fix(examples): redesign gp-lineart-contour as an inked lighthouse diorama

[TMHSDigital]

What and why

The gallery review judged the gp-lineart-contour hero weak: a single teal octahedron with a cyan outline on a small black stand. It was a lone primitive, not a designed subject.

Redesign: the Line Art source is now one cel-shaded mesh, a lighthouse diorama:

• a banded red and white tapered tower with a door and lit windows
• a gallery deck with railing posts, a lantern room with mullions, a roof and a vent ball
• a keeper's cottage with a pitched roof, chimney, door and windows
• a rowboat hauled up on the faceted rocks
• a round diorama sea with white wave crests

The materials are flat toon tones (Diffuse → Shader to RGB → constant ramp, emitted), so the bold black LINEART ink is what makes the render read as an inked comic illustration:

• material-border ink between the tower bands
• intersection ink at the waterline around every rock
• crease ink on the rock facets
• contour ink on every silhouette

The render is the proof. With source_object cleared, the same render (made by the scratch probe .scratch/broken.py, not committed) is a flat lineless low-poly model. There are no band borders, no waterline and no outlines. The difference is obvious at thumbnail size.

• Contact sheet (committed): docs/gallery/contact-sheets/gp-lineart-contour-contact-sheet.webp
• New hero: docs/gallery/assets/gp-lineart-contour-hero.webp (26,966 B)

Stage deviation (documented in README): this is an NPR contract, so the hero materials are cel-shaded emission instead of Principled. The floor, wall, world, key, fill and warm wedge are the default stage. The cel look uses Shader to RGB, which is EEVEE-only.

Asset-sheet gate: not run. This is an NPR / art subject in the rendering section, not a game-prop or kit example, and scripts/asset_sheet.py has no SELECT row for it.

Check changes (stricter, not weaker)

• One camera for the check and the still. Line Art is view-dependent, so the counts the check gates are exactly the strokes the still draws.

• Stroke and point gates. They went from >= 1 stroke / >= 4 points to >= 240 strokes / >= 1320 points. The measured drawing is 255 / 1393 on all three binaries. Dropping any one edge type leaves:

  • contour: 234 strokes
  • crease: 171 / 1099
  • material borders: 247 / 1305
  • intersections: 179 / 921

  Every one of those falls below a gate, so each edge type is load-bearing on the count.

• Flags-off probe. It now turns off every edge type (contour, crease, loose, intersection, material, edge mark), not only contour and crease. The result must still be 0.

• Restore. The rebuilt modifier must recover the exact first-pass counts, not just clear the lower bound.

• Width. radius is 0.021, and the 4.5 thickness is 22. The thickness/radius version-branch check is unchanged.

Gate numbers (5.2.1)

• Framing: fill x=0.475 y=0.844 (band 0.70–0.90) ok; margins left=0.275 right=0.250 bottom=0.050 top=0.106 ok
• Hero drift: mean_abs 0.00334, verdict matches
• Contact sheet:

tile	luma	stage	warmth	sat
gp-lineart-contour (candidate)	0.1995	0.1790	-0.1435	0.1650
armature-bend	0.2787	0.2132	-0.4279	0.1436
damped-track-aim	0.1511	0.1448	+0.2772	0.2305
bmesh-gear	0.2358	0.1772	+0.3027	0.2495

Verdicts:

• Stage darkness: pass. Luma 0.200 is inside the calibration range 0.151–0.279. Under the NPR deviation, luma is information only.
• Wedge warmth: pass. The warm pool is visible top-right behind the tower. The tile's warmth metric is negative because of the blue sea, the same as armature-bend's teal.
• Subject fill: pass, 0.844 vertical.
• Saturation: pass. Saturated red, teal and blue.
• Thumbnail legibility: pass. At 427×240 it reads instantly as an inked lighthouse.

The render is NPR, but it does not look like it belongs to a different gallery: it keeps the same dark stage and the same wedge.

Check runs

binary (path under E:\Blender-Developer-Tools\.scratch\)	reports	result
blender-5.2.1-windows-x64\blender.exe	5.2.1 LTS	PASS: 255 strokes / 1393 points, nosource 0, noflags 0, restored 255/1393
blender-5.1.2-windows-x64\blender.exe	5.1.2	PASS: same counts
blender-4.5.11-windows-x64\blender.exe	4.5.11 LTS	PASS: same counts, thickness=22

tests/smoke/run_example.py passed on all three binaries.

Falsifiers (5.2.1; each was a temporary copy of the script, deleted afterwards)

probe	measured error	exit
--no-contour	use_contour should be True	4
use_material=False in the configured modifier	strokes=247 points=1305 (need >= 240, >= 1320)	5
source "clear" writes the saved source back	cleared source_object still produced strokes=255 points=1393	6
flags-off leaves every type but material borders on	every edge type off still produced strokes=247 points=1305	7
rebuild with use_intersection=False	did not recover after restore: strokes=179 points=921 (first pass 255/1393)	7

Proven by live run vs established by inspection only

Proven by live run:

• all check runs on the three binaries and the smoke runner PASS
• every falsifier above
• the per-edge-type counts
• the framing gate
• render_hero.py, contact_sheet.py and measure_hero_drift.py (matches)
• build_gallery.py, including the alt-length gate
• the broken-state (source_object cleared) render compared with the hero

Established by inspection only:

• the generated <img alt> text in docs/gallery/index.html and docs/gallery/gp-lineart-contour/index.html (read, correct)
• the contact-sheet verdicts are an eye judgement
• CI's floating 4.5.x and 5.2.x point releases have not been run locally

🤖 Generated with Claude Code