fix: showcase quality pass — stone-well and cart

[TMHSDigital]

Showcase visual quality pass: stone-well + cart

Second quality pass under the showcase visual-quality procedure (first was #170, hay-bale + hitching-post). Both pieces went from "passes budgets" to "actually looks right" via an inspection-sheet review: six orthos, two material three-quarter views, clay pass, wireframe overlay, joint close-ups, and a ground-contact close-up, all rendered headless on 5.2.1 into .scratch/ (untracked).

stone-well

Defects found (view that revealed each)

1. Curb floated ~2.5 cm above the top masonry course — visible daylight ring; front ortho + curb-rim close-up.
2. Windlass stubs ended buried inside the posts — the axle never visibly entered its bearings; windlass-post close-up.
3. Crank was a capped stub — no arm, no grip, unreadable as a windlass crank; three-quarter material view.
4. Roof peak was an open pit — shingle courses met at a hole; top ortho + peak close-up.
5. Bucket was sealed shut (capped cone) and its hoops ignored the wall taper — three-quarter + bucket close-up.
6. Bail handle floated off the bucket rim — legs touched nothing; bucket close-up.
7. Rope ended above the handle, attached to nothing — rope-bucket close-up.
8. Every stone identical — instant CG read in all views.
9. Flat single-color materials, no roughness variation or wear — material three-quarter.
10. Razor edges on all woodwork — clay pass.
11. N-gon caps on every cone primitive — wireframe overlay.

Generator changes (structural, not nudges)

• Curb Z is now derived from MASONRY_TOP (computed from course count × course height), so the curb seats on the masonry no matter how the dimensions change.
• Windlass length extended to pass through both posts and protrude; posts carry real bearing holes by construction.
• Crank rebuilt as an arm + grip assembly off the windlass end.
• Peak cap added at the roof apex, derived from the roof geometry.
• Bucket rebuilt as a manifold open vessel: tapered inner/outer walls, rim ring, floor slab; hoops placed from the taper function so they sit on the wall at their Z.
• Bail handle legs pinned at the rim radius; rope terminated at the bail bar.
• Deterministic per-stone width/radial jitter via a fixed-seed hash01 (no RNG state; identical across reruns — verified).
• Small bevels on structural wood (posts, beam, fascia); cap_tris=True on all cone primitives (n-gons now 0).
• Fascia corner joints changed from exact butt joints to an embedded overlap, eliminating 56 doubled vertices at the corners.
• All three materials gained noise-textured roughness/color variation and edge wear.
• Slot assignment rewritten to be index-preserving. obj.data.materials.clear() resets every polygon.material_index to 0 on 5.2.1 — the piece rendered single-material while the slot-count check still passed. The new per-material face-floor budgets (exit 5) catch this class permanently.

New hygiene budgets (measured 4.5.11 / 5.1.2 / 5.2.1)

Budget	Declared	Measured
Hygiene (loose V/E, non-manifold, zero-area, doubles @1e-5, n-gons)	all 0	0 / 0 / 0 on every axis
Grounded zmin	|zmin| ≤ 1e-4	0.0000 / 0.0000 / 0.0000
Material-island BVH gap (stone↔wood, metal↔wood)	≤ 0.008 m	0.00000 all pairs, all binaries
Material face floors	stone ≥ 3000, wood ≥ 600, metal ≥ 100	met on all binaries

Exit codes (4.5.11 / 5.1.2 / 5.2.1)

• default: 0 / 0 / 0
• --skip-decimate: 9 / 9 / 9 (violates the LOD1 ratio band)
• --lift-z: 16 / 16 / 16 (violates the grounded-zmin budget)

Triangles

Base 8388 → 9380 (crank, peak cap, open bucket, bevels); collider 230 → 306 (ceiling raised 260 → 320 accordingly). LOD2 0.2198 of base on all three series.

Rounds

Two: round 1 found the eleven defects above; round 2 verified every fix on a fresh sheet.

cart

Defects found (view that revealed each)

1. Axle interpenetrated the bed floor — bottom ortho + axle-bed close-up.
2. Shafts poked through the bed floor — bed-interior close-up.
3. Corner plates buried inside the wall corners, invisible — top ortho + bed-interior close-up.
4. Shaft straps floated, aligned to nothing — shaft-root close-up.
5. Back wall same height as side walls — no tailgate read; side ortho.
6. Wheels were torus-profiled — round pipe treads instead of flat iron tyres; wheel close-up + clay pass.
7. Spokes floated at both ends — hub-spoke close-up.
8. Flat single-color materials — material three-quarter.
9. N-gon caps on cylinder primitives — wireframe overlay.

Generator changes (structural, not nudges)

• Bed Z is now derived from the axle: bolsters (real beams) sit on the axle and carry the bed, so the axle can never intersect the floor regardless of dimension edits.
• Shafts dropped to run beside the bed edge and angle down to hitch height; straps aligned to the bed bottom plane.
• Corner plates moved onto the bed top at the corners.
• Back wall rebuilt as a lower tailgate (TAILGATE_H); side and front walls flush at WALL_H.
• Wheels rebuilt with add_ring: flat-tread wooden felloes and flat iron tyres; spokes embedded into hub and felloe at both ends.
• Noise/wear variation on both materials; cap_tris=True everywhere (n-gons now 0).

New hygiene budgets (measured 4.5.11 / 5.1.2 / 5.2.1)

Budget	Declared	Measured
Hygiene (loose V/E, non-manifold, zero-area, doubles @1e-5, n-gons)	all 0	0 / 0 / 0 on every axis
Grounded zmin	|zmin| ≤ 1e-4	0.0000 / 0.0000 / 0.0000
Material-island BVH gap (metal↔wood)	≤ 0.008 m	0.00000 on all binaries
Material face floors	metal ≥ 24, wood ≥ 800	met on all binaries

Exit codes (4.5.11 / 5.1.2 / 5.2.1)

• default: 0 / 0 / 0
• --skip-decimate: 9 / 9 / 9 (violates the LOD1 ratio band)
• --lift-z: 16 / 16 / 16 (violates the grounded-zmin budget)

Triangles

Base 2572 → 2600; collider 306 → 122 (the hull no longer contains the interpenetrating shafts/axle geometry). LOD2 is 0.2200 on 4.5.11/5.1.2 and 0.2154 on 5.2.1 — the ratio band absorbs the cross-series DECIMATE difference, as documented.

Rounds

Two: round 1 found the nine defects above; round 2 verified every fix on a fresh sheet.

Convention change

showcase/README.md gains the material face-floor convention: every declared material asserts a named minimum polygon count recomputed from polygon.material_index, and slot assignment must be index-preserving (never materials.clear() + rebuild, which wipes indices to 0 on some versions while the slot count still passes). Existing hygiene/grounded/gap conventions already covered the rest of this pass.

Evidence notes

• All exit codes and measured values above are from live headless runs on .scratch/blender-4.5.11-windows-x64, .scratch/blender-5.1.2-windows-x64, and .scratch/blender-5.2.1-windows-x64 (binary-reported versions).
• Determinism: stone-well default run twice on 5.2.1, measured lines byte-identical.
• Heroes re-rendered on 5.2.1; check_framing passes both (well: fill 0.861, min margin 0.033; cart: fill 0.806, min margin 0.044). No deviation= used.
• Gallery regenerated via scripts/build_gallery.py; generated HTML for both detail pages and the index cards read back and verified (witness strings, alt text, × entities). The rebuild also picks up stale iron-cauldron detail-page drift whose README landed without a gallery regen.
• Inspection sheets live in .scratch/inspect/ (untracked, per procedure).