[ChrisRackauckas-Claude]

Please ignore this draft until reviewed by @ChrisRackauckas.

What changed and why

NonlinearSolution.retcode was declared ::ReturnCode.T, so a return code carried by another scalar type (e.g. a Reactant TracedEnum produced inside a compiled solver loop) cannot be stored in it. This mirrors the ODESolution change from #1564 (released as minor v3.56.0): the field becomes retcode::RC with RC a new trailing type parameter, solution_new_retcode gains a NonlinearSolution method that preserves the carrier type, and successful_retcode's existing convertible-carrier fallback applies. Positional and build_solution constructors are unchanged for ordinary ReturnCode.T values, and no version bump is included per project policy.

Because appending RC turns the ten-parameter spelling NonlinearSolution{T,N,uType,R,P,A,O,uType2,S,Tr} into a non-callable UnionAll, a constructor covering that spelling is added; it takes RC from the retcode argument. Unlike #1563 (which failed released NonlinearSolve precompilation), released downstream code that constructs with all parameters spelled out keeps working — verified below.

This change is non-breaking for ordinary constructors. SciML/NonlinearSolve.jl#1332 keeps vector-form SCC users' legacy ten-parameter FunctionWrapper callbacks converting by matching the declared SubArray element type (including when the legacy wrapper is not the first element of a mixed Any[...] container).

Merge/release only after SciML/NonlinearSolve.jl#1332 is merged and released. Otherwise NonlinearSolve.jl's SCC CI goes red on the next SciMLBase release.

Test evidence

New testset NonlinearSolution retcode type parameter in test/solution_interface.jl. Negative control — same solution_new_retcode call on unmodified origin/master source (src/solutions/nonlinear_solutions.jl checked out from master, everything else identical):

ERROR: LoadError: MethodError: no method matching solution_new_retcode(::NonlinearSolution{Float64, 1, Vector{Float64}, Vector{Float64}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, ::Base.RefValue{SciMLBase.ReturnCode.T})
The function `solution_new_retcode` exists, but no method is defined for this combination of argument types.

With this branch the same script prints retcode fieldtype: Base.RefValue{SciMLBase.ReturnCode.T}, and the full new testset passes inside GROUP=Core below.

Local test runs (Julia 1.12.7, Pkg.test())

Re-run on this head from job scratch (agent/1619-nonlinear-retcode-param):

GROUP=Core:

Remake        | 4605   4605  3m57.7s
     Testing SciMLBase tests passed
SCIMLBASE_CORE_EXIT=0

GROUP=QA (JULIA_CONDAPKG_BACKEND=Null JULIA_PYTHONCALL_EXE=/usr/bin/python3):

Test Summary: | Pass  Total     Time
QA            |  123    123  3m01.5s
     Testing SciMLBase tests passed
SCIMLBASE_QA_EXIT=0

Lint/format on touched files (incl. comment-only follow-up 2fe56b4d): Runic 0, typos 0.

Downstream compatibility

Hit	Verdict
NonlinearSolveBase / SimpleNonlinearSolve ten-parameter constructors	Unaffected — compat constructor; Core suites pass against this branch.
Released SCCNonlinearSolve 1.15.4 Core against this branch	Fails one assertion: isconcretetype(eltype(scc_sol_debug.original)) in core_tests__item4.jl (72 pass / 1 fail / 73). Correct u / Success retcodes; the ten-parameter spelling is a UnionAll, so the buffer/original eltype is abstract. Same concreteness trade-off as #1564. Fixed by SciML/NonlinearSolve.jl#1332.
Code that spells the ten-parameter NonlinearSolution type	Gets an abstract-eltype buffer/original (correct results, loses concreteness) until #1332 is released.
#1332 + this branch	SCC Core green, including legacy homogeneous wrappers and mixed Any[plain, legacy_FW] (14/14 in item8).

Why non-breaking

Every public spelling and call path is preserved: the generic constructor, build_solution, ::NonlinearSolution annotations, and ten-parameter spelled constructor calls (via the compat method). The observable residual is that NonlinearSolution{10 params} denotes a UnionAll rather than a concrete type — the same accepted trade-off as the trailing-RC change to ODESolution in #1564 (minor v3.56.0). Companion #1332 restores concrete eltype for the updated SCC signature and preserves legacy wrapper conversion.

Not verified

• Reactant-side tracing itself.
• GROUP=Downstream/DownstreamAD/Python; docs build (no docstring changes).

Reviewer push-back points

• Confirm abstract eltype for ten-parameter spellers is acceptable under a non-breaking (no bump / minor) release, as for Support compiled ODE return codes and specialization #1564.
• The compat constructor is a catch-all on the 10-parameter spelling; fully-spelled NonlinearSolution{...,RC}(args...) still uses the default inner constructor.

Closes #1619

---

Please ignore this draft until reviewed by @ChrisRackauckas.

Devin CLI 3000.11.3 (original) + Cursor Agent CLI 2026.09.26-dd393fe, model auto (round-2 text/comment + #1332 companion). Transcript: /home/crackauc/sandbox/agent-jobs/NonlinearSolve.jl/jobs/1332-cursor/log.txt on amdci2.julia.csail.mit.edu