forgekernel 0.1.0

Exact-arithmetic B-rep CAD kernel: topology decided in ℚ, never by floating-point epsilon

forge — the gitcad-native B-rep kernel

From-scratch, unencumbered (Apache-2.0), built to pass the bar OCCT sets and then raise it — the plan and the decision record live in gitcad (ADR-0018, coverage plan).

The three-oracle chain

forgekernel.ref   Python + exact rational arithmetic — the executable
                  specification; topological decisions are EXACT, never
                  epsilon-guarded
      ⇅ differential
OCCT              the 30-year-hardened independent oracle, driven from
                  the gitcad benchmark corpus through the Kernel seam
      ⇅ oracle
forge (Rust)      the production port, added operator class by operator
                  class once ref has proven the semantics

How progress is measured

Not vibes: the benchmark trend in the gitcad repo scores every backend on the shared corpus — capability %, torture-case pass rate, correctness deltas, wall time. Day-one baseline: OCCT scores 93.8%, failing swept_channel (sharp-cornered sweep → invalid geometry). Beating that number with exact arithmetic is the first milestone.

Roadmap (gates, not dates)

• K1 exact planar core: rational linear algebra, plane-based polyhedral B-rep, epsilon-free booleans, divergence-theorem mass properties, native lineage. Gate: planar corpus green vs OCCT.
• K2 quadrics + torus with closed-form intersectors — most of real mech. Gate: ≥80% corpus green.
• K3 NURBS + branch-complete surface–surface intersection (the crown jewel). K4 procedural offsets/shell. K5 G2 blends. K6 the surfacing suite — the SolidWorks-class end goal.

Layout

• src/forgekernel/ — the Python reference kernel (ref)
• rust/ — the forge port (arrives with K1 stability)
• gitcad consumes both through its Kernel seam; nothing here depends on gitcad