0023 — Solid generation is an L2 crate, not part of the compiler

• Status: Accepted
• Date: 2026-08-31
• Deciders: Friedrich Schrödter
• Supersedes: 0021
• Package ownership amended by: 0035

axiolid-construct remains an algorithm package. Historical references to a future seam in axiolid-kernel now mean an operation-specific portable contract, introduced only with a real implementation or independent consumer.

Context

ADR 0021 retired the empty axiolid-sweep crate. That part was right: the crate held one trait with no implementors, no tests, and no references.

The rest of it was wrong. It asserted that sweep construction "belongs with the compiler" because it "depends on the compiler's evaluation context, its tolerance policy and its profile machinery". That claim was never measured. It is false.

Measuring the actual import graph of axiolid-mesh-compile:

module	internal deps	external deps
brep	—	core, curve, kernel, mesh, model, surface, topology
compiler	extrude, profile	core, kernel, mesh, model
directrix	—	core, kernel, model
loft	profile	core, kernel, mesh
sweep	loft, profile	core, kernel, mesh
revolve	profile	core, kernel, mesh
extrude	—	core, kernel, mesh, scalar
profile	—	core, curve, kernel, profile, scalar
center_line	profile	core, curve, kernel, profile
half_space	loft, profile	core, kernel, mesh, primitive

The bottom seven modules form a closed cluster. They reference nothing outside themselves, they never touch axiolid-model, and they perform no dispatch — no kernel trait is called anywhere in the cluster. Their only use of axiolid-kernel is the error vocabulary (GeomError/GeomResult, plus BackendId, Operation and Sign, which appear inside those error variants). Only compiler consumes them.

That is not "part of the compiler". That is a self-contained subsystem of roughly 1,400 lines that happened to be born inside an L3 crate.

The distinction matters beyond tidiness. axiolid-mesh-compile is L3 because it walks a DAG, caches by node, and reports as a provider. A caller who has an exact profile and a path and wants a solid should not have to adopt a graph, a cache, and a model vocabulary to get one — which, before this change, was the only way to reach extrude.

Decision

Extract the seven modules into a new crate, axiolid-construct, at L2.

axiolid-mesh-compile keeps exactly what its name implies: compiler (DAG walk, caching, provider identity), brep (B-rep face tessellation) and directrix (model-driven path extraction). It depends on axiolid-construct like any other consumer.

L2 is the correct tier by the workspace's own rule — algorithms over L1 representations. axiolid-construct sits beside axiolid-reference, axiolid-tessellation-contract and axiolid-measure: it consumes representation types and produces meshes, and it solves rather than merely describes. It is not L1 (it is not a value vocabulary) and not L3 (it owns no execution context, selects no backend, and implements no provider trait).

The crate reports its own BACKEND_ID of scalar-generate. Previously the moved code identified its failures as scalar-compile, which is now simply untrue: a refusal raised while building a swept solid did not come from the compiler.

The facade exposes it as an opt-in generate feature and an axiolid::generate module. axiolid-mesh-compile remains unexposed — the graph compiler is an implementation detail; generation is a capability.

On the retired Sweeper trait

ADR 0021's conclusion stands, for a better reason than it gave. A trait is warranted when there is something to swap. axiolid-construct is a concrete scalar implementation with no competing implementor, so it exports plain functions. When a second implementation appears — a GPU tessellator, or an exact B-rep sweep under ADR 0020 — the seam belongs in axiolid-kernel beside MeshBoolean, which is the one part of 0021 that was correct.

Under ADR 0020 the eventual exact sweep returns a B-rep rather than a mesh. That reinforces this placement rather than undermining it: an exact sweep is still an algorithm over representations, so it belongs at L2 too, and having generation already separated from graph evaluation is what makes adding it a local change.

Consequences

• axiolid-mesh-compile drops from 4,595 to roughly 3,200 lines and loses its direct earcut dependency in the generation path.
• Solid generation is usable without the DAG, the node cache, or axiolid-model. The facade's generate feature is verified by a test that builds a solid through axiolid::generate alone.
• Two functions that were pub(crate) — sweep::linear_extrusion_normals and half_space::for_subject — become public API. The extraction converted an implicit intra-crate coupling into an explicit, reviewable interface, which is the point.
• Seven test files moved with their modules. Two tests that cross the generation/boolean seam stayed at L3 in axiolid-mesh-compile, because axiolid-mesh-boolean-boolmesh is L3 and an L2 crate must not depend on it. The layering test caught this: the first attempt moved them wholesale and failed with crates/algorithms/construction/construct (tier 2) depends on axiolid-mesh-boolean-boolmesh (tier 3).
• Test count is unchanged at 570, confirming the split moved work rather than dropping it.

Alternatives considered

Leave it in axiolid-mesh-compile. Rejected: it makes generation unreachable without the graph, and it was the unexamined position ADR 0021 defended.

Move only sweep and loft. Rejected: sweep depends on profile, and half_space depends on both. Splitting the cluster would leave a crate-crossing dependency where there is currently none.

Put it at L3 beside the compiler. Rejected: L3 is for execution contexts and provider implementations. axiolid-construct selects no backend and implements no operation trait.

Reinstate a Sweeper trait in axiolid-kernel now. Rejected: one implementor. This is the mistake axiolid-sweep already made once, and ADR 0017 warns against contracts written before a second implementation exists.