Nesting — cut optimisation with a true-shape engine

Less offcut per sheet is money, on every sheet. Imports the workshop's real CAD output, nests it, and produces the cut list, the labels and the machine files.

Live: https://gervdalius-droid.github.io/nesting/

Why this exists

Sheet material is the biggest consumable in the shop. Commercial nesting is sold per seat per month and still doesn't read our file formats without an export dance. This reads what SWOOD actually produces and nests it in the browser.

What it does

Imports what the shop already has — a SWOOD project (.xlsm / .xlsx), a CSV part list, or shaped parts straight from DXF. Boards are auto-created from the project's materials. It reads the PARTLIST sheet (SA_GA_VIRT / RIMANTE_ZAB formats) and picks up edgebands and QR operation IDs from the LABEL sheet automatically.

Two nesting modes, deliberately different:

	Engine	For
Optimize & Cut	rectangular strip packing	panels — the 95 % case, fast
Shape nest	jagua-rs compiled to WebAssembly	irregular / curved parts from DXF

The true-shape engine is Rust compiled with wasm-bindgen-rayon, so it runs multi-threaded in Web Workers behind SharedArrayBuffer (which is why the page ships coi-serviceworker.js to get itself cross-origin isolated on GitHub Pages).

Outputs the shop actually uses — cut list, Zebra LP2824 labels with QR codes (scan a part at a station and it's tracked), and DXF export for vCarve.

Part groups keep parts that share a grain pattern together, and offcuts/scrap can be fed back in from the warehouse so a usable remnant gets used instead of re-cut from a full sheet.

Things I learned the hard way

• Finer rotation steps make the result worse. Intuitively more angles should pack tighter; in practice the extra candidates starve the left-bottom-fill heuristic. Coarser steps plus more random restarts beat finer steps.
• Demand grouping pays for the search. Collapsing identical parts into a demand count frees enough time budget to run a best-of-many seed search, which is where the real gain is.
• nestPoly must stay full-resolution. Simplifying the polygon before nesting loses exactly the concavities that let parts interlock.
• Adding a DXF entity type touches seven places — and the easy one to miss is the commit() whitelist, which silently drops the entity instead of erroring.

Run it

python3 -m http.server 8080    # then open http://localhost:8080/

Needs cross-origin isolation for the threaded WASM engine — handled by coi-serviceworker.js when served over HTTP.

Built with

Vanilla JavaScript (~17,000 lines) · Rust → WebAssembly (jagua-rs, wasm-bindgen-rayon) · Web Workers · SharedArrayBuffer · Supabase for the shared warehouse/offcut stock.

Note on the warehouse login

This repo contains the client for a workshop-internal Supabase project. The anon key is public by design and the tables are protected by row-level security — see WAREHOUSE_AUTH_SETUP.md. The nesting itself runs entirely locally and doesn't need an account.