LearnAPI.clone calls the advertised keyword constructor positionally

[s-celles]

Hello,

While working on a downstream package (OnlineML.jl), I noticed a discrepancy between the constructor trait documentation and the actual implementation of LearnAPI.clone.

The constructor trait documentation requires a keyword constructor and demonstrates reconstruction with:

LearnAPI.constructor(learner)(; named_properties...)

However, LearnAPI.clone currently collects the learner properties and passes them as positional arguments by splatting a NamedTuple without a semicolon:

LearnAPI.constructor(learner)(NamedTuple{names}(new_values)...)

Minimal reproducer

using LearnAPI

Base.@kwdef struct DemoLearner
    rate::Float64 = 0.1
end

LearnAPI.constructor(::DemoLearner) = DemoLearner

LearnAPI.clone(DemoLearner(); rate=0.2)

Expected Behavior

DemoLearner(0.2) should be returned through the documented keyword-constructor contract (e.g., if clone splatted with a semicolon: ; NamedTuple{names}(new_values)...).

Actual Behavior

clone calls DemoLearner(0.2) positionally. A learner that intentionally provides only the documented keyword constructor (like the one generated by Base.@kwdef without a custom positional fallback) raises a MethodError.

Additional Context

Currently, downstream learners with properties must provide an additional positional constructor solely for compatibility with LearnAPI.clone.

Fixing this by changing clone to splat the named tuple as keywords (;) would align the code with the documentation. However, please note that changing clone upstream may affect existing learners that have already implemented (or only implemented) a positional constructor to work around this, so an upstream compatibility transition/deprecation phase might be necessary.

Affected code

LearnAPI.jl/src/clone.jl

Line 26 in b17f7de

return LearnAPI.constructor(learner)(NamedTuple{names}(new_values)...)

Possible change

return LearnAPI.constructor(learner)(; NamedTuple{names}(new_values)...) 

PS : this issue have been filled with AI assistance