Own the whole export pipeline, colour and mesh repair included

[revarbat]

Export was split between two implementations: this repo wrote STL/OBJ/OFF/3MF as flat geometry, while BelfrySCAD's exporters.py held the actual truth about colour and mesh repair in ~400 lines of Python the CLI never saw. They disagreed:

	before (this repo)	real OpenSCAD
cube(100); cube(100,center=true); → 3MF	2 objects, 24 tris	1 object, 36 tris
OBJ	flat v/f, no o, no usemtl, no .mtl	—
per-triangle colour	dropped	—

Ported from exporters.py, behaviour-for-behaviour

• splitBodiesForExport — the implicit top-level union, cut into objects that never share volume: one per colour (later shape wins an overlap, painter's order), then one per connected component.
• Per-triangle colour, which is what 3MF's model is built for — the spec is explicit that colour describes the surface, not the distribution of material through the volume — so a body whose surface came out of a multi-colour CSG merge is written faithfully with no volume split.
• New writers: PLY, VRML97, X3D 3.3 (Interchange profile), ASCII STL, and OBJ with a companion .mtl.
• exportModel() — one entry point, format from the extension, returning warnings rather than logging them, and carrying the repair policy the GUI used to apply itself (sliver stripping + mesh check on the merged mesh; per-body checks for multi-object formats). Open shells keep their triangles rather than being dropped.

Two rules came across intact that look like details and are not:

• A per-triangle colour array indexes the triangle list it was built against, so it is carried only when the result's triangles compare equal — not by assuming which paths are no-ops.
• The per-colour claim skips the subtraction when bounding boxes cannot overlap, because A - disjoint B returns A's volume but reorders its triangle list, silently destroying those colours.

Python gets a handle, not another array round-trip

bodyToDict flattens every Manifold into numpy for the renderer, which is all the renderer needs — but export has to do real CSG. Evaluator stashes an opaque Geometry handle as .geometry, alongside csg_tree/profile_result, so evaluate()'s own 2-tuple result is unchanged and render-only callers never see it.

bodies, id_to_node = ev.evaluate(path)      # unchanged
warnings = oce.export_model("m.3mf", ev.geometry)

Measured honestly: this is not faster today. Like-for-like on Dalek (224k triangles) it is 1045 ms against Python's 976 ms. The ~146 ms rebuild saving is real in isolation but swamped by the mesh checks and the split, which dominate and which both sides hand to the same Manifold library. The handle is kept for the architecture and the provenance, not for a speedup it does not deliver.

Verification

• Output identical to the Python implementation across all six formats on a two-tone-plus-transparent model: triangle counts, object counts, per-triangle colour indices, material counts.
• The union bug now matches real OpenSCAD exactly (36 triangles, 1 object).
• 1481 tests pass, green across three consecutive -j8 runs — after fixing a test-isolation bug my new CLI format cases introduced (they wrote and deleted a shared cube.scad, so under ctest -j they destroyed each other's input).

CLI now reaches every format: -o out.{stl,obj,off,3mf,ply,wrl,x3d}, plus --ascii-stl.

🤖 Generated with Claude Code