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

[bvdmitri]

Closes #309.

Problem

combine_broadcast_args(args::Vector, kwargs::Vector) (src/model_macro.jl:617-628) was wrong in
both halves, not just the keyword half the issue identifies:

GraphPPL.MixedArguments($(Expr(:tuple, args...)), NamedTuple{$(Tuple(kwargs_keys))}(args))

• the positional half splices the original argument expressions, which refer to the outer,
  un-broadcast collections — not the per-element slots;
• the keyword half builds the NamedTuple from the whole closure tuple, whose arity is
  length(positional) + length(keywords), so the arity mismatches.

Inside the broadcasted expression, args is the varargs tail of the closure at model_macro.jl:734
(do ilhs, args...), populated at :726 as vcat(positional, kwarg_values) — positional slices first,
keyword values after.

Change

Both halves are now sliced out of that runtime tuple by index, emitted as a tuple literal and a
named-tuple literal (mirroring the non-broadcast combine_args) so that proxy_args wraps each element
individually and the two halves stay a Tuple and a NamedTuple, as
MixedArguments{A <: Tuple, K <: NamedTuple} requires:

GraphPPL.MixedArguments((args[1], args[2]), (τ = args[3], θ = args[4]))

Scope: what this does and does not fix

This fixes the lowering. It does not make stochastic mixed-argument broadcasts build, because
MixedArguments is rejected for any node that must be materialized
(graph_engine.jl:2038) — the non-broadcast ~ has exactly the same limitation, so the two paths stay
consistent. Lifting that is a feature, not a bug fix, and is out of scope here.

The user-visible effect is that z .~ sub(mu; y = sg) now reports the real, documented limitation:

MixedArguments not supported for rhs_interfaces when node has to be materialized

instead of MethodError: no method matching tuple(::GraphPPL.VariableRef{...}).

Known remaining gap (pre-existing, not addressed here)

The comma form f(a, b = c) is still not split into positional + keyword at all.
convert_to_kwargs_expression only converts when every argument is a keyword
(is_kwargs_expression(::Vector) requires all), so b = c stays among the positional arguments and the
call still fails with an opaque tuple MethodError. This affects ~, .~ and := alike and is
independent of this fix — x ~ Normal(mixed(1.0, s = 2.0), 1.0) fails identically on main.

I prototyped a split here but backed it out: Julia puts :parameters first in the argument AST, so a
naive trailing-keyword split breaks the existing convert_to_kwargs_expression test at
test/model_macro_tests.jl:553. It deserves its own issue and its own careful patch. Use the explicit
semicolon form f(a; b = c) meanwhile.

Tests

• Updated the two tests that pinned the buggy shape — "combine_broadcast_args"
  (model_macro_tests.jl:1448) and the codegen baseline in "convert_tilde_expression" (Test 13). These
  pinning the wrong output is why the bug survived.
• New graph-level test item in test/graph_construction_tests.jl asserting the mixed broadcast reaches
  the stated MixedArguments limitation rather than a MethodError, and that keyword-only broadcasts
  still build.

Verified the new test fails on main and passes here. Full suite on Julia 1.12:
259/259 test items, all passing.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.73%. Comparing base (4cc7c6c) to head (8c75153).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #327      +/-   ##
==========================================
+ Coverage   90.72%   90.73%   +0.01%     
==========================================
  Files          16       16              
  Lines        2263     2266       +3     
==========================================
+ Hits         2053     2056       +3     
  Misses        210      210              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.