with a wall-clock timeout.
+- The unified task checks the verifier and public contract as read-only and
+  invalidates candidates that modify them.
+- Process isolation is not a hardened security sandbox. Run untrusted candidate
+  code in a container or disposable host.

benchmarks/WindEnergy/WakeSteeringControl/README_zh-CN.md

@@ -0,0 +1,62 @@
+# 风电场尾流偏航控制
+
+本任务要求 Agent 优化一个确定性的偏航控制策略。冻结的 FLORIS 4.6.6
+验证器在多种风向和风速下评估九台风机组成的风电场。
+
+## 目录
+
+```text
+scripts/init.py                 可编辑偏航策略
+references/farm_config.json     公开风场与评分契约
+verification/evaluator.py       冻结的 FLORIS 验证器
+verification/run_candidate.py   候选代码隔离运行器
+frontier_eval/                  unified-task 元数据
+baseline/result_log.json        零偏航基线结果
+```
+
+## 环境
+
+在 Frontier-Engineering 仓库根目录执行：
+
+```bash
+python -m venv .venvs/frontier-wake-steering
+.venvs/frontier-wake-steering/bin/python -m pip install \
+  -r benchmarks/WindEnergy/WakeSteeringControl/verification/requirements.txt
+```
+
+本任务仅需 CPU，FLORIS 不要求 GPU。
+
+## 直接评测
+
+在任务目录中执行：
+
+```bash
+../../../.venvs/frontier-wake-steering/bin/python \
+  verification/evaluator.py scripts/init.py
+```
+
+验证器生成 `metrics.json` 和 `artifacts.json`。仓库自带的零偏航策略保证可行，
+其综合分数按定义约为零。
+
+## Unified 评测
+
+任务 ID 为 `WindEnergy/WakeSteeringControl`：
+
+```bash
+python -m frontier_eval \
+  task=unified \
+  task.benchmark=WindEnergy/WakeSteeringControl \
+  task.runtime.python_path=uv-env:frontier-wake-steering \
+  algorithm=openevolve \
+  algorithm.iterations=0
+```
+
+零迭代验证不需要模型 API Key；正式优化使用 Frontier Eval 的常规模型配置。
+
+## 复现与安全
+
+- FLORIS 固定为 4.6.6。
+- 验证器使用内置 NREL 5 MW 风机和固定风场。
+- 候选代码在带超时的独立子进程中运行。
+- unified task 检查验证器与公开契约是否保持只读，修改这些文件的候选会被判为无效。
+- 进程隔离不是强安全沙箱；不可信候选应在容器或一次性主机中运行。

benchmarks/WindEnergy/WakeSteeringControl/Task.md

@@ -0,0 +1,63 @@
+# Task: Wind-Farm Wake-Steering Control
+
+## Background
+
+An upstream turbine extracts energy and leaves a slower wake. When turbines are
+aligned with the wind, that wake can reduce the production of downstream
+turbines. Deliberately yawing an upstream rotor can redirect its wake. The
+upstream turbine loses some power, but the farm may gain more power overall.
+
+## Objective
+
+Implement `yaw_policy()` in `scripts/init.py`. For each wind condition, return
+one yaw angle per turbine. Maximize frequency-weighted farm energy relative to
+the zero-yaw baseline while avoiding excessive yaw and abrupt changes between
+nearby wind directions.
+
+## Inputs
+
+The policy receives Python lists:
+
+- `wind_directions_deg`: meteorological wind directions in degrees;
+- `wind_speeds_mps`: wind speeds in metres per second;
+- `turbulence_intensities`: dimensionless turbulence intensity values;
+- `layout_x_m`, `layout_y_m`: fixed turbine coordinates in metres.
+
+There are `n_conditions` wind conditions and `n_turbines` layout points.
+
+## Output
+
+Return a rectangular nested list or compatible numeric array with shape:
+
+```text
+(n_conditions, n_turbines)
+```
+
+Every value is a yaw angle in degrees and must lie in `[-25, 25]`. Positive and
+negative signs represent opposite steering directions under FLORIS conventions.
+
+## Score
+
+The verifier computes weighted expected farm power for the candidate and the
+zero-yaw baseline. The higher-is-better score is:
+
+```text
+relative energy gain in percentage points
+- 0.01 * frequency-weighted mean absolute yaw in degrees
+- 0.002 * mean adjacent-direction yaw change in degrees
+```
+
+The zero-yaw baseline therefore scores approximately `0.0`. Raw energy and
+penalty components are reported separately.
+
+## Hard Constraints
+
+- Return the exact required shape and only finite numeric values.
+- Keep every yaw angle within `[-25, 25]` degrees.
+- Return identical results for identical inputs.
+- Complete within the candidate timeout.
+- Do not modify benchmark, reference, or verifier files.
+- Keep the function signature and EVOLVE markers unchanged.
+
+The verifier determines feasibility and score. During unified evaluation,
+Frontier Eval also invalidates candidates that modify files declared read-only.

benchmarks/WindEnergy/WakeSteeringControl/Task_zh-CN.md

@@ -0,0 +1,59 @@
+# 任务：风电场尾流偏航控制
+
+## 背景
+
+上游风机提取能量后会形成低速尾流。当风机沿风向排列时，尾流会降低下游风机
+发电量。上游风机主动偏航可以改变尾流方向；虽然上游风机会损失部分功率，整个
+风电场的总功率可能增加。
+
+## 目标
+
+实现 `scripts/init.py` 中的 `yaw_policy()`。针对每个风况，为每台风机返回一个
+偏航角。在限制偏航幅度和相邻风向策略突变的同时，相对零偏航基线最大化加权
+风电场发电量。
+
+## 输入
+
+策略接收以下 Python 列表：
+
+- `wind_directions_deg`：气象风向，单位为度；
+- `wind_speeds_mps`：风速，单位为米每秒；
+- `turbulence_intensities`：无量纲湍流强度；
+- `layout_x_m`、`layout_y_m`：固定风机坐标，单位为米。
+
+风况数量为 `n_conditions`，风机数量为 `n_turbines`。
+
+## 输出
+
+返回矩形嵌套列表或兼容数值数组，形状必须为：
+
+```text
+(n_conditions, n_turbines)
+```
+
+每个值表示偏航角，范围为 `[-25, 25]` 度。正负号表示 FLORIS 约定下两个相反
+的尾流转向方向。
+
+## 评分
+
+验证器分别计算候选策略和零偏航基线的加权期望功率。综合分数越高越好：
+
+```text
+相对发电量增益（百分点）
+- 0.01 * 频率加权平均绝对偏航角
+- 0.002 * 相邻风向平均偏航变化
+```
+
+因此零偏航基线分数约为 `0.0`。验证器会分别报告原始发电量和惩罚项。
+
+## 硬约束
+
+- 返回精确形状和有限数值；
+- 所有偏航角均在 `[-25, 25]` 度内；
+- 相同输入必须返回相同结果；
+- 必须在候选超时前完成；
+- 不得修改任务、参考资料及验证器文件；
+- 不得改变函数签名和 EVOLVE 标记。
+
+验证器负责判定可行性和最终分数。在 unified 评测中，Frontier Eval 还会将修改
+只读声明文件的候选判为无效。

benchmarks/WindEnergy/WakeSteeringControl/baseline/result_log.json

@@ -0,0 +1,15 @@
+{
+  "baseline_expected_power_w": 14562813.187886339,
+  "candidate_expected_power_w": 14562813.187886339,
+  "candidate_runtime_s": 0.05653033405542374,
+  "combined_score": 0.0,
+  "evaluator_runtime_s": 0.2937989681959152,
+  "n_conditions": 24.0,
+  "n_turbines": 9.0,
+  "relative_energy_gain_pct": 0.0,
+  "valid": 1.0,
+  "weighted_mean_abs_yaw_deg": 0.0,
+  "yaw_effort_penalty": 0.0,
+  "yaw_smoothness_deg": 0.0,
+  "yaw_smoothness_penalty": 0.0
+}

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/agent_files.txt

@@ -0,0 +1,7 @@
+README.md
+README_zh-CN.md
+Task.md
+Task_zh-CN.md
+references/README.md
+references/farm_config.json
+frontier_eval/constraints.txt

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/artifact_files.txt

@@ -0,0 +1,2 @@
+metrics.json
+artifacts.json

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/candidate_destination.txt

@@ -0,0 +1 @@
+scripts/init.py

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/constraints.txt

@@ -0,0 +1,7 @@
+WakeSteeringControl constraints:
+1) Only modify the EVOLVE block in `scripts/init.py`.
+2) Keep the `yaw_policy()` signature unchanged and return shape `(n_conditions, n_turbines)`.
+3) Return only finite yaw angles in `[-25, 25]` degrees.
+4) Keep the policy deterministic and complete within the candidate timeout.
+5) Do not modify task, reference, or verification files; unified evaluation checks them as read-only.
+6) Maximize weighted farm-energy gain while limiting yaw effort and directional discontinuity.

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/copy_files.txt

@@ -0,0 +1,8 @@
+README.md
+README_zh-CN.md
+Task.md
+Task_zh-CN.md
+references
+verification
+scripts
+frontier_eval/run_eval.sh

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/eval_command.txt

@@ -0,0 +1 @@
+bash frontier_eval/run_eval.sh {python} {candidate}

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/eval_cwd.txt

@@ -0,0 +1 @@
+.

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/initial_program.txt

@@ -0,0 +1 @@
+scripts/init.py

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/readonly_files.txt

@@ -0,0 +1,7 @@
+README.md
+README_zh-CN.md
+Task.md
+Task_zh-CN.md
+references
+verification
+frontier_eval/run_eval.sh

benchmarks/WindEnergy/WakeSteeringControl/frontier_eval/run_eval.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+PYTHON_CMD="${1:?missing Python interpreter}"
+CANDIDATE_PATH="${2:?missing candidate path}"
+
+exec "${PYTHON_CMD}" verification/evaluator.py "${CANDIDATE_PATH}" \
+  --json-out metrics.json \
+  --artifacts-out artifacts.json

benchmarks/WindEnergy/WakeSteeringControl/references/README.md

@@ -0,0 +1,8 @@
+# Public Configuration
+
+`farm_config.json` defines the public farm geometry, yaw limits, timeout, and
+score coefficients. Coordinates are expressed in rotor-diameter multiples and
+are converted to metres using the NREL 5 MW rotor diameter loaded by FLORIS.
+
+The evaluator owns the deterministic wind-condition ensemble. A candidate
+receives every condition it must control through the documented policy inputs.

benchmarks/WindEnergy/WakeSteeringControl/references/farm_config.json

@@ -0,0 +1,20 @@
+{
+  "floris_version": "4.6.6",
+  "turbine_model": "nrel_5MW",
+  "layout_rotor_diameters": [
+    [0.0, 0.0],
+    [5.0, 0.0],
+    [10.0, 0.0],
+    [0.0, 5.0],
+    [5.0, 5.0],
+    [10.0, 5.0],
+    [0.0, 10.0],
+    [5.0, 10.0],
+    [10.0, 10.0]
+  ],
+  "yaw_min_deg": -25.0,
+  "yaw_max_deg": 25.0,
+  "candidate_timeout_s": 5.0,
+  "yaw_effort_coefficient": 0.01,
+  "smoothness_coefficient": 0.002
+}