User-reported regressions on the 1:1 legacy ports: slab/ceiling moves
were slow (polygon CSG rebuilds per scene.update tick), fence moves
teleported briefly on commit (residual mesh.position offset survived
the geometry rebuild). All three now use the same live-drag pattern
the legacy fence move was designed for:
- During drag, write only to `sceneRegistry.nodes.get(id).position`
+ `useLiveTransforms`. No `scene.update`, no polygon rebuild, no
React re-render of geometry.
- History stays UNPAUSED — scene state isn't changing.
- On commit, a single `scene.update` writes the translated
polygon (or fence start/end + linked-fence cascade). Recorded as one
natural undo step.
- Tools leave `mesh.position` at the drag delta on commit;
GeometrySystem / CeilingSystem reset it to (0,0,0) when they
rebuild the geometry on the next frame. By the time position
clears, the new geometry is in place — no teleport.
Two framework changes enable this:
- `GeometrySystem` (viewer/systems/geometry) now resets
`group.position` + `group.rotation` after every rebuild, matching
the legacy `FenceSystem.updateFenceGeometry` behavior. Tools that
translate the group during live-drag can rely on the reset.
- Legacy `CeilingSystem.updateCeilingGeometry` extends its existing
`position.y` reset to cover X/Z too — previously it left X/Z at the
drag delta after rebuild, double-translating the visual.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lands the three-checkbox composition runtime documented in
wiki/architecture/node-definitions.md. A kind with only a pure
geometry function now needs zero per-kind React or system code.
Type-side additions (packages/core/src/registry/types.ts):
- New `GeometryContext` (resolve / children / siblings / parent) — read-
only scene access for builders that reference other nodes by ID
(wall miters, door cutouts). Most kinds ignore it.
- New `geometry?: (node, ctx) => Object3D` field on NodeDefinition,
independent of renderer/system. Three orthogonal opt-ins replace the
v0 RendererSource union.
- Re-exported via packages/core/src/registry/index.ts (consumed by
nodes packages through `export * from './registry'`).
Runtime (packages/viewer):
- New <GeometrySystem> (systems/geometry/geometry-system.tsx) walks
dirtyNodes, builds a GeometryContext per dirty node, calls
def.geometry, disposes old children, attaches new ones, clearDirty.
Frame priority 2 (matches the priority shelf's per-kind system had).
Mounted in viewer/index.tsx alongside <RegisteredSystems>.
- New <ParametricNodeRenderer> (components/renderers/parametric-node-
renderer.tsx) — empty <group> + useRegistry + useNodeEvents +
markDirty-on-mount + useLiveTransforms. Mounts hosted children via
<NodeRenderer> recursively. The default renderer for any registered
kind without a custom def.renderer.
- <NodeRenderer> dispatch updated: custom renderer wins, else
geometry-only kinds fall through to ParametricNodeRenderer, else
null (legacy switch fallback). Documented inline.
Shelf migration (proof of the boilerplate collapse):
- Deleted nodes/src/shelf/renderer.tsx (was 45 lines of registry +
handler boilerplate).
- Deleted nodes/src/shelf/system.tsx (was 60 lines of dirty-loop +
dispose plumbing).
- shelfDefinition now: `geometry: buildShelfGeometry`. One line.
buildShelfGeometry is the pure function from geometry.ts that already
existed.
End-to-end effect: registry-driven shelf now mounts via the framework's
generic renderer + system. Parametric edits flow through the same
dirty-driven rebuild path, but the kind ships ~100 fewer lines of
boilerplate. Every future kind that fits the same shape (item, fence
segment, column, etc. as they migrate in Phase 5) follows the same
"one line, one pure function" pattern.
Wall stays on its dedicated def.renderer + def.system — its mitering
needs level-batch context (`ctx.levelData?.miters`, future extension)
that the generic system doesn't yet provide. Decided at Phase 3+, not
blocking Phase 4 acceptance.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wassim SAMAD