0008 — Fixed slot constants for stable relationships

• Status: Accepted
• Date: 2026-08-26
• Deciders: openbimrs contributors
• Supersedes: —

Context

IFC stores no parent pointers. A wall does not name its storey; a separate IfcRelContainedInSpatialStructure entity names both ends. So every containment question is a question about relationship entities and the attribute slots their ends sit in.

Those two slots are not in a consistent order:

IfcRelAggregates                    4 = RelatingObject   5 = RelatedObjects
IfcRelContainedInSpatialStructure   4 = RelatedElements  5 = RelatingStructure

Reading containment as though it matched aggregation inverts the tree: elements become the parents of their storey. Nothing about that failure is loud — the types match, the walk terminates, and every downstream answer is wrong.

The obvious safeguard is to resolve slots through ifc-schema at runtime, the way ADR 0007 does for authoring. That would make containment traversal — the most common query any consumer makes — depend on parsing a 3 MB EXPRESS file, to answer a question whose answer is fixed for the lifetime of the schema.

Decision

ifc-spatial states the slot positions as constants and does not depend on ifc-schema.

The constants are asserted against the shipped IFC2x3, IFC4 and IFC4x3 schemas in tests/slot_layout.rs, including an explicit test that the two relationships still disagree. A slot constant may not be added without extending that test.

Consequences

• Containment traversal works on a model alone; no schema file, no schema feature, no parse cost.
• The correctness argument moves from the type system into a test. That is only acceptable because the test runs against the normative artifacts rather than against a restatement of the same assumption.
• A future schema that reorders these attributes fails the layout test rather than silently inverting a customer's floor plan.
• Relationships whose layout is not stable across versions must not be added as constants. They belong behind a schema lookup, and the boundary note in that crate's AGENTS.md says so.

Alternatives considered

Option	Why not
Resolve every slot via ifc-schema	Forces a schema parse on the most common query in the library to answer a question that cannot change within a schema version
Hard-code without the layout test	The failure mode is silent inversion; an untested constant is the whole risk
Put traversal in ifc-model	ifc-model must not name concrete entity types — enforced by model_invariants.rs