GCNConv fails to compile with Reactant - MethodError

A minimal GCNConv forward pass fails inside Reactant.@compile. The same model works with plain CPU forward passes and with Zygote gradients.

Failure example

using GNNLux, GNNGraphs, Lux, Reactant, Random

rng = Xoshiro(42)
model = GCNConv(3 => 2, relu)
ps, st = Lux.setup(rng, model)

g = GNNGraph(([1, 2, 3, 4, 5, 6], [2, 3, 1, 5, 6, 4]); num_nodes=6)
x = randn(rng, Float32, 3, 6)

dev = reactant_device()
ps_d, st_d = (ps, st) |> dev
x_d = x |> dev

compiled = @compile model(g, x_d, ps_d, st_d)  # MethodError

ERROR: MethodError: no method matching scatter(::typeof(+), ::Vector{Reactant.TracedRArray{Float32, 1}}, ::Reactant.TracedRArray{Int64, 2}, ::Vector{Vector{Reactant.TracedRNumber{Float32}}}; update_window_dims::Vector{Int64}, ...)

Closest candidates are:
  scatter(::F, ::Vector{<:Reactant.TracedRArray{T, N}}, ::Reactant.TracedRArray{Int64}, ::Vector{<:Reactant.TracedRArray{T}}; ...) where {F, T, N}
  scatter(::Vector{<:Reactant.TracedRArray{T, N}}, ::Reactant.TracedRArray{TI}, ::Vector{<:Reactant.TracedRArray{T}}; ...) where {T, TI, N}

The error originates in the degree computation (GNNGraphs.degree called from GNNlib.gcn_conv), where the edge weights are materialized as a plain Vector{TracedRNumber{Float32}}, which Ops.scatter method does not accept.

GNNLux & Enzyme (just CPU arrays without Reactant) does work, when Enzyme.set_runtime_activity(Enzyme.Reverse) is used.

Version info

Julia 1.12.6 (aarch64-apple), GNNLux v0.2.0, GNNGraphs v1.5.1, Lux v1.31.4, Reactant v0.2.285, Enzyme v0.13.204, Zygote v0.7.13 (macOS, Apple Silicon, CPU).

Start by reproducing the example at the GCNConv forward pass under Reactant.@compile, then trace GNNGraphs.degree through GNNlib.gcn_conv to the Ops.scatter call. Confirm the traced edge-weight representation and scatter signature involved. Done means the supplied GCNConv compilation succeeds without the MethodError while preserving the existing CPU and Zygote behavior.