ConvexOptimization: depend on released SciMLBase 3.36; unify onto OptimizationSolution

SciMLBase v3.36.0 registers ConvexOptimizationProblem and folds the convex
solution into the single OptimizationSolution (with a type-stable `dual` field),
removing ConvexOptimizationSolution. So:

- Drop the temporary [sources] git-rev to the SciMLBase #1440 branch and bump
  the compat floor to SciMLBase = "3.36" (the release carrying the types).
- Use OptimizationSolution instead of ConvexOptimizationSolution in the backend
  and its test; the solve already returns the one struct with duals populated
  (build_convex_solution defaults calculate_dual = Val(true)).

Verified against released SciMLBase 3.36.0: resolves, LP/SOC solves return
correct primal + dual, non-convex objective rejected; all core tests pass.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: ChrisRackauckas

lib/ConvexOptimization/Project.toml

@@ -13,17 +13,12 @@ SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 SymbolicAnalysis = "4297ee4d-0239-47d8-ba5d-195ecdf594fe"
 Symbolics = "0c5d862f-8b57-4792-8d23-62f2024744c7"
 
-# Temporary: ConvexOptimizationProblem/Solution are unreleased (SciML/SciMLBase.jl#1440).
-# Drop this [sources] once a SciMLBase with those types is registered.
-[sources]
-SciMLBase = {url = "https://github.com/ChrisRackauckas-Claude/SciMLBase.jl", rev = "convex-optimization-problem"}
-
 [compat]
 Clarabel = "0.10, 0.11"
 LinearAlgebra = "1.10"
 MathOptInterface = "1.45"
 Reexport = "1.2"
-SciMLBase = "3"
+SciMLBase = "3.36"
 SparseArrays = "1.10"
 SymbolicAnalysis = "0.3"
 Symbolics = "7"

lib/ConvexOptimization/src/ConvexOptimization.jl

@@ -2,7 +2,7 @@ module ConvexOptimization
 
 using Reexport
 @reexport using SciMLBase
-using SciMLBase: ConvexOptimizationProblem, ConvexOptimizationSolution,
+using SciMLBase: ConvexOptimizationProblem, OptimizationSolution,
     OptimizationFunction, AbstractOptimizationCache, AbstractOptimizationAlgorithm,
     NullParameters, ReturnCode
 import MathOptInterface as MOI
@@ -21,8 +21,8 @@ returns the affine map whose image must lie in the MathOptInterface vector cone
 …). The output length of `g` must equal `MOI.dimension(set)`.
 
 The backend traces `g` on its own symbolic variables, so each `ConeConstraint`
-maps to exactly one MOI constraint and therefore one entry of
-`ConvexOptimizationSolution.dual`, in the order the constraints are given.
+maps to exactly one MOI constraint and therefore one entry of the returned
+`OptimizationSolution.dual`, in the order the constraints are given.
 Because the cone is named explicitly, the returned dual is already expressed in
 the user's variables (no sign remap): `>=` → `MOI.Nonnegatives`, `<=` →
 `MOI.Nonpositives`, `==` → `MOI.Zeros`.

lib/ConvexOptimization/test/core_tests.jl

@@ -1,6 +1,6 @@
 using ConvexOptimization
 using SciMLBase
-using SciMLBase: ConvexOptimizationProblem, ConvexOptimizationSolution
+using SciMLBase: ConvexOptimizationProblem, OptimizationSolution
 import MathOptInterface as MOI
 import Clarabel
 using Test
@@ -20,7 +20,7 @@ c = [1.0, 2.0]
 
     sol = solve(prob, ConvexMOI(Clarabel.Optimizer))
 
-    @test sol isa ConvexOptimizationSolution
+    @test sol isa OptimizationSolution
     @test SciMLBase.successful_retcode(sol.retcode)
     @test isapprox(sol.u, [1.0, 0.0]; atol = 1.0e-6)
     @test isapprox(sol.objective, 1.0; atol = 1.0e-6)