Support map
What the kernel can do today, what it partly does, and what it does not do yet.
Every row is one capability. The legend is deliberately blunt:
| Meaning | |
|---|---|
| 🟢 | Implemented. Executable behaviour, tests, and at least one consumer. You can use it now. |
| 🟡 | Partial. Something real exists, but it is narrower than the name suggests. The note says exactly how. |
| 🔴 | Not implemented. Either deliberately out of scope, or plausible future work. The note says which. |
A 🔴 row is not a promise. It is here so you can tell "we decided against this" from "nobody has built it yet" without reading the issue tracker.
For the longer, prose-form status — including the exact/discrete split and what "exact" is claimed to mean — see Capabilities and status. Where the two disagree, that page wins.
How to read the optional column
Axiolid is a kernel, not an application. Almost everything below is behind an additive Cargo feature: you compile the geometry you use and nothing else. The Optional column names the feature, or core when it is always present.
toml
# Points only. Resolves to axiolid + axiolid-core + axiolid-pointcloud.
axiolid = { version = "0.11", default-features = false, features = ["pointcloud"] }Spatial structures
Indices that answer which candidates, never what the intersection is.
| Capability | Support | Optional | Notes |
|---|---|---|---|
| Bounding volume hierarchy (BVH) | 🟢 | spatial | Bvh<K> over anything with an AABB. Deterministic median split; AABB, ray, overlap-pair, and filtered-nearest queries. Used by clash, ray casting, and healing. |
| Uniform point grid (KNN / radius) | 🟢 | pointcloud-queries | PointIndex. Exact distances, not broad-phase bounds, with deterministic tie-breaking by point index. Verified against brute force. |
| Edge adjacency over a mesh | 🟢 | core | EdgeAdjacency: boundary, non-manifold, inconsistent-winding, vertex/triangle neighbours, Euler characteristic. Derived once and shared, so algorithms stop rebuilding it. |
| Octree | 🔴 | — | Not implemented, and not currently justified: the BVH covers object queries and the grid covers point queries. An octree's advantage is sparse volumetric subdivision. It would arrive with a measured workload, not before. |
| k-d tree | 🔴 | — | Not implemented. Would overlap PointIndex for the queries we actually make. |
| R-tree / quadtree / BSP | 🔴 | — | Not implemented, no consumer. |
| Morton / Z-order codes | 🔴 | — | Not computed by us. The boolmesh provider builds its own internally. |
| Half-edge with mutable connectivity | 🔴 | — | EdgeAdjacency is an immutable derived index; every consumer reads adjacency and writes a new mesh. In-place connectivity editing would be a separate type, not a field on this one. |
Geometry representation
| Capability | Support | Optional | Notes |
|---|---|---|---|
| Scalars, frames, transforms, bounds, tolerance | 🟢 | core | axiolid-core. The vocabulary everything else is written in. |
| Triangle and polygon meshes | 🟢 | mesh | TriMesh, PolygonMesh, named attribute channels (per vertex or per triangle corner), zero-copy views. N-gons stay N-gons until you triangulate. |
| Point clouds | 🟢 | pointcloud | PointCloud: positions plus optional normal, colour, intensity. No topology, no adjacency, no source-format types. |
| Layered scalar fields (2.5D grid) | 🟢 | field | LayeredField, row-major cells with a validated configuration. |
| Curves and surfaces | 🟢 | curves, surfaces | Neutral value types. Representation is not an evaluator claim — see the algorithms table. |
| Intrinsic (natural-equation) curves | 🟢 | curves | Curve2::Intrinsic + CurvatureLaw — curvature as a function of arc length. Constant, polynomial (clothoid/Bloss/cubic), sinusoid, composite polynomial-plus-harmonics, and piecewise laws over one domain. Exact and symbolic: differentiate, mirror, and total turning in closed form. Position is deliberately absent — it is a Fresnel-type integral that belongs to an evaluator that can state a tolerance. |
| Profiles and contours | 🟢 | profiles | Validated closed/open contour values. |
| Primitive solids and half-spaces | 🟢 | primitives | Neutral values with validation. |
| Exact B-rep topology | 🟢 | brep | BRep<Curve3, Curve2, Surface> — faces, edges, vertices with exact underlying geometry. |
| Authored geometry graph | 🟢 | model | GeometryGraph, the immutable authored modelling representation. |
| Signed distance fields | 🟢 | field-ops | Sampled Fn(Point3) -> Scalar, the bridge between points, fields, and meshes. |
| Voxel / dense 3D grid | 🟡 | — | Used transiently inside level-set extraction and B-rep compilation. There is no owned, queryable voxel value type. |
| Tetrahedral meshes | 🟡 | — | Tetrahedra are a sampling decomposition inside marching tetrahedra. There is no tet-mesh representation you can hold. |
| Per-point scalar fields as a channel | 🔴 | — | Point clouds carry intensity and nothing more. Arbitrary named scalar channels with gradients/histograms would be new work. |
| Gaussian splats | 🔴 | — | Not implemented. Would be a sibling representation, not an extension of PointCloud. |
Topology and validity
| Capability | Support | Optional | Notes |
|---|---|---|---|
| Mesh audit (manifoldness, winding, degeneracy) | 🟢 | mesh | audit_mesh reports MeshHealth counts rather than a pass/fail verdict, so a caller decides what is acceptable. |
| Mesh healing and diagnosis | 🟢 | heal | Named defects with the offending element, deterministic order. Diagnosis is separate from repair. |
| Connected components | 🟢 | mesh | component_count, decompose, compose. |
| Genus / Euler characteristic | 🟢 | axiolid-inspect † | Refuses anything that is not a closed two-manifold instead of returning a meaningless integer. |
| Boundary and non-manifold edge extraction | 🟢 | core | Via EdgeAdjacency. |
| Containment and winding number | 🟢 | axiolid-inspect † | Robust point-in-solid via generalised winding number. |
| Convex decomposition | 🟢 | axiolid-decompose † | Splits an arbitrary closed solid into convex parts; volume-conserving, every part a closed solid. |
| B-rep validity audit | 🟢 | topology | Face/edge/vertex consistency checks. |
| Non-manifold repair | 🟡 | heal | Defects are detected and named; automatic repair covers a subset. Some defects are reported for the caller to resolve. |
| Topological simplification (feature removal) | 🔴 | — | Not implemented. Plausible future work for model preparation. |
Algorithms
| Capability | Support | Optional | Notes |
|---|---|---|---|
| Exact orientation / in-circle / in-sphere predicates | 🟢 | predicates | Certified signs with degeneracy and filter tests. Depends only on axiolid-core, so you can get exact predicates without meshes. |
| Polygon triangulation | 🟢 | tessellation | Ear clipping through the tessellation contract. Note: not Delaunay. |
| Convex hull (3D) | 🟢 | generate | Incremental hull decided by exact predicates; refuses degenerate input by name. |
| Mesh boolean (union / difference / intersection) | 🟢 | dispatch-mesh-boolean | Through a swappable provider contract, with evidence and typed refusal. |
| Mesh plane section | 🟢 | mesh-section | Exact binary64 plane-side classification with source-topology stitching. |
| Minkowski sum and difference | 🟢 | axiolid-minkowski † | Convex sums exactly via hull of pairwise vertex sums; general operands via convex decomposition. Difference is erosion, and refuses a non-convex subject rather than returning a too-large result. |
| Offset and shelling | 🟢 | generate | Constant-distance offset with miter handling. |
| Mesh decimation | 🟢 | axiolid-decimate † | Edge collapse with a bounded, reported deviation; cumulative per vertex so repeated collapses cannot drift. |
| Mesh refinement and smoothing | 🟢 | axiolid-refine † | Loop-style subdivision and Laplacian smoothing with pinned boundaries. |
| Level-set extraction | 🟢 | axiolid-levelset † | Marching tetrahedra, chosen over marching cubes because it is manifold by construction. |
| Point-set reconstruction | 🟢 | pointcloud-provider | SDF estimated from samples, extracted as a level set. Composed from capabilities the kernel already owns — no third-party numerics adopted. |
| Ray–mesh intersection | 🟢 | ray-mesh | Nearest hit and filtered variants over the BVH. |
| Distance and proximity queries | 🟢 | measure | Point-to-triangle, segment-to-segment, mesh-to-mesh distance with witness points. |
| Volume, area, second moments | 🟢 | measure | Closed-form over closed meshes, plus an exact path for supported B-rep families. |
| NURBS evaluation and differential analysis | 🟢 | nurbs | Curve tangents/curvature, surface fundamental forms, Gaussian/mean/principal curvature. |
| Curve flattening | 🟢 | evaluate | Adaptive, with a chord-error budget; fails closed rather than exceeding it. |
| 2D overlay / boolean on regions | 🟢 | overlay | Sweep-line arrangement with exact predicates. |
| Linear intersection (2D) | 🟢 | linear-intersection | Certified classification with typed refusals naming the operand at fault. |
| Delaunay triangulation | 🟡 | predicates | The exact insphere predicate exists and is tested against near-degenerate cases. No triangulator is built on it — polygon triangulation is ear clipping. |
| 3D curve/surface intersection | 🔴 | — | Not implemented. 2D linear intersection is; the 3D analytic case is not. |
| Normals estimation from raw points | 🔴 | — | We consume normals but cannot compute them from a bare point set. This is the main gap blocking good reconstruction from unoriented captures. |
| Cloud-to-cloud / cloud-to-mesh distance | 🔴 | — | The KNN index makes this cheap to add; it does not exist as an operation yet. |
| Point-set subsampling and outlier filtering | 🔴 | — | Not implemented. Standard first step in any scan pipeline. |
| Registration / ICP alignment | 🔴 | — | Not implemented, and a genuine project rather than a small addition. |
| Voronoi diagrams | 🔴 | — | Not implemented. |
| Mesh parameterisation / UV unwrapping | 🔴 | — | Not implemented. |
| Geodesic distance on surfaces | 🔴 | — | Not implemented. |
† Not re-exported through the axiolid facade yet — depend on the leaf crate directly. This is the ADR 0036 pattern: the facade is a convenience, never a required route, and a leaf consumer should not pay for capabilities it does not use.
Execution and integration
| Capability | Support | Optional | Notes |
|---|---|---|---|
| Provider contracts with typed refusal | 🟢 | contracts | Every operation that can fail says why, by name. An empty result never stands in for "could not". |
| Conformance suites | 🟢 | contracts | Exported as library code, generic over the trait — an out-of-tree provider runs the identical checks. Skips are recorded separately, so a provider cannot reach "conformant" by refusing everything. |
| Runtime provider dispatch with fallback | 🟢 | dispatch-* | Priority ordering, device matching, and memory-budget admission before dispatch. Fallback happens only for Unsupported/Unavailable. |
| Conformance-gated registration | 🟢 | dispatch-* | A non-conformant provider is rejected at registration with its failing report, not discovered later by a caller receiving wrong geometry. |
| Cancellation and scratch budgets | 🟢 | contracts | Providers declare their granularity honestly, including "never polls". |
| Deterministic output | 🟢 | core | Declared per provider, from best-effort through topological to bit-exact. Routing compares strength, so a stronger provider is never rejected for being too good. |
| C ABI | 🟢 | — | Stable v0.4 surface for native consumers. |
| Parallel CPU execution | 🟡 | parallel | The seam exists and some paths use it; most algorithms are still serial. Parallelism is a runtime choice, never a build-for-one-host decision. |
| SIMD | 🟡 | simd | Selected at runtime behind the same seam. Coverage is partial. |
| GPU execution | 🟡 | gpu | Contract and graph-compile seam exist. Not a claim that we bundle GPU kernels. |
| Python bindings | 🔴 | — | Not implemented. |
| WASM target | 🔴 | — | Untested. The pure-Rust core has no obvious blocker, but no claim without evidence. |
Deliberately out of scope
These are not gaps. They are decisions, and they are what keeps the kernel a kernel.
| Not in the kernel | Why | Where it belongs |
|---|---|---|
| LAS / LAZ / E57 / PCD / COPC parsing | Format neutrality. Admitting source-format types is the mistake ADR 0044 exists to prevent. | A sibling ingestion crate outside crates/. |
| STEP / IFC parsing | Same rule, longer standing. | openbimrs/*. |
| Rendering and visualisation | A kernel computes geometry; it does not draw it. | Application layer. |
| Scan registration UI, cleanup workflows | Application policy built on kernel primitives. | Application layer. |
| Statistical testing, Kriging | Analysis of measurements, not geometry. | Application or a stats library. |
| Navigation meshes | Application semantics on top of geometry. | Application layer. |
Honest limits
- "Exact" is scoped. It applies to the certified predicates and the supported exact B-rep families, not to every operation. The capabilities page draws that line precisely.
- Scale is untested above ~10⁶ points.
PointIndexis a uniform grid, appropriate for 10⁵–10⁶ scattered points. We have not measured it at scan scale and do not claim it. - Reconstruction is an estimate. A point set does not determine a unique surface. Results carry evidence — including how many triangles were interpolated across gaps in the capture — so a caller can tell measured surface from invented surface.
- Partial rows are partial on purpose. A 🟡 means the note is load-bearing. Read it before depending on the row.