geo/Merge

By default the inputs simply CONCATENATE: N objects in, N objects out, each keeping its own transform, material, groups and parenting. The object groups of every input are unioned, and so are the stage attributes (first input to name one wins).

With `compact` on, the node instead collapses the meshes it receives into a SINGLE mesh PER GEOMETRY TYPE: the point sets and index buffers are concatenated into one geometry, with each source's matrix baked into its points. Only meshes of the same type join: triangles read their index buffer three at a time, segments two at a time, so triangles and segments coming in together give one merged triangle mesh and one merged segment mesh, in the order those types first appear. That is what a renderer wants (one draw call instead of N), and what the mesh operators that work on a single geometry want. The price is that per-object data cannot survive a collapse: transforms, materials, per-object hierarchy and object groups are all gone. Stage attributes still forward.

SDF objects compact too, as a hard UNION of their fields, ONE PER DIMENSION: 2D and 3D fields are different geometry types and cannot become one object, so an input carrying both gives a 3D union and a 2D union. SDF objects cannot join a mesh either, so a compact merge carrying meshes and both SDF dimensions gives three objects: the 3D union, the 2D union, then the merged mesh.

Only geometry is compacted. Lights, cameras and controls pass straight through, and volumes are rejected: their per-voxel data lives on a grid, with no point set to concatenate.