fix: seat the ladder rungs and stop the iron chamfers reading as wood

[TMHSDigital]

Visual quality pass on showcase/wooden-ladder. Four inspection rounds. Everything below was proven by live run on three local binaries unless labelled otherwise.

Binaries, versions as the binaries report them:

• .scratch/blender-4.5.11-windows-x64/blender.exe → Blender 4.5.11 LTS
• .scratch/blender-5.1.2-windows-x64/blender.exe → Blender 5.1.2
• .scratch/blender-5.2.1-windows-x64/blender.exe → Blender 5.2.1 LTS

Defect list

Inspection sheet: six orthos, two three-quarters, clay, wireframe, twelve rung-joint close-ups, two top-joint, two foot-joint, one ground-contact, 1024², rendered to .scratch/ladder-inspection/round-{0..4}/ and never committed.

Rounds 0 and 1 landed in the prior session; rounds 2 to 4 are this one.

#	Part	Defect	Revealed by
1	Stile top	A second wood block sat on top of each stile as a stub, reading as a broken-off extension	joint_top_*, three_quarter_a
2	Iron fittings	Four fittings per stile (base collar, two side straps, top band) clustered into an unreadable mass and protruded past the silhouette as loose tabs	ortho_front, committed hero
3	Feet	No shoe: the raked stile met the floor on a bare angled end, so contact was a line, not a pad	ground_contact
4	Whole piece	zmin was forced to 0 by clamping stray vertices upward, deforming the foot instead of grounding the piece	ortho_left, ground_contact
5	Rungs	All six rungs identical in radius and perfectly even in spacing, reading as CG	ortho_front
6	Stiles	Bevel applied via every wood vertex, so the rung cap rims were beveled too and shaded broad faces as faceted	clay pass
7	Rungs / stiles	Rung diameter equalled the full 34 mm stile depth, so each dowel broke through the stile's front and back chamfer as a thin spike	joint_rung_* close-ups, rounds 1 and 2
8	Iron fittings	Every chamfer face on the caps and shoes rendered as wood, framing each black fitting in tan	joint_top_left, joint_foot_left, ground_contact, round 2
9	Whole piece	The piece failed its own gates at exit 4 — round 1 removed geometry without re-fitting the triangle band or outer AABB	default run, round 2
10	Hero set	The wall's left edge sat in frame as a bright vertical band in the corner	committed hero and re-render
11	Falsifier	--lift-z exited 0 — the grounded budget could not fail	falsifier matrix, round 3

Defects 1 to 6 were fixed in the prior session's round 1. Defects 7 to 11 are this PR.

What changed in the generator, and why it is structural

7 — rung diameter. RUNG_R 0.017 → 0.013, a 26 mm dowel in a 34 mm stile. Structural because it corrects a proportion that was wrong: the dowel was as deep as the member it passes through, so no material, camera angle, or chamfer tweak could stop it erupting. The clearance is now 3.61 mm against a 2.5 mm chamfer, so the dowel lands on the stile's flat face with margin rather than grazing the chamfer.

8 — chamfer material. bmesh.ops.bevel does not inherit material_index onto the faces it creates, and the piece claimed its metal faces from a face-set snapshot taken before the bevel. Every chamfer therefore fell back to slot 0. Now claimed from the bevel op's own return["faces"]. Structural because it derives the face set from the operation that created it, rather than from a snapshot that is correct only if bevel never adds faces.

9 — stale fitted constants. BASE_TRIS 2100–2280 → 900–975 and OUTER_SIZE (0.474, 0.361, 1.482) → (0.472, 0.365, 1.460), both re-fitted to the measured result. Not a loosened assertion: the band is the same ±4% relative width, and the triangle count fell because geometry was removed, not because a gate was widened.

10 — set edge. Floor and wall grid 14 → 60. Structural because it removes the class (no edge of the set can enter frame at any framing) rather than nudging the wall out of this one shot.

11 — vacuous falsifier. world_bbox reads matrix_world, which is evaluated data; the object was moved with no view-layer update, so the cached matrix hid the lift. Added bpy.context.view_layer.update() before measuring.

Also replaced the bevel's wood_verts[:16] slice with an explicit stile_verts list. Identical geometry — the slice happened to be the two stiles — but it no longer depends on the stiles being constructed first and contributing exactly sixteen vertices.

New budgets, measured on all three binaries

All three agree exactly except glTF byte count.

Budget	Declared	4.5.11	5.1.2	5.2.1
Shell count	exactly 12	12	12	12
Rung depth clearance	≥ 0.003 m	0.00361	0.00361	0.00361
Tenon engagement	≥ 0.006 m	0.01200	0.01200	0.01200
Tenon breakout margin	≥ 0.005 m	0.04600	0.04600	0.04600
Wood faces	≥ 280	324	324	324
Metal faces	≥ 180	216	216	216
Non-manifold / loose v / loose e / doubles @1e-5 / zero-area / n-gons	0	all 0	all 0	all 0
Grounded abs(zmin)	≤ 1e-4	0.0000	0.0000	0.0000
Base triangles	900–975	936	936	936
Outer AABB	(0.472, 0.365, 1.460) ± 0.01	(0.4720, 0.3649, 1.4597)	same	same
Collider tris	≤ 120	24	24	24
glTF bytes	> 0	75140	75140	75128

The joint budgets are measured per rung end against each stile from vertex positions, un-raked into the construction frame so the 12° rake does not inflate the stile's own AABB, and reported as the worst case over all twelve ends.

Exit codes, all three binaries

Invocation	Budget violated	4.5.11	5.1.2	5.2.1
default	—	0	0	0
--skip-decimate	LOD1 ratio band	9	9	9
--stray-vert	mesh hygiene, loose verts	15	15	15
--lift-z	grounded zmin	16	16	16
--fat-rungs	rung depth clearance	17	17	17

Measured failures: LOD1 ratio 1.0000, loose_v=1, zmin 0.050000, clearance -0.00051.

--fat-rungs widens the dowels to the full stile depth, reproducing defect 7 exactly, so the assertion is proven against the real defect rather than a synthetic one.

--stray-vert and --fat-rungs are new. --lift-z existed but exited 0 on all three binaries before this PR.

Falsification of the material-face floor

The floor that locks defect 8 in was proven to fire by breaking the fix: dropping metal_faces.update(bevelled["faces"]) gives mat_index_counts={0: 516, 1: 24} and exits 5 with metal faces 24 < 180 on 5.2.1. Restored.

Triangle counts

2184 → 936. Cause is round 1's fitting removal, not hidden-face culling. The rung end rims stay capped and buried inside the stiles — 144 triangles nobody sees — because an open rim is a non-manifold boundary and the hygiene budget forbids it.

Hero still

Re-rendered on 5.2.1. Framing gate passes without deviation=: fill x=0.166 y=0.839 in band 0.70–0.90, margins left 0.419 right 0.416 bottom 0.072 top 0.089 against a 0.020 minimum, no edge touched or crossed.

Conventions

showcase/README.md gains the joint-fit budget class (shell count plus the three named minima, measured in the construction frame) and a named falsifier for every hygiene budget, which previously had none. Every future piece inherits both.

Not done

No contact sheet or asset sheet. Those gates cover examples/; no showcase piece has ever carried one, including the four prior showcase quality passes (#167, #169, #170, #171). Showcase pieces are governed by showcase/README.md instead.