The generated promoting constructor for `@data` has an unbound type parameter

[SebastianM-C]

Since 0.3.10, @data emits a "promoting" variant constructor for parametric self-referential fields (the fix for #34). When a variant has two or more bare self-referential fields and no other field pins the type parameter, that generated method has an unbound type parameter, so Test.detect_unbound_args fails for any package defining such an ADT.

MWE

import Moshi
using Moshi.Data: @data
using Test

@data Tree{T} begin
    Empty()
    Leaf(T)
    Node(Tree.Type{T}, Tree.Type{T})
end

Test.detect_unbound_args(Tree; recursive = true)

Output on Moshi 0.3.12 / Julia 1.12.7:

1-element Vector{Method}:
 Main.Tree.Node(var"1"::Union{Main.Tree.var"typeof(Tree)"{T}, Main.Tree.var"typeof(Tree)"{Union{}}},
                var"2"::Union{Main.Tree.var"typeof(Tree)"{T}, Main.Tree.var"typeof(Tree)"{Union{}}}) where T
   @ Main.Tree ~/.julia/packages/Moshi/jb5et/src/data/emit/cons.jl:203

If both arguments take the Tree.Type{Union{}} branch, nothing determines T.

As far as I can tell this is a static-analysis problem, not a runtime bug.
Because the variant has ≥2 self-refs, emit_each_variant_cons_inferred emits both the exact and the promoting constructor, and the exact one claims the all-bottom case:

e, l = Tree.Empty(), Tree.Leaf(3)

Tree.Node(l, e)   # -> Tree.Type{Int64}   dispatches to cons.jl:203 (promoting)  ← what #34 needs
Tree.Node(e, e)   # -> Tree.Type{Union{}} dispatches to cons.jl:176 (exact, T = Union{})

So the one argument shape that leaves T unbound in the promoting signature is exactly the shape the more-specific exact method already handles. The method is selectable only when at least one position is non-bottom, and then T is pinned.
detect_unbound_args works on signatures alone and can't see that.

Downstream impact

SymbolicUtils.jl caps Moshi = "0.3.6 - 0.3.9" for this reason (SymbolicUtils#1042):
its BasicSymbolicImpl.Div variant has num::BasicSymbolicImpl.Type{T} and den::BasicSymbolicImpl.Type{T}. Its full test suite passes on 0.3.12; only the QA lane fails. Any package with a binary-node ADT and an Aqua QA lane will hit this.

I have some possible suggestions from Claude, but I'm not sure if they are useful:

1. Bind every self-ref position independently and reconcile in the body, e.g.
   Node(a::Tree.Type{A}, b::Tree.Type{B}) where {A, B}, resolving T from the
   non-bottom parameters and throwing when they're irreconcilable. No unbound
   parameters, and it subsumes the exact constructor. Downside: it accepts
   Node(::Tree.Type{Int}, ::Tree.Type{Float64}) at the signature level, so the
   rejection moves from dispatch to a runtime error.

2. Emit one promoting method per self-ref position, with that position pinned to
   Tree.Type{T} and the rest left as the union. For two self-refs the existing
   exact method resolves the ambiguity at the fully-pinned point. This gets murky for
   three or more self-refs, where partial-bottom combinations are genuinely ambiguous.

[ChrisRackauckas]

#92

Yeah it's not being reviewed, I think we either need to upper bound the package