Bookshelf

A bookcase: a stained carcass with two shelves housed in its sides, a base on a recessed plinth, lapped back boards and a stepped cornice, filled with thirty-eight hardbacks. Twenty-nine stand upright, seven lie in three stacks, and two lean on a neighbour. Every book is a rounded-spine cover in cloth or leather, a page block glued into it, and two gilt bands tooled round the spine. A showcase piece, not an example — it witnesses no API contract. It asserts that generated geometry meets declared asset budgets, recomputed from the finished mesh.

What it composes#

The budget that matters#

A leaning book has to rest on its neighbour's head edge and on the shelf, not in either. Tip it a few degrees too far and its board cuts through the neighbour's corner; not far enough and it stands in the air against nothing. Neither is visible to any other budget:

• the leaning book still stands on its shelf, so the seat budget passes;
• it stays inside the carcass, so the bounding box does not move;
• the triangle count is the same, because only a position changed;
• the book-overlap budget has to exempt this very contact, so it cannot see how deep the contact goes.

The piece finds the leaning books on the finished mesh (a book whose largest face, a board, is tilted between 3° and 80°) and the upright neighbour on the side its top leans toward, on the same shelf. It takes the plane of the leaning board's outer face off the mesh and measures how far the neighbour's deepest vertex sits behind it, as a band of 0.3–2.0 mm. Both lean at 0.80 mm.

A plane, not the cover solid: the board is 2.8 mm thick, and a head edge driven right through it lies in the hollow of the book, outside the cover again, so a solid test reads a deep cut as no contact at all. That is how the first --deep-lean passed.

--air-lean and --deep-lean are the falsifiers built for exactly this. One stops each leaning book 3 mm short of its neighbour, the other drives it 4.5 mm in. Both exit 22 while every seat, the envelope and the book overlap budget still pass.

Budgets#

Declared in the script as named constants, recomputed from the generated mesh. Measured values are from Blender 5.2.1; see the cross-version table below.

Real-world size: 0.97 m wide over the cornice, 1.18 m tall and 0.33 m deep. Each compartment is 318–348 mm clear, and the books run from a 186 mm octavo to a 305 mm folio.

Construction#

• Carcass. Two full-height sides carry everything and are the only named supports. The plinth, the base and the two shelves are each housed DADO (8 mm) into both sides and stop 14 mm short of the outer faces. The plinth stands 18 mm back from the base's front and 4 mm off the floor, so only the sides ground. The top is seated over the sides and the back boards; the cornice steps 10 mm out past it and is seated CORN_BITE into it, its back held 4 mm short of the top's back so the two do not share a plane. Every board is chamfered at 2.5 mm with material= pinned to wood, edges sorted by index.
• Back boards. Five vertical boards, lapped 2 mm into one another and every other one stepped forward 1.5 mm, as on apothecary-shelf.
• Covers. One shell per book: a U of front board, half-round spine and back board, extruded up the height (SPINE_ARC = 6 segments). The boards are tangent to the spine, so it reads round and the boards flat. Every edge over 40° is chamfered 0.8 mm; the spine facets (30°) are not, and shade smooth.
• Page blocks. A block inside the U, SQUARE (4 mm) short of the cover at head, tail and fore-edge, biting PAGE_BITE (0.6 mm) into both boards. Its back stops 1.5 mm short of the line where boards meet spine, which is where the band ends lie.
• Gilt bands. Two per book, built on the cover's own outer chain round the spine: every band vertex sits on a cover vertex's normal, inner face 0.3 mm inside the spine and outer face 0.8 mm proud.
• Standing books. Set along each shelf with a 2.5–4.5 mm gap, spines a stepped 6–14 mm back from the shelf front, each seated a stepped 1.0–3.3 mm into the shelf. Each stands a degree or two off square, in a four-step cycle (−2.4°, −0.8°, 0.8°, 2.4°).
• Stacks. Books laid on a board, spine out, yawed ±3° alternately and shifted a few millimetres. The bottom one seats into the shelf; each one above takes a 1.0–1.4 mm seat into the board below it.
• Leaning books. Turned about Y by 18° (bottom shelf, onto the book on its left) and 22° (middle shelf, onto the book on its right), seated on the shelf by the lowest vertex of the chamfered cover. Then the leaning book (or, leaning right, its neighbour) is slid along the shelf until the neighbour's deepest vertex is LEAN_BITE (0.8 mm) behind the leaning board's outer face. Each leaning book is taller than its neighbour's head divided by the cosine of its lean, so the head edge meets the board and not its top corner.

Findings the budgets forced#

• Coplanar page blocks and shelves. The first page blocks sat 3 mm short of the cover's tail, and one book seated 2.98 mm into its shelf put its page block's underside 0.02 mm off the shelf top: a coplanar pair. The square is now 4 mm, clear of every seat.
• Every spine has the same facet angles. A half-round spine of six facets has the same six facet normals on every book. Square to the shelf, two neighbours share a facet plane whenever their offsets happen to line up, and the uniform and crowded falsifiers did so (five and twelve coplanar pairs). Each upright now stands a degree or two off square in a four-step cycle. A half-degree cycle was not enough: on a 45° chamfer the normals differ by half the yaw, inside the 1e-4 parallel tolerance.
• Planes facing the front. Spine lines, band ends and fore-edges face the shelf front, and two neighbours landed band ends on one plane. Each book is stepped back 0.9 mm at a time until none of its front-facing planes is within 0.4 mm of a neighbour's.
• A stacked book on a page block's plane. A book stacked on another with the running seat (up to 3.3 mm) could land its underside on the plane of the lower book's page block, 2.2 mm inside its board. A stacked book now takes its own 1.0–1.4 mm seat.
• Matching parts to books. Page blocks were first matched to the cover nearest their centroid. In a stack, the centroid of a thick book's page block is nearer the board of the book above than its own. Parts are now matched on the mean distance of all their vertices.
• Deep lean through the board. Measured against the cover solid, a head edge driven 4.5 mm into a 2.8 mm board came out in the hollow of the leaning book and read as no contact, so --deep-lean exited 0. The contact is now measured against the plane of the board's outer face.
• Black gilt. Fully metallic bands facing the camera mirrored the dark stage and read black except at their edges. The gilt is half metallic.

Conventions walked#

Every convention in showcase/README.md, and whether it applies here.

Falsifiers#

Each breaks one pipeline stage so a named budget fails. All fifteen were run on 4.5.11, 5.1.2 and 5.2.1 and exited the same declared code on all three.

Exit codes#

File-local and sequential. 9 is a valid check code. 1 is the FATAL wrapper — a crash, never a named check. 19 (plumb and real-world size) is reserved across pieces and unused here.

Run it#

# Budget check, no render.
blender --background --python bookshelf.py --

# Falsifier: each leaning book stops short of its neighbour. Must exit 22.
blender --background --python bookshelf.py -- --air-lean

# Falsifier: each leaning book cuts into its neighbour. Must exit 22.
blender --background --python bookshelf.py -- --deep-lean

# Render the gallery still (EEVEE; --engine cycles on a GPU-less host).
blender --background --python bookshelf.py -- --output bookshelf.webp

Smoke runs the check-only path. It does not pass --output or any falsifier.

Cross-version measurements#

LOD2 differs by ten triangles on 5.2 (DECIMATE COLLAPSE, the known divergence; the gate is a ratio band) and the glTF size by 8 bytes (exporter metadata, not a budget). Every other value is identical.