Imported from snjsomnath/foton (.agents/skills/daylight-benchmark-scene/SKILL.md). Install upstream with npx skills add snjsomnath/foton --skill daylight-benchmark-scene. Copyright stays with the author.

Daylight Benchmark Scene

Generate a fixed shoebox before tuning GPU code. Keep geometry, materials, sensor positions, sky basis, and Radiance options versioned together so benchmark changes are intentional.

Run scripts/generate_shoebox.py to create:

• materials.rad with 70% walls, 20% floor, and 80% ceiling.
• room.rad for a 6 m × 9 m × 3 m room with one open south aperture.
• sensors.pts at 0.75 m with a 0.5 m grid.
• scene.json containing dimensions, counts, and coordinate conventions.

python scripts/generate_shoebox.py --output-dir benchmarks/shoebox-v1

Treat this as a geometry/transport fixture, not an LM-83 model. Add glazing, exterior obstructions, blinds, and occupancy only in separately named benchmark variants.

For Honeybee direct-visibility fixtures, clone and convert the model to metres without mutating the caller. Triangulate punched opaque parent faces, omit aperture polygons so openings remain open, and include overhangs, fins, orphaned shades, and shade meshes as opaque geometry. Feed identical patch-center directions, flattened sensor ordering, and geometry semantics to the GPU and Radiance paths.

Maintain a complete canonical benchmark suite rather than a single fixture:

• A shaded Honeybee shoebox with aperture, overhang, fins, and 216 sensors for direct visibility.
• The same material/geometry contract with diffuse and thin-glass transport.
• Annual illuminance and DA/sDA derived from matching coefficient and sky matrices.
• An instanced 1,000-room scene with 25 sensors per room for cold AS construction.
• An unchanged rerun of that scene for resident BLAS/TLAS cache validation.

Version the fixture identity and all scale/sample parameters in each report. Do not replace the 1,000-room fixture with a smaller smoke scene in published hardware tables.