showcase

Showcase

Budget-conformance props. Not examples.

An example witnesses one API contract and carries a falsifier that makes a real assertion fail. A recognizable crate witnesses no API contract. Forcing one into examples/ produces a vacuous check. Showcase pieces assert that generated geometry meets declared asset budgets.

"It rendered without error" is not an assertion. A tolerance so wide nothing can violate it is not an assertion.

This directory is a sibling of examples/, not nested under it. The manifest key is showcase. Counts are separate from the example total.

Conventions

Every piece is a directory showcase/<name>/ with a script, a README that includes an exit-code table, a falsifier, a catalog.json row, a gallery entry in showcase/gallery.json, and a rendered still.

• Deterministic. Fixed seed (or no RNG). Identical output across runs on the same binary, and across Blender 4.5, 5.1, and 5.2. If a value legitimately cannot match across versions, the piece README names it, states a tolerance, and justifies it. DECIMATE COLLAPSE triangle counts are the usual suspect — prefer ratio bands, not exact counts.

• Budgets declared in the script as named constants and documented in the piece README. Suggested axes: triangle count, material count, UV bounds, bounding-box dimensions, LOD ratios, collider triangle ceiling, export file written.

• Assertions recompute those budgets from the generated result. They never restate constants the script set (if n == DECLARED where n was assigned DECLARED is not a check).

• Falsifier breaks one pipeline stage so a named budget fails and the piece exits its documented code. Prove default and falsifier on 4.5.11, 5.1.2, and 5.2.1. --skip-decimate is the LOD-ratio falsifier (exit 9), --stray-vert the mesh-hygiene falsifier (exit 15), --lift-z the grounded-zmin falsifier (exit 16), and --fat-rungs (or the piece's equivalent) the joint-fit falsifier (exit 17), and --round-band (or equivalent) the seat-conformance falsifier (exit 18). --twin-sole duplicates a ground plate so the coplanar-face budget fails (exit 15). --short-stile / --short-post lift a member out of its cup (exit 18 or 19). --clip-ring pulls a hung ring off the eye centerline (exit 18). --flush-tyre / --stand-posts restore a flush surface so the coplanar budget fails (exit 15). --float-wheel / --float-stone lift one named support while the rest still ground the AABB (exit 16). --sink-tyre / --shallow-tenon put a seat outside its band (exit 18). --drop-bucket hides the hung subject (exit 18). --skew-wheel / --turn-posts break a mirror or placement budget (exit 19). --float-rivets lifts fasteners off their host (exit 18). --round-haft turns an oval section round (exit 19). --float-nails lifts nail heads off their strap (exit 18). --short-mortar stops every mortar joint shy of its stones (exit 18). --open-ends restores open U-band ends so the water is not contained (exit 18). --pile-rocks draws scattered stones together so they interpenetrate (file-local code; terrain-scatter uses 20). --straddle-handle mounts a fixture across a board gap (file-local code; shipping-crate uses 20). --odd-loaf breaks a body's own mirror symmetry (exit 19). --lift-lid-bands floats bands off a curved host (exit 18). --sharp-iron skips a chamfer pass so the edge-treatment budget fails (file-local code; crate-stack uses 21). A budget with no falsifier witnesses nothing: prove each one fails once, and check the exit code, not just non-zero.

• A falsifier must fail the budget it targets. Declare the target in the piece README's falsifier table — flag, target budget, exit code — and make the run exit that code. A falsifier that trips an earlier check is red for the wrong reason and proves nothing about its target. Fix a collision by changing the model or the falsifier; never widen a band so an ill-aimed falsifier lands. stone-archway's --flat-arch laid the voussoirs as a lintel, which is 0.62 m shorter and so failed the bounding box (exit 8) instead of the intrados circle fit (exit 19); --off-circle replaced it by keeping the whole envelope and wandering only the radius. A falsifier that thickens a member scales only the axis its budget measures: chopping-block's --fat-haft scaled the whole haft section, and once the section became an oval the swing-plane width grew the knob 16 mm past the bounding box, so it exited 8 instead of 17 on eye clearance. Scaling across the cheeks alone restored exit 17 with the identical measured clearance. tests/check_falsifier_targets.py checks the declaration statically, and with --run BLENDER executes each falsifier and asserts the observed exit matches.

• Hygiene budgets. Copied combinatorics from examples/mesh-hygiene-audit (do not import the example). Every piece asserts on the generated mesh: non-manifold edges 0, loose verts 0, loose edges 0, doubles at 1e-5 0, zero-area faces 0, n-gons 0, world AABB min Z within 1e-4 of 0, and coplanar disjoint face pairs 0. Count the last one over faces that share no vertex: the triangles of one flat fan cap are coplanar and close-centred by construction, and counting those makes the budget unsatisfiable rather than meaningful. Count the coplanar pairs cross-shell, not merely over faces that share no vertex: two quads two steps apart on one flat cap share no vertex and are coplanar and close-centred by construction. Counting those reported 96 pairs on hay-bale, none of them a hazard, which is the same unsatisfiable-budget trap in a second disguise. Z-fighting is two separate bodies landing on one plane, and that is a cross-shell pair by definition. Multi-body props do not require Euler characteristic 2 — that is a single-shell contract. Two boxes that share a coplanar face z-fight; two that share a vertex weld into one shell. Seat a joint with a named overlap or rebate computed from the host (a cage post into its rail, a glass pane behind its muntin, a plinth step into its base) rather than hoping the numbers happen not to coincide. A hang-station versus arm-station split lets a contact falsifier move one member without dragging the rest of the assembly. Pairs of parts meant to touch assert a BVH surface gap below a named epsilon (vert-vert is the wrong metric for thin straps and collars). A spanning board whose only vertices sit at the far ends will report a false gap to a mid-span knuckle; put a local shell at the joint station, or measure a slice there.

• Named supports (exit 16). AABB zmin is necessary but not sufficient on a multi-support prop: a wheel, a sled floor, or one planted foot can ground the box while the other feet float. Split the mesh into shells and assert each named support (shoe, tyre, foot) has its own zmin within epsilon of 0. --short-legs (or the piece's equivalent) is the falsifier: float the named supports while leaving something else on the ground so the AABB gate would still pass.

• Joint-fit budgets. Parts that interpenetrate on purpose — a tenon in a mortise, a peg in a hub — assert how deep the overlap goes, not just that it exists. Split the mesh into shells by edge connectivity, assert the expected shell count, and for each joint assert three named minima recomputed from vertex positions: the inserted part stays clear of the host's chamfer, it seats far enough past the host's near face, and it stops short of the host's far face. A part exactly as thick as its host passes a "parts overlap" test while erupting through the host's chamfer as a spike — that is the class this catches. Measure in the construction frame (un-rotate by any rake) so a tilted host does not inflate its own AABB.

• Diagonal from stations (exit 17). A brace, gooseneck or scroll is an oriented primitive or a tube between named host endpoints, not a hypot-length box rotated about its centroid. The centroid form is short of one host and punches through the other. --short-brace / --float-spout (or the piece's equivalent) is the falsifier.

• Even shaping terms and mirror symmetry (exit 19). Where a prop has mirrored members — two belts, two brackets, a pair of feet — every closed-form term in the host's shaping function must be even in the mirrored axis. hay-bale used sin(x*24), which is odd, so its two belt stations sampled different surface heights and the wraps shipped 5.4 mm out of step; cos fixed it exactly. Assert it: pair the mirrored shells and compare their axis positions and extents within a named epsilon. A bounding box cannot see this, and neither can a per-member budget that only ever looks at one member.

• Wrappers follow the host's profile, and the host is finished first (exit 16/18). Build, bevel and ground the host, then hang the wrapper on the finished surface by raycasting the host's own cross-section at the wrapper's station. Placing a wrapper against half-built geometry and shifting everything afterwards leaves each one a different distance off the floor. A constant-section rectangle on a rounded host stands proud at the middle of each face and is swallowed at the corners — hay-bale's belts stopped 2.9–8.3 mm above the floor with a visible notch, while the AABB zmin gate stayed green because the loaf grounded the box. Assert per wrapper shell that the loop passes under the host, not just that something touches Z=0.

• Seat conformance (exit 18). A band, hoop, strap or collar wrapped around a host asserts a banded seat depth — a minimum so it cannot float and a maximum so it cannot sink — sampled per angular segment rather than as a single global figure. Derive the wrapper's profile from the host's own radius function instead of a circle, and classify inner versus outer vertices against the host surface at each vertex's own angle. A global midpoint radius misclassifies outer chamfer vertices as inner ones the moment the host is out of round, which is how a hoop that visibly gapped on one side still passed. The paired falsifier makes the wrapper a true circle on an out-of-round host. Measure the depth station-locally, along the vertex's own direction from the station axis — not by nearest-surface distance. A wrapper seated in a concave waist has its nearest host face on the bulge shoulder at a neighbouring station, and the wrap then reads as sitting 10 mm outside a host it is in fact hugging. Bin the samples by rounding to the nearest station, never by flooring: a station on a bin boundary puts its inner and outer rings in different bins, and the inner-only bin reports the wrapper's outer offset as its seat depth.

• Level sole on a raked leg (exit 18). A raked stile does not sit in a world-axis cube. Build the sleeve in the member frame so it follows the rake; add the tread after the rake as its own axis-aligned plate, from Z=0 up to a height that covers the tilted sleeve. The member end stays inside the sleeve by a named bite and above the tread. --short-stile starts the member above the sleeve so the bite fails while the AABB zmin still passes.

• Hung ring (exit 18). A hitching ring is a torus whose center is the eye center plus the hang direction times the major radius, so the eye lies on the ring centerline, and the tube is thinner than the eye's inner radius. The shank stops in the top of the eye tube; continuing it to the eye center runs the pin through the ring. A torus parked on a face is not a hung ring. --clip-ring lifts the ring off that centerline. --short-post starts the post above the shoe cup so the seated-post budget fails (exit 19) while the sole still grounds the AABB.

• Sampled host seat (exit 17/18 on scatter). Instanced scatter seats by sampling the host surface, then clamping instance verts above the slab floor. A closed-form icosphere (or equivalent) replaces the GN cube so the scatter is stones, not open crates; a per-shell face floor is the budget that catches a cube leftover (exit 19). --poke-rock skips the clamp and over-bites; --float-rocks seats above the host.

• A seated part is sealed all round (file-local code). Sampling the host once, at the part's centre, and biting the lowest point below it leaves daylight under the downhill side of anything on a slope: terrain-scatter's stones were sealed in at most 6 of 8 sectors, two in only 4, and the gap showed in the ground-contact view. Split the azimuth around each part's centroid into sectors, sample the host under every vertex (a ray down onto the host alone), and sink the part until each sector's most-buried vertex is below the ground. Assert every part is sealed in every sector. The falsifier restores the centre-sample rule (--perch-rocks).

• Orthogonal members (exit 19). Boards, planks or arms that should sit on a world axis assert their AABB secondary extent stays within a named bound of the member thickness. A closed-form yaw of 0.12 rad reads as an accident, not weathering. --yaw-boards is the falsifier. Fingerboard tips are a wedge whose base is the board's own thickness×height, bitten into the board so the base face is inside it — not a 45° cube of a different size glued on the end.

• Wall-mount zmin (exit 16). A wall sconce plants the plaque bottom at Z=0. Lifting every vert by a hanging height so the asset floats in its own space is not the same as placing it on a wall in the engine.

• Plumb and real-world size (exit 19). Assert that the axis of a turned or lofted body is vertical, by comparing the XY centroid of a bottom slab against a top slab — not the exact zmin and zmax rings, which may be a handful of vertices once a rim is notched or chipped. Separately assert the body's own diameter and height against the dimensions the README states in metres, with a named tolerance. The outer AABB does not cover this: on a prop with an appendage the AABB is the appendage, and the body can drift to any size underneath it.

• A band is hooped onto its host, never set flush against it (exit 15/18). A tyre, ferrule, hoop or collar derives its inner radius from the host's outer radius minus a named interference, and its outer radius from the host plus its own thickness. Writing r_mid = host_out + t/2, radial_t = t/2 — the obvious spelling — makes the band's inner cylinder and the host's tread the same surface, so every segment is a coplanar cross-shell pair around the whole circumference. cart shipped that way and measured 32 pairs (16 per wheel); the speckle was visible on the committed hero and in a clay pass, and no budget could see it. --flush-tyre restores the equality and is the falsifier.

• A member is tenoned into its seat, never stood on it (exit 15/18). A post, leg or stile whose bottom face lands exactly on its host's top face puts both on one plane. Derive the member's bottom from the host's top minus a named seat depth, and hold the member's top fixed so nothing above it moves. stone-well stood its four roof posts on the coping and measured 4 pairs, one per post. Assert the seat as a band recomputed against the host surface read off the generated mesh — max z over the host's material — not against the constant the builder used, which witnesses nothing.

• Posts go under the roof's corners, not the middle of its eaves (exit 19). Where a hip or pyramid roof is carried on four posts, assert each post's plan bearing against a hip corner recomputed from the generated mesh, as a wrapped angular difference. Take the corners from the pooled vertices at the eave line: pulling them from a single shell picks one fascia board, whose own extremes sit 90° off the corners it is nailed to, and the budget then fails on a correct model. stone-well placed its posts on the axes, which cantilevered the roof's corners 0.80 m and stood one post dead centre in the well mouth in every orthographic view.

• A roof is sized from what it must cover, not from what carries it. Deriving the eave reach from the post ring plus an overhang gave stone-well a 1.64 m roof over a 1.08 m drum — an umbrella. Derive it from the covered body's own radius plus a named clearance (EAVE_HALF = R_OUTER + CURB_OUT + EAVE_CLEAR), and let the overhang fall out of that. Expect the convex-hull collider budget to move when the roof does; re-fit it and say so.

• The subject hangs where it can be seen (exit 18). A prop whose story is one small part — a bucket on a rope, a lantern on a hook — asserts that part's clearance above the body it hangs over, as a band, measured against the body read off the mesh. stone-well's bucket sat down the shaft with only its rim level with the coping: invisible in the hero and in all six orthographic views, and no budget noticed. --drop-bucket is the falsifier.

• Mirrored assemblies (exit 19). Where a prop has a left and a right of the same part — two wheels, two brackets — pair the shells and assert they match in the two axes they share and in their extents, within a named epsilon, and are opposite in the mirrored axis. Size the falsifier's displacement to stay inside BBOX_TOL so the AABB gate cannot steal the failure: cart's --skew-wheel moves one wheel 6 mm along a track whose tolerance is 10 mm.

• Segment counts are a silhouette budget, not a triangle budget. A 16-gon felloe reads as a polygon at hero size, and the flat facet facing the key light renders as a hard white plate. cart went to 24 and the chords disappeared. Re-fit the triangle band around the new measured count rather than leaving the old one — and prefer a band narrower than the one it replaces, centred on the measurement.

• Material face floors. Every declared material asserts a named minimum face count on the finished mesh, recomputed from polygon.material_index. This catches the slot-assignment wipe class: obj.data.materials.clear() resets every polygon's material_index to 0 on some versions, which renders the piece single-material while the slot count still passes. Assign slots by index-preserving append/replace, never clear-and-rebuild.

• Exit codes are file-local: 0 success, argparse 2, 3 and above in check order. 9 is legal. FATAL sys.exit(1) is a crash, never a named check. 15–19 are reserved across pieces for the hygiene family above: 15 hygiene, 16 grounded, 17 joint fit, 18 seat conformance and contact, 19 plumb and real-world size.

• Shading is part of the model. A flat-shaded low-poly body and a smooth-shaded one are different objects to a viewer, so decide per part and say why. Facet what is faceted in life — a chamfered axe head, a dressed stone. Do not facet an organic or turned surface with equal facets: chopping-block's flat-shaded 36-gon log read as coopered staves and the whole block as a tub, and its flat-shaded iron hoop threw one highlight per facet, a row of piano keys. Smooth those and carry the surface — bark furrows, growth rings, drying checks — in the material, while the wobble keeps reading in the silhouette. Mark every material boundary as a hard edge so a sawn rim stays crisp against the bark. Nothing here is measurable, which is exactly why it has to be looked at.

• One substance, one material slot. A part made of something else gets its own slot and its own face floor, even when the colours are close. chopping-block's hickory haft shared the bark slot, which made the handle the darkest wood on the piece and would have fissured it with the bark texture.

• Fasteners are aimed down the host's surface normal (exit 18). A rivet, bolt head or stud on an out-of-round host takes its axis from the normal of the host's own surface function at its station — not the radial — and its station is a host section, so the surface under it is exact rather than a chord. Assert the seat per fastener as a band measured radially at each vertex's own angle against the host read off the mesh, plus a proud minimum so the head is visible. chopping-block's radially aimed rivets seated 1.87 and 2.82 mm against a 1.5 mm design bite, one edge sunk and the other lifted, because the log's wobble drops the surface 1.4 mm across one head. Aimed down the normal both seat at 1.50–1.54 mm. --float-rivets is the falsifier.

• Edge treatment: no right angles (file-local code). Real objects have chamfers that catch light. Where every box in a piece is chamfered, count manifold edges whose faces meet within 5° of 90° and assert 0 — a one-segment chamfer turns every 90° edge into two 45° ones, so a survivor is a bevel pass that was skipped. Chamfer thin stock at its own offset: crate-stack's 3.5 mm iron takes 0.8 mm, where the timber's 1.8 mm would leave no flat. Pass material= to bmesh.ops.bevel: left at its default the chamfer faces took slot 0, and the iron plates rendered — and were classified — as timber. --sharp-iron is the falsifier. 15–19 are reserved, so the code is the piece's next free one.

• A rim over a slope is bounded by its tilt step (file-local code). The count above assumes box edges. Where a curved or sloping surface turns down into a wall — a terrain tile's edge, a mound's skirt — the fold runs from well under 90° to well over it, so bevel it with enough segments and bound the largest step in normal elevation between adjacent faces. Use elevation, not the dihedral: at a tile's corners two chamfer strips meet at a right angle in plan, which is a rounded corner, not a knife. Bevel gives the middle folds twice the end ones, so two segments leave 45° of a 90° edge in one step. The falsifier skips the bevel (--sharp-rim).

• An L-section is one shell, not two boxes (exit 15). Two overlapping boxes for the legs of an angle strap share the outer corner edge; chamfer them and both lay a strip on the same line — a coplanar cross-shell pair at every corner (crate-stack: 12). Extrude the L profile once, each cap as two convex quads meeting on the inner-corner diagonal, and skip that flat diagonal in the bevel.

• Classify small parts by a size derived from the host. A nail, rivet or stud told apart from its plate by world-AABB extent needs a threshold taken from the plate (IRON_WRAP * 0.5), not from the fastener: a yawed 10 mm head has a world AABB wider than 10 mm, and a fixed 10 mm threshold silently dropped 8 of 48 nails.

• A joint is filled, not open (exit 18). Masonry, brick and stacked-stone pieces separate their blocks with a named joint, and a joint left as air is a stack of floating blocks with daylight through every course. stone-archway shipped that way: 7 mm bed joints and 9.7 mm voussoir joints, none of them touching. Fill each joint with its own mortar shell, sized from the two blocks it sits between, recessed behind their faces by more than the chamfer so it reads as a raked joint, and biting a named depth into both. Assert every mortar shell BVH-overlaps exactly two stones — one means a bed left hanging, none means the air gap came back. Keep the mortar shells out of any classifier that names stone parts; left in, a bed classifies as a course and every downstream stone budget measures the wrong neighbour. Add mortar after the chamfer pass, unchamfered: a 6 mm chamfer on a 10 mm bed eats it.

• A vessel holds its contents (exit 18). A trough, tub, barrel or bucket that shows a liquid must close every side of it. Touching the hull is not holding: water-trough passed a water-to-hull gap budget for a whole release with open U-band ends and the water's flat end face on show. Assert containment by casting rays outward from across each exposed face of the contents — perimeter and interior points, since a rim band covers the perimeter only — and require every ray to hit the vessel within a named reach. --open-ends is the falsifier.

• Scattered parts do not interpenetrate (file-local code). Jittered scatter pushes neighbours into each other, and nothing else in the hygiene family compares one scattered part with another. BVH-test every pair and assert 0 overlaps. Space the parts structurally: relax each pair of centres apart to the sum of the two parts' own radius bounds, each derived from the same constants that size that part, so scaling a part widens the spacing around it. A falsifier that only skips the relaxation proves nothing if the parts happen to miss anyway; terrain-scatter's --pile-rocks also draws the scatter inward so collisions are guaranteed.

• Scatter varies in size (file-local code). Nine stones within 20% of one size, on a relaxed grid, read as planted. Give each scatter point a size factor (boulders among cobbles), and assert the ratio of the largest to the smallest footprint clears a floor. The falsifier sets every factor to 1 (--uniform-rocks).

• Assert a body's symmetry on the body (exit 19). Where mirrored parts are hung on a body, the body itself must be mirror-symmetric: give every vertex a partner at its mirror position within a named epsilon (a KD-tree lookup). A budget on the hung parts sees an odd shaping term only second-hand. Triangulate non-planar quads along their short diagonals before raycasting a body or shipping it: a fixed split is not mirror-symmetric, and hay-bale sampled surfaces 5.6 mm apart at mirrored stations on vertices that matched to 71 um. Apply an odd-term falsifier after any angle-selected bevel, or the asymmetry changes which edges the bevel picks and the bounding box fires first.

• Bands on a curved host are built on the host's arc (exit 18). A band on a vault, drum or hoop is one slab on the host's own radius function, with the host's segment count, not a chain of boxes rotated per segment. Rotating a box to a circle's tangent at angle t about X needs -t when the arc is parameterised as (sin t, cos t), and treasure-chest's +t fanned every segment out like feathers. Assert the band radially in the host's construction frame, un-rotating any hinge swing and undoing any re-centring from a part known to be symmetric.

• A fixture is mounted on a member, not across a gap (file-local code). Handles, hinges, hasps and plates screwed to planked faces must land on one board. A height fixed as a fraction of the frame lands wherever the plank layout happens to put a gap: shipping-crate's handle plates sat across the slot between two end boards. Snap the fixture to the centre of the nearest board, from the same layout function that places the boards, and assert each mounting plate sits inside a single board's extent with a named clearance at both edges.

• Rock is broken, not smooth. A smooth-shaded ellipsoid is an egg. Cleave it with a few closed-form planes (project every vertex beyond a plane onto it) and shade it flat, so it reads as broken stone. Cleaving takes volume off, so re-check scale against the host afterwards.

• Keep a falsifier's envelope still. A falsifier that moves the support the piece grounds on re-grounds the whole piece and moves the AABB with it. water-trough's --short-legs lifted the shoes; the piece re-grounded on the legs and shrank 10 mm against a 10 mm tolerance, so a 1.5 mm model change tipped it from exit 16 to exit 8. Make the falsifier move only what its budget measures — the shoes float, the legs stretch to the floor — and check its AABB delta against the tolerance, not just its exit code. Re-check it whenever a design change moves which part sets the envelope. hay-bale's ridge dropped from 14 mm to 6 mm, which made the knot loops the top of the box. The twine falsifiers, sized at 20 mm and 10 mm when the loaf's crest was the top, then lifted the knots past BBOX_TOL and exited 8. The fix was to size each one to its own budget, 8 mm against a 6 mm wrap gate and 6 mm against a 2 mm seat floor, never to widen the tolerance.

• Stone is not noise stretched over a block. A noise-driven colour mix sampled so its features streak read as wood grain on stone-archway at hero scale. Stone wants isotropic object-space mottling, fine speckle driving roughness and a small bump, and a seeded tone per block as a face attribute so no two stones match. Chain a baked normal map under that bump rather than replacing it.

• Identical boards read as CG. Planks cut from one material are one plank repeated. Give each shell a seeded tone and its own grain direction — its long axis, recovered from its vertices — as face attributes, and have the shader stretch its grain along that axis rather than a world axis, which a yawed member is off by its yaw. Nothing here is an assertion; it is found on the inspection sheet. The same holds for coopered staves: wooden-barrel and wooden-bucket were one flat tone until each stave got its own. Hoops are forged iron, dark and rusted, not the bright polished metal both pieces first shipped with.

• A head or bottom is boards, not a slab (exit 17). A barrel head or bucket bottom cut from one disk reads as a lid. Cut the croze outline across into boards with a named seam, so every board's outer edge is still the croze and the seat does not change. Assert the board count and that every seam lies in a band. Classify boards by their span in plan, not by a radius: a quarter of the head's width is smaller than any "round" test built for the whole disk. --one-piece-head / --one-piece-bottom are the falsifiers.

  The same holds for any end that closes a vessel. water-trough's ends were one D-shaped slab each; they are three vertical boards. Where the boards must also hold water, lap them a millimetre into each other rather than leaving a seam: a through gap leaks, and lets a containment ray through, while the bevel on the boards' face edges still shows the joint. Read any rim test on the union of an end's boards, or a board's inner edge at a joint is taken for the rim (--slab-ends).

• Rope is laid, not piped. A smooth tube with a rope material is a plastic hose. Loft a three-lobed section and turn it one vertex step per ring, so each lobe winds along the path as a strand. That costs the same triangles as a round section of the same count. Fix the section's frame to the path's plane, or the lay jumps where the path turns vertical.

• What passes through a part must clear its mount (exit 17). A bail through an ear ring has to clear the plate the ring hangs on. Ring offsets sized to the ring (RING_MAJOR * 0.25) put the rope's axis 3 mm off the plate and its body through it. Size the offset from the thing passing through, and assert that no vertex of it lies inside the mount. --sunk-rope restores the old offset.

• Variation goes into surface, never into function (exit 19). A ladder's rung heights were jittered up to ±18 mm for "variation"; a climber's feet find rungs blind, so the pitch is part of what a ladder is. Jitter tone, grain and a few percent of turning, not spacing that something uses. Assert the pitch (every gap within a tolerance of the mean); --drift-rungs restores the old heights.

• A plank wall is boards, and extra boards are paid for (exit 17). A tray wall cut from one slab beside a planked floor reads as a box. Stack it from boards with a named seam, let the audits that find walls take the union of a wall's boards, and assert count and seam. If that breaks the triangle ceiling, take the triangles back from faces nobody sees (bottoms buried in a seat, undersides facing the chassis) rather than raising the ceiling. The falsifier keeps the triangles (--wide-seams), so it cannot trip the triangle floor instead.

• Sort bmesh operator inputs. A Python set of BMEdges iterates in memory-address order. Handed to bmesh.ops.bevel it gives the same geometry in a different face order from run to run, which the mat_index_counts print exposed on wooden-ladder. Call bm.edges.index_update() and sort by index first.

• A band wraps its host; it is never sunk into it (exit 18). hitching-post's bands were sized from the post's inner offset, so they sat inside the post and only the chamfered corners broke the surface, as black slits. Size a band from the host's face: inner face a named bite inside it, outer face proud of it. Assert the band's outer half-width against the host's (--sunk-bands). An overlap test alone passes a sunk band, because a sunk band overlaps too.

• A forging is not a turning (exit 19). A lathed, elliptical waist under a rectangular anvil face reads as a funnel. Loft forged parts from a squared section (a superellipse) and assert it: corner reach, how far a section vertex fills its bounding rectangle's corner, is 0.707 for an ellipse. --round-waist is the falsifier.

• Paint wears through to what is under it. A painted board is one flat colour until a mask lets the wood show at the edges and pores; shade the paint per board.

• Iron is not chrome. Cast and forged iron is near-black or rusted and rough. Metallic 1.0 with roughness under 0.4 reads as polished steel or plastic (hand-pump read as chrome).

• The stage is bigger than the frame. A 14 m set lets the back wall's edge into the corner of a three-quarter hero. Use 60 m.

• A leaning prop is staged against something. A ladder raked 12° in open air reads as falling. Render-only: put a wall section on the side it rakes toward, with its face through the top, and turn the piece so that side faces away from the camera.

• A ring is threaded across its hole, never in its plane (exit 17). iron-cauldron's ear rings stood in the bail's own plane, so the bail ran through each ring's tube. Orient a ring so its hole axis follows whatever passes through it at that point. Assert that no ring vertex lies inside the thing passing through, by the nearest face's normal, and that it passes within the hole. --edge-on-ears is the falsifier.

• A vessel has a base and a belly (exit 19). A profile interpolated from a tiny bottom radius runs to a point: the cauldron was an onion. Build the belly from a stated base radius and belly height, offset the inner wall along the profile's normal (a sideways offset thins the wall to nothing where it runs flat), and assert the base radius and where the widest ring sits. --pointed-pot is the falsifier.

• Stones are lumps, and no two match (file-local code). A curved brick with a flat top reads as a kerb. Loft a chamfered section that swells mid-stone and pinches at capped ends, with seeded radial offset and a crown on the top course. Keep the beds flat so courses still seat. Assert the spread of stone radii (and top-course crowns) clears a floor; the falsifier gives every stone the same draw at the jitter's upper bound, so the envelope stays put (--uniform-stones).

• A section that carries the read is a budget (exit 19). Where a cross-section is what makes a part recognisable — an axe handle is oval, a broom handle is round — assert it as a ratio band at a named station, measured in the construction frame from a slab that holds exactly one ring. --round-haft is the falsifier.

• A brace joins two members (exit 17). A diagonal that stops in the air is a peg, not a brace. Assert that both ends reach past the inner face of the member each tenons into, measured per side, and count the braces so a classifier that silently drops one (a level brace fails a "diagonal" shape test) cannot pass vacuously. The falsifier stops one end short inside the envelope (--short-brace).

• A joint bites; touching is not joining (file-local code or 17). A BVH gap of zero is satisfied by one grazing bevel corner, which reads as a loose stick. Where a member tenons into another, assert the deepest vertex of the member inside the host shell (signed distance to the host's surface), at every end. Run it after plumb so a raked host reports as plumb. --shallow-brace (street-lantern) restores the grazing ends.

• A platform bears on something (exit 17). Every deck, floor or shelf needs bearers that tenon into the frame, and every joist must sit in a bearer at both ends. Count the bearers and joists so a dropped member cannot pass vacuously, and aim the falsifier at the joist span with the bearers kept, so the triangle floor and envelope do not steal it (--short-joists, watchtower).

• A bent bar is one sweep (file-local code). Wrought iron that bends from a foot into a post, or from a post into a back, is one bar. Sweep it along one path with parallel-transported frames. Do not stand a post on a separate foot arc. park-bench built its scroll feet that way, stopped each arc 8% short, and started each post at the arc's top. That left 14 mm of daylight under all four posts. A BVH gap budget can't see this, because the toe still grounds and the post still meets its stretcher. Assert it from shells instead: the shell under each toe station must be a leg that reaches the seat. --split-feet is the falsifier.

  The same holds for timber that carries a load end to end. A barrow's shaft is the handle, the bearer under the tray and the fork, in one piece. wheelbarrow built it as three boxes and hid the middle one exactly under the tray's side wall, so the handle appeared to stop at the tray corner and the tray to carry everything. Sweep it through its stations with mitred joints, route it where it can be seen, and assert the count of shells that span the whole run (--split-shafts).

• A tray that tips is a hopper, not a box (file-local code). Build a tray, bin or trough that is emptied by tipping from a bottom outline and a top outline: flared sides, a raked tipping end. Each wall is the planar band between its two edges, the end boards housed into the sides. Measure the lean from the boards' own largest faces and assert a floor; the falsifier stands the walls vertical on the same top outline, so the envelope and size budgets stay put (--box-tray).

• Measure a leaning board in its own plane. A flared or raked board's outer face sits lower than its inner one, so AABB tests lie: a seam in Z between two stacked boards reads negative, and a zmin seat reads the outer bottom edge. Take each wall's up direction from its boards' largest face and measure seams along it; seat the wall on its lowest board's inner face. Derive both from the mesh, not the build constants, or the box-tray falsifier trips the seam gate first.

• A carried member bites its carrier, measured where the vertices are (file-local code). A slat on a rail, a back slat on an upright and an armrest on its leg each assert an overlap depth against a named minimum, and a count of the members, so a dropped slat cannot pass vacuously. A round rail has vertices only at its end rings, so the deepest rail vertex inside a mid-span slat is nowhere near the slat. Take the overlap either way round: under a slat, the slat's corner sits in the rail. Where neither member has a vertex at the crossing — a brace over a rail, mid-span — read both faces off the mesh and compare the planes, only where their heights overlap. --float-slats and --float-brace are the falsifiers.

• A kit section fits its tile (file-local code). A section that claims to tile at a pitch asserts its width along that axis is at most the tile and within a named band of it. Derive the post stations from the tile minus the reach of whatever is proudest, not from the post's own half width. fence-kit placed its posts from the post width alone. Its iron bands then stood 12 mm past each end, so adjacent copies interpenetrated by 24 mm, while its README said the width was the tile. --wide-tile restores the old stations. Check it before the AABB gate, or the envelope check fires first.

• A fixture is bolted to the face it is seen on (file-local code). A wall plate seated on the backing slab, behind the face of the stones laid over it, is buried: wall-torch showed only its four bolt heads. A plate-to-plaque gap budget passed, because the plate still touched the slab. Derive the seat from the proudest face (stone_face_y()), and assert two things from the mesh: the fixture stands proud of that face by a named minimum, and its back bites the face within a band. --sink-plate restores the old seat.

• A ferrule grips the leg on the leg's axis (exit 18/20). A vertical cup on the floor under a splayed leg can't hold it. The leg either saws through the wall or, as tavern-stool shipped, stops above an open ring with air around the foot. Build the sleeve on the leg's own axis, with its inner radius a named grip inside the leg. Bury its raked bottom in a level tread, and end the leg inside the sleeve above the tread. Measure the grip radially about the leg's principal axis, not by nearest-face signed distance: points in an open sleeve's hollow read as inside its wall.

• Bake texels per UV cell (file-local code). One UV cell per face and a 256 px bake gives a 1,700-face piece about 6 px per cell. The render's bilinear lookup then reads the next cell's normals across the border, as dark slivers on the hero (tavern-stool). Assert the smallest UV cell's extent in baked texels, from the mesh's UVs and the image's actual size, against a floor (12 px). --low-bake bakes at 256 px and is the falsifier.

• Level on the stage. A hero that tilts the piece about X, even by 2°, sinks one row of feet into the floor and lifts the other. Turn the piece only about Z.

• Laid stone and cut timber vary piece by piece. One flat grey on every stone reads as a moulded ring; one brown on every member reads as a single casting. Drive a per-shell tone (PlankTone) and, for wood, grain along each shell's long axis (GrainDir) from face attributes, seeded so the render is the same every run.

• One connected assembly (exit 18, file-local). Every joint in a piece is a bite, so joined members' surfaces cross. Build a BVH per shell, union the shells whose trees overlap, and assert one component. Per-part budgets pass a box resting a hair above its rails, because every plank still touches a wall. The contact graph shows the box as a second component. cart's --lift-bed raises planks, walls and straps 7 mm and splits the graph in two.

• A z-fight budget matches planes, not centres (exit 15). A test that counts only faces whose centres coincide within 0.1 mm never fires on a real overlap. grindstone reported 0 for a whole release while its iron shoes, exactly as wide as the sill and flush with its end, shared three planes with every sill (a lit slit in each foot close-up). Use the cross-shell coplanar budget above, with a KD-tree range query at COPLANAR_CENTRE_MAX so it stays linear on a 5k-triangle piece. A shoe, cap or ferrule is a cup a named reveal proud of its host on every side; --flush-shoes restores the flush shoe.

• Braces that meet tenon into the host's faces (exit 15). Two diagonals run to the same centre line meet at one apex. Their chamfered end faces then land on each other inside the host: 6 coplanar pairs on grindstone, hidden in the bearing block. End each one a named bite inside the king post's face.

• Carried parts bite their bearers (exit 18). A tub, bed or shelf asserts how deep it bites each member that carries it, as a band, with the bearers counted. Stretchers left at sill height stopped 6 mm under grindstone's trough, a dark slit that the sill seat and the dip band both passed. --low-stretchers restores them.

• Chamfer n-gon caps, then triangulate. A triangle-fan cap on a lathe, chamfered 5 mm at its rim, folded its centre out through the cap: 8 non-manifold edges per nave on cart. An L-strap cap pre-split on its inner-corner diagonal left 24 per strap once the reflex corner was chamfered. Build caps as single n-gons, run the bevel passes, then bmesh.ops.triangulate every face over four corners, so the shipped mesh still has no n-gon.

• The bake cage is narrower than the nearest neighbour. cage_extrusion only has to cover the difference between high and low, under 2 mm for a chamfer. At 0.08 m, cart's cage reached past the 74 mm between a side wall and its wheel. Rays from the wall hit the high-poly felloe, and the rim baked onto the board as a black staircase that no budget measures. Raising the bake resolution made it sharper, not smaller. Start from 0.01 m and look at the hero.

• Aim an edge falsifier at one member. Skipping a whole chamfer pass moves the triangle count by hundreds: cart's whole-iron skip dropped 1200 and exited 4 on the triangle floor. Leave one small member square instead (--sharp-bar, --sharp-handle, 20–24 right-angle edges and about 48 triangles), so only the edge budget can see it.

• Rendered still and gallery entry. Showcase pieces are visual by definition. The pathology / sidecar exemption does not apply. Call examples/gallery_framing.check_framing on the --output path only. The deviation= parameter is gone; bleed compositions call measure_framing_deviation and assert at the call site. Do not move or modify gallery_framing.py — import it by resolving the repo root (see the shipping-crate script).

• Contact sheet (required). Composite the candidate hero beside the pinned calibration set — canonical membership is in CLAUDE.md § Quality Gates — commit it as docs/gallery/contact-sheets/<name>-contact-sheet.webp, link it in the PR body, and report a per-criterion verdict: stage darkness, wedge warmth, subject fill, saturation, thumbnail legibility, plus mean luminance against the calibration band. A claim without the committed composite is not evidence.

  Required because it has caught a real defect in showcase work. The first sheets for crate-stack and stone-archway showed both wedge pools reading as cool grey bands rather than the warm pool the house style calls for; both were relit as a result. Nothing else in the pipeline looks at the still beside its peers, so nothing else could have seen it.

• Asset sheet (required). Render the hero alone — neutral three-quarter view, plain studio lighting, no staging tricks, no labels, no comparison props — composite it beside the pinned asset-quality reference set rendered the same way, commit under docs/gallery/asset-sheets/, and report a verdict. The piece ships only if it is not identifiable as the least-designed object in that lineup.

  Required because showcase is entirely game props, which is exactly the scope docs/VISUAL-STYLE.md § Asset quality names, and because it covers something no other gate here does. Budgets measure geometry conformance; the contact sheet measures staged presentation. Neither removes the scene, and a strong scene carries a weak model. The recorded evidence is socket-attach-points: it passed every measurable floor on its first draft — edge90 0.000, ten materials, no default datablock names — and was then judged bad by eye and rebuilt from scratch. The floors scored the bevels, not the design.

  examples/gallery_asset_quality.check_asset_quality returns 11 on violation, the same call pattern as gallery_framing. Showcase numbering already spends 11 on the collider-triangle ceiling, so remap the return at the call site rather than letting two budgets share a code.

• Composition. The README names which shipped skills and snippets the piece composes. Duplicated helpers stay inlined or copied; showcase scripts do not import snippets as a package.

Layout

showcase/
  README.md              # this file
  gallery.json           # this tree's gallery index (pieces[])
  <name>/
    <name>.py
    README.md
    preview.webp

Hero stills live at docs/gallery/assets/<name>-hero.webp like examples. scripts/build_gallery.py merges showcase/gallery.json into the same docs/gallery/ site as examples, tagged showcase.

Smoke

tests/smoke/catalog.json takes opaque script paths. A showcase row is enough; blender-smoke.yml has no path filter and runs the whole catalog on every PR. Measure wall-clock before adding the next piece.