feat(mcp): STEP import keeps B-rep instead of baking to mesh

[ecto]

What

import_step tessellated every body on the way in, so an imported part arrived as a triangle soup: no analytic faces, booleans and fillets degraded to mesh operations, and export_cad STEP refused the part it had just imported. Round-tripping a vendor model through vcad was impossible.

Imports now default to B-rep. Each body becomes a lazy step_import node carrying a path plus a solid index, resolved through a kernel-side registry of STEP bytes (the WASM kernel has no filesystem, so the bytes are handed to it explicitly). as_mesh: true keeps the old tessellated import for the cases that actually want a mesh — rendering, physics colliders.

Why the registry needs a rebind hook

The registry lives in WASM memory, so it is empty for any document this process did not import itself. A document-entry hook re-registers a document's STEP sources when it is opened or rehydrated after a restart — otherwise a reopened save would evaluate to nothing. step_import resolution failures are a hard error naming the path, not a silent "no geometry", because a document that quietly evaluates to nothing because its STEP moved is far harder to diagnose.

open_in_browser bakes step_import nodes to meshes for the trip only — the browser cannot resolve a reference into this server's registry and would render an empty part. The session's own document keeps its B-rep.

Two pre-existing bugs this surfaced

1. STEP export of any imported solid panicked. Topology::half_edge_dest asserted next.is_some(), but a STEP import builds both directions of every edge and only loop-used half-edges get next wired — so an edge on a skipped face has a loop-less side. Exporting an imported body died with half-edge has no next. It now falls back to the twin's origin, which is the same vertex the loop would have named. This blocked the CLI too; it was just unreachable through MCP while import was mesh-only.

2. A STEP assembly reads as "no geometry." The flat reader follows the product anchor exactly one hop, so bodies reached via SHAPE_REPRESENTATION_RELATIONSHIP are invisible to it — and vendor files routinely are assemblies. The failure now names the assembly and its part/placement counts instead of claiming the file is empty.

Reviewer notes

• Security: registerDocumentStepSources reads paths out of a caller-supplied document, so it is confined to the same export root import_step writes into — a crafted document cannot turn a reopen into an arbitrary file read.
• The import_step hook is registered once per engine (createServer runs per connection).
• tool-surface.fixture.json is regenerated for the intended import_step description + as_mesh schema change, and nothing else.
• Three artifact-store tests now pass as_mesh: true: they cover report surfacing and large-result offload against stub engines that only implement the mesh entry point. Real B-rep import is covered in step-import.test.ts against the live kernel.
• Generated packages/kernel-wasm/vcad_kernel_wasm* artifacts are deliberately not committed, per CLAUDE.md — the kernel needs a rebuild (or the wasm-refresh workflow on main) for registerStepSource to exist.

Known gap

The named verification target — a SolidWorks RMD-X6-P20-60 vendor file — still cannot be imported end-to-end, because it is an assembly (16 part definitions, 24 placements). This is not a regression: the old mesh import failed on it identically. Real assembly import needs per-instance placement in the IR and world-transform composition through nested instances; placing bodies without transforms would stack them at the origin, so it is left as follow-up rather than half-implemented.

Testing

• cargo test -p vcad-eval -p vcad-ir -p vcad-kernel-step -p vcad-kernel-topo — green, including 3 new tests in crates/vcad-eval/tests/step_import_brep.rs (B-rep survives + can_export_step; solid_index selects the right body; a missing source errors with the path named).
• cargo clippy on all touched crates — clean.
• Full TS suite: 1572 passed. The 6 remaining failures are all in packages/app (last-opened, circuit-build, ecad-feature-inspector, parameter-influence, pcb-interference) and were confirmed failing identically on a stashed clean tree — pre-existing, untouched by this branch.
• STEP round-trip verified against the live kernel: import → boolean → export_cad STEP now succeeds where it previously panicked.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vcad-mcp	Building	Preview	Aug 16, 2026 2:14am

3 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
mecheval	Ignored		Aug 16, 2026 2:14am
vcad	Ignored		Aug 16, 2026 2:14am
vcad-docs	Ignored		Aug 16, 2026 2:14am