Modular Kit Snap

A runnable example building the asset a tiling modular kit lives or dies by: a corridor segment whose open-end boundary vertices sit exactly on the declared tile grid, so instances placed at 4 m multiples share boundary positions with no gap and no overlap. Interior geometry is free-form (beveled detail boxes); only the boundary is snapped — the snap is a deliberate authoring pass, and the check is what catches you skipping it.

The asset, for reuse: a 4 × 3 × 3 m corridor segment (hollow-rectangle shell plus twenty detail parts — frame rib, walkway plate, wall panel assemblies (backing plate + inset panel + four bolt heads each), trim rails, emissive light strips). Origin at the connection pivot (floor-center of the start edge, so instance n places at x = n·4), identity transforms by construction, datablocks under Kit.CorridorSeg.*, manifold everywhere except the two intentional open ends, and every detail part contained strictly inside the tile so it can never break a joint. Drop the hierarchy into a scene and array it along X.

Pipeline arc neighbors: watertight parametric solids in bmesh-gear, topology gates in mesh-hygiene-audit, origin/pivot discipline in prop-origin-transform, collision packaging in collision-hull-proxy.

What it witnesses (all closed form or independently re-derived):

• Snap. Exactly 16 boundary verts (8 per open end, from the hollow-rectangle profile), every one on its end plane x ∈ {0, 4} within 1e-6 m; every boundary edge lies on an end plane.
• Loop coincidence. The two end rings, matched by nearest (y, z) key, agree within 1e-6 — opposing loops are coincident under the tile offset.
• Tiling. A linked duplicate offset by (4, 0, 0) places its start ring at world positions matching the original's end ring within 1e-6 — no gap, no overlap at the joint.
• Bounding box. The shell's local bbox equals the declared tile exactly (0..4 × ±1.5 × 0..3, all float32-exact values, deviation 0.0), and the whole-asset bbox matches it because detail is contained.
• Manifold. Every edge has exactly 2 link faces except the 16 boundary ring edges (1 face each) — the open ends are the only boundaries.
• Reuse hygiene. Identity scales, Kit.CorridorSeg.* names, all part origins at the pivot.

What each check catches on failure (probed with an unsnapped variant built by the same code minus the snap pass — 3 mm end-ring skew, 2 mm y-nudge on two verts): boundary verts off the end planes, worst 3.000e-03 m (exit 3); opposing rings displaced 2.000e-03 (exit 4); tiled joint gap/overlap (exit 5); bbox off the declared tile by 3.000e-03 (exit 6); boundary edges torn off the rim (exit 7); non-watertight detail (exit 8); detail escaping the tile (exit 9); unapplied transforms, default names, or wandering origins (exit 12).

Version witness: check output is byte-identical on Blender 4.5.11 LTS and 5.1.2 — same counts, same zero measured deviations.

Render as proof: a four-segment run — the joints vanish. The falsification variant (--falsify) accumulates a 120 mm gap, 50 mm lateral jogs, and 40 mm floor steps at each joint: floor plates split with dark seams and the trim rails visibly break. The check measures the same failure class at 3 mm; the render exaggerates it to read at frame scale. The render path also gates the asset through examples/gallery_asset_quality.py (naming, material variation, edge treatment — exit 11). An earlier revision shipped flat grey wall panels and was remodeled under that gate: each panel is now an assembly (backing plate + inset panel + four bolt heads), the palette shifted teal-slate so the card does not twin with lightmap-uv-channel's warm cart, and mean luminance was brought into the calibration range (82.7 → 66.1, ceiling 77.7) — no luminance deviation needed. The surface language is procedural and reads Object coordinates: every part's origin is the segment pivot, so painted-panel wear, the walkway's diamond-plate bump and the diagonal hazard stripes on the trim rails are functions of segment-local position — identical in every instance and continuous across a snapped joint, exactly like the geometry.

Framing deviation: the still is an interior corridor run — the envelope surrounds the camera on five sides and the tiling joints are the proof, so the subject reads as extending past the frame. The helper measures; this example enforces a projection-space fill-over cap (FRAME_DEV_MAX = 8.0). Gallery camera scores ~5.06 (pass). --close-camera scores ~30.9 and exits 10. Silhouette saturates at fill 1.0, so the cap is measured with strategy="projection".

Run#

blender --background --python modular_kit_snap.py --
blender --background --python modular_kit_snap.py -- --seam-gap   # falsifier: tiled duplicate 120 mm past the joint; exits 5
blender --background --python modular_kit_snap.py -- --output corridor.png
blender --background --python modular_kit_snap.py -- --falsify seams.png
blender --background --python modular_kit_snap.py -- --output close.png --close-camera

Exit codes#

Per-script sequential checks. 9 is a valid check code; there is no rule against it. 10 is this example's framing-deviation cap (call site, not the shared Layer 1 helper). 11 is the shared asset-quality helper.

The blender-smoke workflow runs the check on Blender 5.2 LTS and 4.5 LTS (5.1 on the weekly cron, the needs-5.1 PR label, or manual dispatch). Smoke does not pass --output, --falsify, or --close-camera. Its catalog falsifier is --seam-gap (expect exit 5): the --falsify render's 120 mm gap and 50 mm jog, applied to the duplicate the joint check measures.