Carry dimension types through the axis interface: dimtype and a 6-arg yaxcreate

[bjarthur]

This is the alternative to JuliaDataCubes/YAXArrays.jl#626 described in this comment there, implemented so that the two approaches to JuliaDataCubes/YAXArrays.jl#362 can be compared as code. One of the two should be taken, not both.

Problem

yaxconvert reduces a source array to dimension names and values and hands them to yaxcreate(T, data, dnames, dvals, attrs). A target that models dimensions as types, DimensionalData, has nothing to go on but the name, so it rebuilds every axis as Dim{name}(values): a DimArray with a user's @dim D1 converts to one with Dim{:D1}, and X, Ti become Dim{:X}, Dim{:Ti}.

Change

• A new optional interface function dimtype(x, i): the dimension type of axis i, or nothing when the source only has a name (the fallback, so nothing changes for existing implementers). The DimensionalData extension defines it as DimensionalData.basetypeof(dims(x)[i]), which gives D1 for a D1, X for an X, Dim{:lon} for a Dim{:lon}.
• yaxconvert now calls a six-argument yaxcreate(T, data, dnames, dtypes, dvals, attrs). Its default forwards to the five-argument method, so targets that do not model dimension types (AxisArrays, NamedDims, KeyedArray, NamedTuple) are untouched. The DimensionalData extension implements it as dtypes[i] === nothing ? Dim{dnames[i]}(dvals[i]) : dtypes[i](dvals[i]).
• dimtype is exported alongside dimname and dimvals.

The YAXArrays side (defining dimtype(::YAXArray, i) and the six-argument yaxcreate(::Type{YAXArray}, ...)) is in JuliaDataCubes/YAXArrays.jl#TBD, which needs this released first.

Compared with the direct methods in YAXArrays#626

That PR adds yaxconvert(::Type{YAXArray}, ::AbstractDimArray) and yaxconvert(::Type{<:DimArray}, ::YAXArray) passing the Dimension objects through, in YAXArrays alone. This one extends the protocol so any source that can supply dimension types has them preserved through the shared pipeline. Trade-offs: the direct methods also keep lookups (Sampled vs Categorical, order, span) and metadata exactly, which this still rebuilds from values; this one touches the public interface and needs a release before YAXArrays can use it, but has no special case for a pair of types. Only DimensionalData-based sources carry dimension types today, so the two are equivalent in what they fix.

Metadata normalisation (NoMetadata to a Dict) is a separate defect handled by #70 either way.

Tests

The DimensionalData test item checks dimtype on a plain source (nothing) and on a DimArray (D1, X), that a DimArray round trip through yaxconvert keeps typeof(dims(...)), and that the plain-array conversion still produces Dim{:x} axes.

🤖 Generated with Claude Code