World-space hatching lab: four strategies compared on an architectural test scene

[cbgbt]

What this is

An experimentation harness for the "beautiful architectural renderings" direction: one test scene (gabled house with door/window openings, chimney, sphere-canopy tree, ground plane, single point light) rendered through four hatching strategies so they can be compared visually. Run with:

cargo run --release -p hatch-lab

Outputs land in hatch-lab/renders/*.svg (gitignored, like all render outputs in this repo).

Core seams added to raydeon

Two minimal public APIs that world-space hatching needs, both extracted from existing internals rather than new logic:

• Scene::illumination_for_hit — computes total light arriving at a caller-constructed HitShape, so hatch generators can sample illumination at surface points they already know.
• SceneCamera::clip_and_project — the existing chop → occlusion-clip → join → project pipeline, now callable with arbitrary caller-generated world-space segments. render() is reimplemented on top of it.

Strategies

	Strategy	Idea
a	screen_space	Existing screen-space hatching (baseline), seeded
b	world_uniform	45° hatch lines generated on each face in world space; stochastic keep driven by sampled illumination
c	crosshatch_tonal	Discrete tone bands accumulate cross directions at tightening spacing (engraving look)
d	light_flow	Hatch direction = light direction projected onto each face; sphere rings wrap the light axis

Hatch lines are generated in face-plane coordinates, clipped to the convex outline minus door/window holes, lifted ~0.006 world units off the surface (so their visibility rays don't re-intersect their own face), then illumination-filtered per ~0.16-unit sample using the same slice/join machinery as the main renderer.

Notable findings while building this

• Hatch lines lying exactly on a face self-occlude — the midpoint visibility ray hits its own surface at distance ~0 and gets clipped. The existing shapes dodge this by expanding outlines outside their face; surface detail needs a normal lift. A first-class world-space hatching API in raydeon should handle this internally.
• join_slices drops the shape reference when rebuilding segments, so render output can't currently be traced back to materials — this blocks per-pen/per-layer SVG output and is worth fixing in the planned cleanup.
• World-space strategies are markedly cheaper than the screen-space pass here (~10–40ms vs ~120ms) because they only sample where surfaces exist, and the screen-space hatch filter is currently single-threaded due to the shared RNG.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WT2VUpRTEKguqL42ehtXdn

[cbgbt]

Closing per Sean's preference to work from branches without PRs — the branch stays pushed.