0018 — Kernel-free geometry is a feature, not a crate

• Status: Accepted
• Date: 2026-09-27
• Deciders: openbimrs contributors
• Supersedes: —

Context

ifc-geometry serves two kinds of consumer. A drawing or quantity tool needs representation contexts, plan and body selection, placements, units and the typed resource views: slot reading over ifc-model. A viewer or kernel needs those same views lowered into the format-neutral Axiolid DAG. Only the second needs the axiolid-* representation crates, and linking them costs the first consumer a geometry kernel it never calls.

The split between the two halves has to live somewhere. The obvious candidate is a second crate (an ifc-geometry-lower beside a kernel-free ifc-geometry). The alternative is one crate with an optional, default-on lowering feature that gates the neutral dependencies.

This decision was made when the feature was introduced and recorded only in the crate's context file. It is written down here so it survives that file.

Decision

We will keep lowering inside ifc-geometry, behind the default-on lowering feature, rather than in a separate crate.

• lower is a leaf module: nothing else in the crate depends on it, so the feature boundary holds structurally. Outside lower and the compile layer built on it, only feature-gated conversions such as Transform::to_geom and the IfcBooleanOperator conversion name an axiolid-* type; everything else reads ifc-model slots and compiles with the feature off.
• tests/kernel_free_build.rs checks the resolved dependency graph under --no-default-features, so a stray unconditional use axiolid_* fails the gate instead of silently relinking the kernel for 2D consumers.
• Placement resolution (constraint::placement) stays outside lower, because world coordinates are needed with lowering off.
• The compile features (ADR 0004, ADR 0012) sit on top of lowering; they do not change this split.

Alternatives considered

Option	Why not
A separate lowering crate	Costs a published crate name, its own release and changelog, and a manifest edit for every consumer that wants lowering, to enforce a boundary one test already enforces.
One crate with unconditional kernel dependencies	Every drawing and quantity consumer would link the geometry kernel.
Kernel-free crate by convention only	An optional dependency can be re-enabled through a feature edge; only the resolver sees it, so the boundary needs the dependency-graph test either way.

Consequences

Positive

• One crate name, one version, one changelog; ifc-geometry with default features keeps working unchanged for existing consumers.
• The kernel-free build is proven by the resolver, not by review.

Negative / costs

• --all-features cannot see a break that exists only with the feature off, so the gate builds, tests, lints and documents --no-default-features separately.
• Intra-doc links to feature-gated items need that separate rustdoc run.

Follow-ups / risks to watch

• Revisit if the kernel-free half gains heavy dependencies of its own, or needs a release cadence independent of lowering. Either would make a separate crate pay for itself.

Relation to existing code

• crates/ifc-geometry/Cargo.toml: the lowering, compile and compile-reference-backend features.
• crates/ifc-geometry/src/lib.rs: #[cfg(feature = "lowering")] pub mod lower.
• crates/ifc-geometry/tests/kernel_free_build.rs: the dependency-graph proof.
• scripts/gate.sh, section features: the kernel-free column.