fix: correct the lowering of mixed positional/keyword broadcast arguments

src/model_macro.jl

@@ -614,16 +614,25 @@ combine_broadcast_args(args::Vector, kwargs::Nothing) = quote
     args
 end
 
+# Inside the broadcasted expression `args` is the varargs tail of the broadcast closure, holding the
+# per-element slice of every combinable argument: the positional ones first, then the keyword values, in
+# declaration order. Both halves therefore have to be sliced out of that runtime tuple. Splicing the
+# original positional expressions here instead would capture the outer, un-broadcast collections, and
+# building the keyword `NamedTuple` out of the whole tuple mismatches its arity.
 function combine_broadcast_args(args::Vector, kwargs::Vector)
     kwargs_keys = [arg.args[1] for arg in kwargs]
     if length(args) == 0
         return quote
             NamedTuple{$(Tuple(kwargs_keys))}(args)
         end
     else
-        return quote
-            GraphPPL.MixedArguments($(Expr(:tuple, args...)), NamedTuple{$(Tuple(kwargs_keys))}(args))
-        end
+        npositional = length(args)
+        # A tuple literal and a named-tuple literal, mirroring the non-broadcast `combine_args`, so that
+        # `proxy_args` wraps each element in its own `proxylabel` and the two halves stay a `Tuple` and a
+        # `NamedTuple` as `MixedArguments{A <: Tuple, K <: NamedTuple}` requires
+        positional_args = Expr(:tuple, [:(args[$i]) for i in 1:npositional]...)
+        keyword_args = Expr(:tuple, [Expr(:(=), key, :(args[$(npositional + j)])) for (j, key) in enumerate(kwargs_keys)]...)
+        return :(GraphPPL.MixedArguments($positional_args, $keyword_args))
     end
 end
 

test/graph_construction_tests.jl

@@ -2019,6 +2019,41 @@ end
     end
 end
 
+@testitem "Broadcasting with mixed positional and keyword arguments" begin
+    using Distributions
+    import GraphPPL: create_model
+
+    include("testutils.jl")
+
+    @model function bc_mixed_sub(out, x, y)
+        out ~ Normal(x, y)
+    end
+
+    @model function bc_mixed_main()
+        local mu
+        local sg
+        for i in 1:5
+            mu[i] ~ Normal(0, 1)
+            sg[i] ~ Gamma(1, 1)
+        end
+        z .~ bc_mixed_sub(mu; y = sg)
+        out ~ Normal(z[5], 1)
+    end
+
+    # A broadcast mixing positional and keyword arguments now lowers to a well-formed `MixedArguments`,
+    # built by slicing the broadcast closure's `args` tuple. Materializing a node from `MixedArguments`
+    # is still unsupported (same as for the non-broadcast `~`), but the user gets that stated limitation
+    # instead of an opaque `MethodError: no method matching tuple(...)` from the broken lowering.
+    @test_throws "MixedArguments not supported" create_model(bc_mixed_main())
+
+    # Keyword-only and positional-only broadcasts are unaffected
+    @model function bc_kwargs_only()
+        y .~ Normal(fill(0.0, 5), 1.0)
+        z .~ Normal(mean = y, var = fill(1.0, 5))
+    end
+    @test create_model(bc_kwargs_only()) isa GraphPPL.Model
+end
+
 @testitem "Multiple anonymous variables in one context should not collapse to a single key" begin
     using Distributions
     import GraphPPL:

test/model_macro_tests.jl

@@ -1456,8 +1456,14 @@ end
     @test_expression_generating combine_broadcast_args([], [Expr(:kw, :μ, :μ), Expr(:kw, :σ, :σ)]) quote
         NamedTuple{$(:μ, :σ)}(args)
     end
-    @test_expression_generating combine_broadcast_args([:μ, :σ], [Expr(:kw, :μ, :μ), Expr(:kw, :σ, :σ)]) quote
-        GraphPPL.MixedArguments((μ, σ), NamedTuple{$(:μ, :σ)}(args))
+    # Both halves are sliced out of the broadcast closure's `args` tuple: the positional arguments come
+    # first, the keyword values after them. Splicing the original expressions (`(μ, σ)`) would capture the
+    # outer, un-broadcast collections instead of the per-element slots.
+    @test_expression_generating combine_broadcast_args([:μ, :σ], [Expr(:kw, :τ, :τ), Expr(:kw, :θ, :θ)]) quote
+        GraphPPL.MixedArguments((args[1], args[2]), (τ = args[3], θ = args[4]))
+    end
+    @test_expression_generating combine_broadcast_args([:μ], [Expr(:kw, :σ, :σ)]) quote
+        GraphPPL.MixedArguments((args[1],), (σ = args[2],))
     end
 end
 
@@ -1792,8 +1798,14 @@ end
                 some_node,
                 ilhs,
                 GraphPPL.MixedArguments(
-                    (GraphPPL.proxylabel(:a, a, nothing, GraphPPL.False()), GraphPPL.proxylabel(:b, b, nothing, GraphPPL.False())),
-                    GraphPPL.proxylabel(:anonymous, NamedTuple{$(:μ, :σ)}(args), nothing, GraphPPL.False())
+                    (
+                        GraphPPL.proxylabel(:args, args, (1,), GraphPPL.False()),
+                        GraphPPL.proxylabel(:args, args, (2,), GraphPPL.False())
+                    ),
+                    (
+                        μ = GraphPPL.proxylabel(:args, args, (3,), GraphPPL.False()),
+                        σ = GraphPPL.proxylabel(:args, args, (4,), GraphPPL.False())
+                    )
                 )
             )
         end