[revarbat]

An AMF object whose triangles don't share vertex indices -- a triangle soup, or a duplicated vertex list -- imported as an open surface ("N boundary edges, M unwelded vertices") that could not take part in CSG. OpenSCAD 2026.02.01 merges an object's vertices at exactly the same position, so the same file imports closed there, including a solid with a through-hole (genus 1).

Fix

loadAmf now reads each object into its own LoadedMesh, runs the STL importer's existing weldVertices on it, then appends it with the index offset: exact-only and per object, as OpenSCAD does (objects are never welded to each other). weldVertices moves out of mesh_import.cpp's anonymous namespace into mesh_import.hpp. CLAUDE.md's AMF note is corrected (unit is ignored, as in OpenSCAD) and describes the weld.

What OpenSCAD does (from running the binary only)

• Exact-position duplicates within an object: merged → closed (soup cube/sphere/ring, duplicated list, far from origin; a triangle degenerate after welding is dropped silently).
• Across objects: no merging.
• Genuinely open meshes stay open; nothing is patched.
• Near-coincident vertices: not merged by import (only by OpenSCAD's generic CSG-time repair, ≤ ~3e-7, same for polyhedron() -- a separate item).

Verification

• vs OpenSCAD: every exact-duplicate case now matches its volume and vertex/face counts; open meshes, near-coincident jitter and closed controls unchanged.
• Regression: 148 files (30 seeded models exported as STL/OFF/3MF/AMF/OBJ) imported by old and new and re-exported to OFF and 3MF: all byte-identical, identical messages.
• 3 new gtests in tests/test_import_export.cpp (2 fail on the old code; the third pins the per-object boundary). Full suite: 1471 passed.
• All corpus runs under a timeout + 1.5 GB RSS watchdog; no runaways.

🤖 Generated with Claude Code