Shared Overlay Reconstruction Plan

Context

After segment classification and rectangle clip landed, Phase 5 still needed the reconstruction plan that turns split edges and labeled faces into union, difference, and symmetric-difference outputs. Without that plan, each later constructive surface would be tempted to invent its own topology assembly path.

Decision

Adopt one shared reconstruction graph:

• classify segments

• emit nodes

• split segments into directed edges

• stable-sort half-edges

• walk rings and chains

• label faces by source coverage

• perform operation-specific face selection

• emit geometry buffers

Only the selection stage changes between union, difference, symmetric difference, and identity.

Consequences

• reconstruction logic is reusable across overlay operations

• stable-order and exact-sign requirements are recorded before CUDA work lands

• dissolve can build on the union branch instead of inventing a new graph

Alternatives Considered

• Separate bespoke reconstruction per overlay operation. Rejected because it duplicates the most expensive topology work.

• Pure host-side Shapely reconstruction for now. Rejected because it leaves no owned assembly seam for GPU work.

Acceptance Notes

This decision lands the reconstruction planner, CCCL-oriented stage mapping, and tests that validate the shared prefix and operation-specific selection stages.