You are a mechanical design engineer. A customer sends you catalogue drawings of standard parts and needs editable 3D models.

## Input
- Twelve parts in `/data/cad/ale_runs/benchcad_views_to_step/20260909T011320_gpt-5.6-terra/input/parts/<stem>/`: `view_0.png` .. `view_3.png` are orthographic renders, `composite.png` puts them side by side, and `meta.json` names the part family, the governing standard when there is one, and a difficulty tag.

## What you must do
1. Study the renders and the standard to infer the geometry and proportions of each part.
2. Model each part as a single closed solid with any open-source CAD tool.
3. Export every part to `/data/cad/ale_runs/benchcad_views_to_step/20260909T011320_gpt-5.6-terra/output/<stem>.step` (STEP AP214, one solid per file, any units).

## Scoring
Each output solid is centred and scaled to unit bounding-box diagonal and compared with the reference solid by volume IoU; the best of the 24 axis-aligned orientations counts. Part score = (IoU - 0.5) / 0.5 clipped to [0, 1]; task score = mean over the twelve parts.

## Environment
Linux, 2 vCPU, 1.9 GB RAM (stream large files; do not load 8 GB into memory). `python` is a venv with cadquery/OCP, numpy, scipy, pandas, scikit-learn; pip install is allowed. FreeCAD/OpenSCAD are NOT installed; use CadQuery for geometry. No GUI. The task README is at /data/cad/ale_runs/benchcad_views_to_step/20260909T011320_gpt-5.6-terra/input/README.md. Verify your deliverable file exists and parses before finishing.
