Crate axiolid_pointcloud_reconstruction_sdf 

Reference reconstruction: a signed-distance field from samples, extracted as a level set.

Why this method

The kernel already owns both halves of this: PointIndex answers nearest-neighbour queries, and axiolid-levelset extracts a closed manifold surface from any scalar field. Composing them gives a reconstruction with no external dependency and no new geometric machinery to verify.

That matters more than raw quality. Poisson reconstruction produces a smoother surface, but adopting it would mean vendoring a solver whose numerics we cannot audit, to satisfy a contract whose whole purpose is swappability. This provider exists so the contract is verifiable — a better one can replace it without any consumer noticing.

How it works

For a query point p, find the nearest samples and estimate the signed distance to the surface they lie on:

• With normals, project onto the neighbour’s tangent plane. The sign is which side of that plane p falls on, so the surface passes exactly through the samples.
• Without normals, use unsigned distance offset by the sample spacing. This produces a surface around the points rather than through them, which is honest: with no orientation information there is no way to say which side is inside.

The distinction is reported in the evidence, never hidden: a positions-only reconstruction is a genuinely weaker result.

What it is not

Not a hole filler. Where the capture has no data the field is extrapolated from distant samples, and those triangles are counted in interpolated_triangles so a caller can see how much of the surface is inference rather than measurement.

Structs

SdfReconstruction

Reference reconstruction provider.