feat(StructuralOptimization): add composite laminate stacking benchmark

TASK_DETAILS.md

@@ -286,7 +286,7 @@ We welcome new engineering problem ideas — even without complete verification
       <td>JSSP YN family (Yamada and Nakano 1992)</td>
     </tr>
     <tr>
-      <td rowspan="4"><b>StructuralOptimization</b></td>
+      <td rowspan="5"><b>StructuralOptimization</b></td>
       <td><code>ISCSO2015</code></td>
       <td>Minimize weight of 45-bar 2D truss under stress / displacement constraints</td>
     </tr>
@@ -302,6 +302,10 @@ We welcome new engineering problem ideas — even without complete verification
       <td><code>PyMOTOSIMPCompliance</code></td>
       <td>pyMOTO-based 2D beam topology optimization (SIMP + OC/MMA) under a volume-fraction constraint</td>
     </tr>
+    <tr>
+      <td><code>CompositeLaminateStacking</code></td>
+      <td>Optimize balanced, symmetric 48-ply composite laminates for buckling and maximum-strain failure across ten plate cases</td>
+    </tr>
     <tr>
       <td rowspan="6"><b>Robotics</b></td>
       <td><code>DynamicObstacleAvoidanceNavigation</code></td>

TASK_DETAILS_zh-CN.md

@@ -286,7 +286,7 @@ Frontier-Eng 目前已覆盖以下领域的任务。每个任务均配有可运
       <td>JSSP YN 家族（Yamada &amp; Nakano 1992）</td>
     </tr>
     <tr>
-      <td rowspan="4"><b>StructuralOptimization</b></td>
+      <td rowspan="5"><b>StructuralOptimization</b></td>
       <td><code>ISCSO2015</code></td>
       <td>在应力 / 位移约束下最小化 45 杆 2D 桁架重量</td>
     </tr>
@@ -302,6 +302,10 @@ Frontier-Eng 目前已覆盖以下领域的任务。每个任务均配有可运
       <td><code>PyMOTOSIMPCompliance</code></td>
       <td>基于 pyMOTO 的 2D 梁拓扑优化（SIMP + OC/MMA），体积分数约束</td>
     </tr>
+    <tr>
+      <td><code>CompositeLaminateStacking</code></td>
+      <td>在十个板工况下优化平衡、对称的 48 层复合材料铺层，同时提高屈曲和最大应变失效性能</td>
+    </tr>
     <tr>
       <td rowspan="6"><b>Robotics</b></td>
       <td><code>DynamicObstacleAvoidanceNavigation</code></td>

benchmarks/StructuralOptimization/CompositeLaminateStacking/README.md

@@ -0,0 +1,53 @@
+# CompositeLaminateStacking
+
+Optimize balanced, symmetric 48-ply composite laminates across ten simply supported
+plate cases derived from an MIT-licensed published benchmark.
+
+## What the agent edits
+
+Only [`scripts/init.py`](scripts/init.py) is editable. Implement:
+
+```python
+def design_laminates(cases: list[dict]) -> dict[str, list[int]]:
+    ...
+```
+
+Return one list of 12 integer base angles in `[0, 90]` for every `case_id`. The evaluator
+expands the list into a balanced, symmetric 48-ply stack. See [`Task.md`](Task.md) for the
+complete contract and [`references/design_notes.md`](references/design_notes.md) for the
+mechanics and source cross-check.
+
+## Requirements
+
+- Python 3.10+
+- NumPy
+- no GPU, Docker, external data, or proprietary solver
+
+Install task-local requirements if needed:
+
+```bash
+python -m pip install -r verification/requirements.txt
+```
+
+## Direct evaluation
+
+From this task directory:
+
+```bash
+python verification/evaluator.py scripts/init.py
+python -m unittest discover -s verification -p "test_*.py" -v
+```
+
+## Unified evaluation
+
+From the repository root:
+
+```bash
+python -m frontier_eval \
+  task=unified \
+  task.benchmark=StructuralOptimization/CompositeLaminateStacking \
+  algorithm=openevolve \
+  algorithm.iterations=0
+```
+
+The baseline is feasible in all ten cases and scores 50. Higher is better.

benchmarks/StructuralOptimization/CompositeLaminateStacking/README_zh-CN.md

@@ -0,0 +1,52 @@
+# CompositeLaminateStacking
+
+在十个简支板工况上优化平衡、对称的 48 层复合材料铺层。任务参数来自一个采用
+MIT 许可证发布的真实研究 benchmark。
+
+## Agent 修改内容
+
+仅允许修改 [`scripts/init.py`](scripts/init.py)，并实现：
+
+```python
+def design_laminates(cases: list[dict]) -> dict[str, list[int]]:
+    ...
+```
+
+必须为每个 `case_id` 返回一个包含 12 个 `[0, 90]` 区间整数角度的列表。评测器会
+把它展开为平衡、对称的 48 层铺层。完整接口见 [`Task_zh-CN.md`](Task_zh-CN.md)，
+力学模型与来源交叉验证见 [`references/design_notes.md`](references/design_notes.md)。
+
+## 环境要求
+
+- Python 3.10+
+- NumPy
+- 不需要 GPU、Docker、外部数据或商业求解器
+
+如有需要，可安装任务依赖：
+
+```bash
+python -m pip install -r verification/requirements.txt
+```
+
+## 直接评测
+
+在当前任务目录运行：
+
+```bash
+python verification/evaluator.py scripts/init.py
+python -m unittest discover -s verification -p "test_*.py" -v
+```
+
+## Unified 评测
+
+在仓库根目录运行：
+
+```bash
+python -m frontier_eval \
+  task=unified \
+  task.benchmark=StructuralOptimization/CompositeLaminateStacking \
+  algorithm=openevolve \
+  algorithm.iterations=0
+```
+
+baseline 在全部十个工况中均可行，得分为 50；分数越高越好。

benchmarks/StructuralOptimization/CompositeLaminateStacking/Task.md

@@ -0,0 +1,127 @@
+# Task: Composite Laminate Stacking Optimization
+
+## 1. Engineering problem
+
+Design the stacking sequence of a 48-ply orthotropic composite plate. A good laminate
+must resist both material failure and plate buckling under several geometries and load
+ratios. This is a discrete, coupled structural-design problem: moving stiff plies toward
+the surfaces can help buckling, while changing fiber directions alters in-plane strength
+and can hurt another load case.
+
+The task is derived from the MIT-licensed Zenodo release *Beyond Double-Double Theory:
+n-Directional Stacking Sequence Optimisation in Composite Laminates*
+([doi:10.5281/zenodo.15864525](https://doi.org/10.5281/zenodo.15864525)). It preserves
+the released 48-ply construction, orthotropic material, ply thickness, five aspect ratios,
+uniaxial/biaxial compression, strain allowables, and Haftka reference layup. The evaluator
+is an independent NumPy implementation documented in `references/design_notes.md`.
+
+## 2. Candidate API
+
+Implement in `scripts/init.py`:
+
+```python
+def design_laminates(cases: list[dict]) -> dict[str, list[int]]:
+    ...
+```
+
+The function receives all public cases in one call. Each case contains only:
+
+```python
+{
+    "case_id": str,
+    "aspect_ratio": float,
+    "a_mm": float,
+    "b_mm": float,
+    "Nx_N_per_mm": float,
+    "Ny_N_per_mm": float,
+}
+```
+
+Compression is negative. `Ny_N_per_mm == 0` denotes uniaxial loading; a negative value
+denotes biaxial compression.
+
+Return a dictionary with exactly the supplied `case_id` values. Each value must be a JSON
+list containing exactly 12 finite integer angles in the inclusive interval `[0, 90]`.
+
+## 3. Layup construction
+
+For candidate angles `theta_1 ... theta_12`, the verifier constructs the first half as:
+
+```text
+[theta_1, -theta_1, theta_2, -theta_2, ..., theta_12, -theta_12]
+```
+
+and appends its reverse. The result has 48 plies and is balanced and symmetric by
+construction. The candidate therefore optimizes both angle selection and through-thickness
+ordering without needing to implement manufacturing-constraint repair.
+
+## 4. Evaluation cases
+
+Ten deterministic cases combine aspect ratios `0.5, 1, 2, 3, 4` with:
+
+- uniaxial compression: `Nx = -0.175126835 N/mm`, `Ny = 0`;
+- biaxial compression: `Nx = -0.175126835 N/mm`, `Ny = -0.087563418 N/mm`.
+
+The plate widths are fixed at `127 mm`; lengths vary with aspect ratio. Material and
+allowable values are committed in `references/config.json`.
+
+## 5. Physics
+
+The verifier applies classical lamination theory to compute `A`, `B`, and `D`. It then
+computes:
+
+1. a maximum-strain failure load factor from the most critical longitudinal, transverse,
+   or shear ply strain under the released 1.5 design-load factor;
+2. a simply supported plate buckling load factor from a double-sine Ritz basis, numerical
+   quadrature, and a symmetric generalized eigenproblem.
+
+The governing reserve factor is:
+
+```text
+reserve = min(failure_load_factor, buckling_load_factor)
+```
+
+## 6. Score
+
+Each case is normalized against the released Haftka reference stacking sequence:
+
+```text
+case_score = clip(50 + 50*tanh(log(reserve/anchor_reserve)/0.5), 0, 100)
+```
+
+The final diagnostic score is:
+
+```text
+0.75 * mean(case_scores) + 0.25 * percentile(case_scores, 20)
+```
+
+A valid candidate receives that value as `combined_score`. Missing cases, extra case ids,
+bad types, non-integer/out-of-range angles, timeouts, import failures, or invalid mechanics
+make the candidate invalid and set `combined_score` to zero. Per-case diagnostic feedback
+remains available.
+
+## 7. Runtime and integrity contract
+
+- Keep `design_laminates` deterministic and self-contained.
+- Candidate import and execution occur in a separate Python process with a bounded runtime.
+- The process receives cases and returns designs through a JSON-lines protocol.
+- Candidate stdout is discarded so it cannot corrupt the evaluator protocol.
+- This is process isolation, not an operating-system security sandbox.
+- Do not read or modify evaluator, reference, output, or environment-secret files.
+- Edit only the region between the `EVOLVE-BLOCK` markers in `scripts/init.py`.
+
+## 8. Commands
+
+Direct evaluation, from this directory:
+
+```bash
+python verification/evaluator.py scripts/init.py
+```
+
+Unified evaluation, from the repository root:
+
+```bash
+python -m frontier_eval task=unified \
+  task.benchmark=StructuralOptimization/CompositeLaminateStacking \
+  algorithm=openevolve algorithm.iterations=0
+```

benchmarks/StructuralOptimization/CompositeLaminateStacking/Task_zh-CN.md

@@ -0,0 +1,87 @@
+# 任务：复合材料层合板铺层优化
+
+## 1. 工程问题
+
+设计一个 48 层正交各向异性复合材料板的铺层顺序。铺层需要在多种几何尺寸和载荷
+比例下同时抵抗材料失效和板屈曲。这是一个耦合的离散结构设计问题：把高刚度方向
+的铺层放到外侧可能提高屈曲性能，但改变纤维方向也会改变面内强度，并可能损害另一
+载荷工况。
+
+本任务源自采用 MIT 许可证发布的 Zenodo 数据集 *Beyond Double-Double Theory:
+n-Directional Stacking Sequence Optimisation in Composite Laminates*
+（[doi:10.5281/zenodo.15864525](https://doi.org/10.5281/zenodo.15864525)）。任务保留了
+原发布代码中的 48 层构造、正交各向异性材料、单层厚度、五种长宽比、单轴/双轴压缩、
+应变许用值和 Haftka 参考铺层。评测器采用独立的 NumPy 实现，详见
+`references/design_notes.md`。
+
+## 2. 候选程序接口
+
+在 `scripts/init.py` 中实现：
+
+```python
+def design_laminates(cases: list[dict]) -> dict[str, list[int]]:
+    ...
+```
+
+函数一次收到所有公开工况。每个工况仅包含 `case_id`、长宽比、板长宽和 `Nx/Ny`
+膜内载荷。负值表示压缩；`Ny_N_per_mm == 0` 表示单轴载荷。
+
+返回字典必须恰好包含输入给出的全部 `case_id`。每个值必须是可 JSON 序列化的列表，
+包含恰好 12 个 `[0, 90]` 闭区间内的有限整数角度。
+
+## 3. 铺层构造
+
+对于候选角度 `theta_1 ... theta_12`，评测器先构造：
+
+```text
+[theta_1, -theta_1, theta_2, -theta_2, ..., theta_12, -theta_12]
+```
+
+再追加其逆序，得到平衡、对称的 48 层铺层。候选程序需要优化角度选择和厚度方向顺序，
+但不需要自行修复制造约束。
+
+## 4. 评测工况与物理模型
+
+十个确定性工况由长宽比 `0.5, 1, 2, 3, 4` 与单轴、双轴压缩组合而成。板宽固定为
+`127 mm`，长度随长宽比变化。材料、载荷和许用值记录在 `references/config.json`。
+
+评测器使用经典层合板理论计算 `A`、`B`、`D`，然后求得：
+
+1. 由最危险的纵向、横向或剪切层内应变确定的最大应变失效载荷因子；
+2. 通过双正弦 Ritz 基、数值积分和对称广义特征值问题得到的简支板屈曲载荷因子。
+
+控制储备因子为：
+
+```text
+reserve = min(failure_load_factor, buckling_load_factor)
+```
+
+## 5. 评分
+
+每个工况相对于已发布的 Haftka 参考铺层归一化：
+
+```text
+case_score = clip(50 + 50*tanh(log(reserve/anchor_reserve)/0.5), 0, 100)
+```
+
+最终诊断分数为 `75%` 平均工况分加 `25%` 的第 20 百分位工况分。缺少工况、多余 ID、
+错误类型、非整数或越界角度、超时、导入失败或非法力学结果都会令 `combined_score=0`；
+逐工况诊断反馈仍会保留。
+
+## 6. 运行与完整性约束
+
+- `design_laminates` 必须确定性、可独立运行。
+- 候选程序在受时间限制的独立 Python 进程中导入和执行。
+- 工况和设计通过 JSON-lines 协议交换；候选 stdout 会被丢弃。
+- 这是进程隔离，并非操作系统安全沙箱。
+- 不得读取或修改 evaluator、参考参数、输出或环境机密文件。
+- 仅修改 `scripts/init.py` 内 `EVOLVE-BLOCK` 标记之间的区域。
+
+## 7. 运行命令
+
+```bash
+python verification/evaluator.py scripts/init.py
+python -m frontier_eval task=unified \
+  task.benchmark=StructuralOptimization/CompositeLaminateStacking \
+  algorithm=openevolve algorithm.iterations=0
+```

benchmarks/StructuralOptimization/CompositeLaminateStacking/baseline/result_log.txt

@@ -0,0 +1,32 @@
+CompositeLaminateStacking baseline verification
+================================================
+
+Reference dataset:
+  Dataset for Beyond Double-Double Theory: n-Directional Stacking Sequence
+  Optimisation in Composite Laminates
+  DOI: 10.5281/zenodo.15864525
+  License: MIT
+
+Verification date: 2026-08-13
+
+Direct evaluator:
+  python verification/evaluator.py scripts/init.py
+  feasible_cases: 10 / 10
+  diagnostic_score: 50.0
+  combined_score: 50.0
+
+Unit tests:
+  python -m unittest discover -s verification -p "test_*.py" -v
+  result: 13 tests passed
+
+Unified evaluator:
+  python -m frontier_eval task=unified \
+    task.benchmark=StructuralOptimization/CompositeLaminateStacking \
+    algorithm=openevolve algorithm.iterations=0
+  benchmark_returncode: 0
+  valid: 1
+  feasible_cases: 10
+  combined_score: 50.0
+
+The unified run was also exercised on Windows with platform-specific shell and
+Python-path overrides; the benchmark output and score matched the direct run.