fix: wheelbarrow and gltf-skin-roundtrip quality pass

[TMHSDigital]

A quality pass on two items, both taken from the last run's (#248) list of weakest remaining content:

• wheelbarrow (Track A, showcase)
• gltf-skin-roundtrip (Track B, gallery example)

Binaries:

• .scratch/blender-4.5.11-windows-x64 (reports Blender 4.5.11 LTS)
• blender-5.1.2-windows-x64 (Blender 5.1.2)
• blender-5.2.1-windows-x64 (Blender 5.2.1 LTS)

The heroes were rendered on 5.2.1 through scripts/render_hero.py. The inspection renders stayed in .scratch/.

wheelbarrow (Track A)

Inspection sheet: six orthographic views, two three-quarter views, clay, wireframe, six joint close-ups and a ground-contact view.

Defects (round 0):

1. The frame was three sticks. Each shaft was three separate boxes: a handle that stopped at the tray's rear corner, a level run at 0.255 m from the centreline, and a fork from the tray's front corner. The level run sat exactly under the tray's side wall, which was also at 0.255 m, so it was hidden. From outside, the tray carried everything and the barrow read as a crate on legs. Revealed by the front and top orthos, the three-quarter views and the code.
2. Box tray. The sides were vertical and the front square (end orthos, three-quarter views).
3. Abrupt fork. The fork turned in a 50° V over 0.18 m to reach the hub (top ortho).
4. Spreader inside the wheel. The spreader between the forks stood inside the wheel's disc and touched neither fork arm: it spanned ±0.038 m at x = 0.44, where the arms are 0.21 m out (joint_wheel_axle).
5. Strap tabs. The iron straps ran 12 mm above the wall tops (joint_tray_corner, joint_tray_floor, the nubs in ortho_top). They were sized from the floor bottom plus the wall height, but each wall sits WALL_SEAT into the floor.

Generator changes, and why each is structural:

1. Shafts. Each shaft is one mitred sweep through the grip, the tray's rear, its front and the axle, so it reads as one timber.
   • Its plan narrows the whole way: 0.27 m at the grips, 0.19 m and 0.15 m under the tray, 0.038 m at the hub. It therefore shows beneath the tray.
   • The legs, their stretcher, and a bearer under the tray front all hang from it. The bearer's ends are positioned on the shafts' own centrelines, so it replaces the spreader (4).
   • This also fixes the fork (3): the convergence is spread over the whole shaft instead of one short V.
2. Hopper tray. The tray is built from a bottom outline and a top outline: sides flared 15°, the front raked 30° out over the wheel for tipping, the low rear leaning back 12°.
   • Each wall is the planar band between its two edges, so a board's corners follow the geometry.
   • The end boards are housed half a board into the sides, and the sides run past the ends' outer faces.
3. Straps follow the flared wall's outer face and stop 10 mm below its top (5).
4. Audits rebuilt for leaning boards. A leaning board's outer face sits lower than its inner one, so AABB-based tests lie: a Z seam reads negative and a zmin seat reads the outer edge.
   • Seams are now measured along each wall's own up direction, and the seat on each wall's lowest board's inner face.
   • Both directions are taken from the boards' own faces, not the build constants, so the vertical-wall falsifier still reaches its own gate.
5. WALL_SEAT 12 → 8 mm. A raked board's outer bottom edge sits WALL_T·sin(RAKE) = 11 mm below its inner one, so a 12 mm seat would break through the 22 mm floor's underside.

Removed budget: the handle-join budget is gone, because the handle no longer joins anything; it is the shaft. The shaft-run budget replaces it.

New budgets. Every value is identical on 4.5.11, 5.1.2 and 5.2.1:

Budget	Declared	Measured
Shaft runs: wood shells unbroken from behind the tray to past its front	exactly 2	2
Tray flare / rake, from the boards' faces	≥ 10° / ≥ 20°	15.00° / 30.00°

Existing budgets, re-measured:

• Hygiene is 0 on every axis.
• Named shoe and tyre supports: z 0.
• Spoke clearance: 52.1 mm.
• Wall-floor seat: 8.0 mm.
• Metal-wood gap: 0.
• Wall boards: 2 / 2 / 2, with seams of 2.07–2.31 mm.
• Tread aspect: 5.111.

Re-fitted:

• Outer AABB: (1.558, 0.630, 0.574) → (1.563, 0.594, 0.577) m.
• Tray size: (0.720, 0.532, 0.220) → (0.939, 0.578, 0.225) m. The tray is now measured at its flared top edge.

Triangles: base 2536 → 2440, inside the unchanged 2300–2800 band. The collider went from 90 to 80.

Falsifiers. Each new one violates only its own new budget:

• --split-shafts builds the three boxes again: 0 shaft runs.
• --box-tray stands every wall vertical on the flared tray's top outline, so the envelope and tray size stay put: 0° / 0°.

Because exit 17 covers several joint gates, I checked that each 17 falsifier fails on its own gate's message.

	default	--skip-decimate	--stray-vert	--lift-z	--short-legs	--fat-spokes	--float-walls	--pipe-rim	--wide-seams	--split-shafts	--box-tray	--output
4.5.11	0	9	15	16	16	17	17	18	17	20	21	0
5.1.2	0	9	15	16	16	17	17	18	17	20	21	0
5.2.1	0	9	15	16	16	17	17	18	17	20	21	0

Inspection rounds: 1 for geometry. The round-1 sheet showed the flared hopper, the shafts running unbroken under the tray, and every joint closed. The hero read as a wheelbarrow with no further change.

Framing: fill 0.728, minimum margin 0.067, with no deviation=.

Determinism: the default and --output runs on 5.2.1 report identical measured lines. Across binaries, only the LOD2 decimation count (inside its ratio band) and the export byte count differ, as before.

Contact sheet: wheelbarrow.

	before	after
luma	0.172	0.171
stage	0.165	0.164
warmth	+0.068	+0.055
sat	0.132	0.130

It sits inside the band and is legible at card size.

Asset sheet: wheelbarrow. Verdict: passes clearly. The piece is rendered alone and neutrally beside the four pinned asset references. The flared, strapped, raked tray with per-board tone puts it among the more designed objects in the lineup, level with the hydrant.

Convention changes (their own commit, in showcase/README.md):

• "A bent bar is one sweep" now also covers timber that carries a load end to end.
• New: "A tray that tips is a hopper, not a box".
• New: "Measure a leaning board in its own plane".

gltf-skin-roundtrip (Track B)

Defect: the camera looked at both scorpions head-on. The tail curls in each rig's own Y-Z plane, so it was seen end-on: it stood straight up behind the body like a chimney, and the pair read as boxy robots in cone hats.

Not changed, because it is the contract:

• The upright, blocky body is deliberate: a rearing mech scorpion on a base plate.
• Its vertices, weights and posed deformation are what the check asserts, so the mesh is untouched.
• The dark legs are the same asserted mesh.

Presentation change: both armatures are yawed −50° about their own origins and set 1.0 m either side of centre (they were 1.1 m). The tail now arcs over the head in profile, with the claws forward and the legs down the flanks.

• The yaw is composed into each world matrix. The glTF importer leaves its armature in quaternion mode, and my first try, setting rotation_euler, turned only the authored rig.
• It took three rounds: the yaw, the quaternion fix, then the spacing to bring width fill from 0.913 back into band.
• Out-of-band framing: fill 0.828, minimum margin 0.066.

Contract confirmation: the contract, assertions and exit codes are unchanged. I diffed every measured-value line across all 9 runs against the pre-change baseline logs, and there are no differences.

	default	--no-skins	--output
before, 4.5.11 / 5.1.2 / 5.2.1	0 / 0 / 0	5 / 5 / 5	0 / 0 / 0
after, 4.5.11 / 5.1.2 / 5.2.1	0 / 0 / 0	5 / 5 / 5	0 / 0 / 0

Contact sheet: gltf-skin-roundtrip.

	before	after
luma	0.202	0.200
stage	0.193	0.194
warmth	+0.245	+0.217
sat	0.130	0.136

The existing alt text ("…with curled tails and glowing cyan stingers…") now matches what the still shows.

Checks

Run locally and passing:

• check_exit_code_readme.py
• check_gallery_images.py
• check_falsifier_targets.py (229 falsifiers across 28 pieces)

build_gallery.py was run last. I read the generated alt text in docs/gallery/index.html and on the wheelbarrow detail page.

🤖 Generated with Claude Code