0014 — Net geometry is an explicit request; openings are exact booleans

• Status: Accepted
• Date: 2026-09-24
• Deciders: openbimrs contributors
• Supersedes: —
• Amends: ADR 0012 — adds a second compile entry point with the same backend seam

Context

A product's Body representation is its gross shape. Openings are separate products (IfcOpeningElement, IfcVoidingFeature, any IfcFeatureElementSubtraction) related to their host by IfcRelVoidsElement. The file never authors the net result.

Until #44, compile_product_mesh compiled only the Body, so every wall came back solid through its doors. On the local real-model corpus the relation appears 49 to 775 times per model. No compiled mesh reflected any of them.

Both answers are legitimate. Quantity takeoff is usually gross, with openings reported separately. Clearances, net areas, window-to-floor ratios and clash tests through an opening need net. Neither can replace the other.

Decision

We will keep compile_product_mesh gross and unchanged, and add a separate net entry point:

• lower::lower_product_net lowers the host Body, then each voiding opening's Body through the opening's own placement chain, and appends exact SolidOperation::Boolean { Difference } nodes. Nothing is evaluated in the bridge, per ADR 0004.
• compile::compile_product_mesh_net and compile_product_mesh_net_with compile that graph. They return NetMesh { mesh, openings }, where openings lists every subtracted opening in ascending id. The _with form takes any MeshCompiler, as ADR 0012 requires.
• openings_of(model, host) is a kernel-free read of the relation.
• Refusal, not fallback. If any opening cannot be removed, the result is GeometryError::OpeningNotSubtracted { host, opening, cause }. Returning the gross body would hand back geometry that claims to be net and is not.
• Multi-item bodies are split into solid parts. The neutral graph accepts only solids as boolean operands, while a multi-item Body lowers to a Collection (or an Instance of one, for a mapped item). Both host and opening are flattened to their solid parts, with instance transforms carried onto each part. Each host part is cut on its own, since (A ∪ B) − O = (A − O) ∪ (B − O).
• Order is content-defined. Openings are subtracted in ascending opening id, and an opening named by two relations is subtracted once.

Alternatives considered

Option	Why not
Make compile_product_mesh net by default	Silently changes every existing caller's volumes, and gross is the right answer for takeoff.
A bool / options flag on the existing function	Hides a different return contract (the subtracted-opening list, a new error) behind a flag. Two named functions are clearer to callers, most of them agents.
Skip an opening that fails, return the rest	The mesh would claim to be net while still solid through that opening. This is the failure #44 was filed to remove.
Evaluate the subtraction in the bridge	Violates ADR 0004; the backend decides how booleans are computed.
Union all opening parts into one tool first	Adds a union the file never authored and a second failure mode for no benefit. Sequential differences give the same region.

Consequences

Positive

• Net geometry is reachable with the same backend seam as gross.
• Every refusal names the opening at fault. On a backend refusal, the host is compiled alone first, so a host that cannot mesh even gross is blamed as the host rather than on its first opening.
• The success path compiles the net graph once.

Negative / costs

• A refusal pays a diagnostic pass: the host alone, then each chain prefix until one fails. That cost only lands on calls that already failed.
• voidings_of scans all relations per call. Netting every product of a model is O(products × relations). That is fine at the measured hundreds of relations; an index is the fix if it ever shows up in a profile.
• Net quality is bounded by the backend's boolean on real exporter data. Non-manifold operands are refused by the reference kernel, and those refusals surface as OpeningNotSubtracted. On the local corpus that is 14 of 423 hosts with axiolid-construct 0.3.1.

Follow-ups / risks to watch

• IfcRelProjectsElement (additions) is the mirror case and is not handled.
• axiolid-construct 0.3.0 wound a downward extrusion inside-out, which is how many exporters author opening bodies, so those subtractions were refused (never inverted). Fixed in 0.3.1 (axiolid/kernel#166), now the workspace floor.

Relation to existing code

• ifc-geometry/src/input/openings.rs — kernel-free relation read; slot constants asserted in tests/context_slots.rs (ADR 0008 pattern).
• ifc-geometry/src/lower/net.rs — graph construction and body splitting.
• ifc-geometry/src/lower/session.rs — NodeShape side table recorded at push.
• ifc-geometry/src/compile.rs — net entry points and refusal attribution.
• Tests: tests/opening_subtraction_compile.rs (analytic volumes), tests/opening_attribution.rs (backend refusals named precisely).