Part-kit assets — manifest-driven multi-part assembly

The higher-order layer on the art pipeline for assets that are kits of parts (vehicles from chassis/wheels/doors; procedural buildings) rather than one fused mesh.

The contract: the generator emits a manifest.json next to the meshes, and code consumes ONLY the manifest — never inferred render bounds.

1. Pivots at joints

Author each part with its pivot AT its joint: wheels at hub centre, doors on the hinge line, bumpers at the mount plane, buildings at footprint-centre@ground. Detach/hinge/spin later = transform math around an origin that is already correct.

2. Manifest records real-world meters

Per part, in real-world meters: the backing vmdl path, a semantic kind, pivot_semantics, rotation_axis_local, dims_m, local bounds, the attach point in the AUTHORING frame (attach_author_m), and a suggested mass_fraction (unused until parts detach and need their own rigidbodies).

For buildings the same idea for placement: footprint, collision height, front axis, door centre + door-clearance rectangle. Generate a C# catalog from it so code never parses paths at runtime.

3. Kit vmdls are render-only

No physics shapes on the kit meshes. Code composes simple BoxColliders from dims_m / footprint+height at assembly time: box colliders sized from the manifest are cheaper and cleaner than per-poly hulls, and never drift from the placement contract. Do NOT infer collision from decorative render bounds and do NOT add an indiscriminate concave collider — see model-no-collision and scaled-collider-didnt-scale.

4. Position through ONE tested frame mapping

Position from attach_author_m through a single, empirically proven author→model-local mapping. The instructive failure: a mapping AND yaw that are both 180° off cancel in POSITION and add in ORIENTATION — every pivot lands world-correct while every mesh sits spun in place. So the conversion constants and the facing yaw must always flip together, and position-correctness alone proves nothing about frame-correctness — verify facing with a screenshot + an asymmetric-feature probe. See model-imported-wrong-facing.

5. Assembly is transactional

Preload/validate every required model BEFORE creating any GameObject; on any required-part failure create NOTHING and fall back to the fused/blockout path — a broken kit must never brick a spawn or ship a half-assembled compound collider. Distinguish required (structural/collision-bearing) from optional (cosmetic) kinds. Validate the manifest strictly at load (schema version, finite length-3 arrays, min ≤ max bounds, exactly the expected wheel/part set) so a parseable-but-truncated JSON is rejected at TryLoad, not thrown at spawn.

6. Mirror the validator offline

A dependency-free twin of the C# rules (python or otherwise), run against the generator's output + malformed fixtures, with a LOCKSTEP comment on both sides — the generator gets a CI gate without a headless C# test runner.

7. Loose JSON under Assets/

Readable via FileSystem.Mounted in-editor (Json.Deserialize with DTO property names byte-matching the JSON keys, zero serializer attributes). Whether loose json ships in a packaged build is a packaging question to re-verify — dev-host streaming to a joiner does NOT send loose files.

8. Animated sub-parts mount on live simulation state

Reuse the existing visual component: put animated channels on an identity-based "Visual" GO under the physics part, with kit-specific alignment as a static child below it — the shared visual stays byte-identical for fused and kit paths.

9. Version the manifest schema

Normalize old versions at load. Additive optional fields need no bump. Same append-only discipline as every other spec in the house — and the same spirit as delta-log save and spec-replication networking.

Units reminder: author in meters, convert once at the engine boundary — sbox-units-are-inches.