Remove broad ModelingToolkitBase reexport, export only what docs/tests need

[ChrisRackauckas-Claude]

Summary

@reexport using ModelingToolkitBase leaked the entire MTKB/Symbolics export surface (~300 names, including transitive reexports like solve, Num, structural_simplify, ODESystem, @syms, @mtkbuild, unwrap, ...) through using Catalyst.

This PR replaces it with an explicit export list of just the dependency-owned names that Catalyst's built documentation and test suite actually exercise through using Catalyst:

• DSL macros: @parameters, @variables, @independent_variables, @discretes, @brownians, @poissonians, @named, @nonamespace, @mtkcomplete, @register_symbolic, @unpack
• Module bindings: ModelingToolkitBase, Symbolics, SymbolicUtils
• Types: System, Equation, Differential, Initial, Pre, GlobalScope, ParentScope
• Accessors/builders: complete, compose, continuous_events, equations, extend, full_equations, getmetadata, hasmetadata, independent_variables, isinitial, mtkcompile, observed, operation, parameters, setmetadata, simplify, substitute, unknowns
• Problem constructors Catalyst already exported and extends: ODEProblem, SDEProblem, JumpProblem, NonlinearProblem, SteadyStateProblem (unchanged pre-existing exports)
• Spatial grids: CartesianGrid, CartesianGridRej (unchanged pre-existing exports)

test/qa/qa.jl's LEGACY_DEPENDENCY_REEXPORTS allowlist is shrunk to the same set, and the undefined_exports Aqua exception is removed since the stale Variable reexport it tracked is gone. Reexport is dropped as a direct dependency. HISTORY.md documents the change under "Unreleased"; no version bump — not marked breaking.

Test plan

• GROUP=QA julia --project -e 'using Pkg; Pkg.test()' — passes (14 pass / 6 expected-broken; ExplicitImports reexport check matches the allowlist exactly, Aqua undefined_exports now clean)
• GROUP=Extensions julia --project -e 'using Pkg; Pkg.test()' — passes
• Every file in the Modeling / Simulation / Misc / Spatial / upstream test sets run individually — all pass (one missing name, independent_variables, was caught this way and added)
• julia --project=docs docs/make.jl — full docs build completes; all 168 @example blocks evaluate
• Runic format check clean; typos clean
• GROUP=All via Pkg.test() not run end-to-end (files were run individually instead; the Hybrid lane is separate from All and its file passes)

🤖 Generated with Devin CLI (model: SWE-2 Max), harness version 3000.10.21
Session: local session 7cb1e8c5a1cf4f23 (transcript: /home/crackauc/.local/share/devin/cli/summaries/history_7cb1e8c5a1cf4f23.md)

[isaacsas]

Wouldn’t this be breaking?

[isaacsas]

More generally is there a context where one wants Catalyst but not MTKBase?

[ChrisRackauckas]

Yeah we can make this a breaking release. But I think the bigger thing is, Catalyst should probably be choosing what it considers its public API, rather than deriving it from whatever happens downstream (that makes the definition of breaking very loose), so the question is moreso whether this is the right set.

[ChrisRackauckas-Claude]

Heads-up on overlap: this and #1559 address the same QA failure — limited/limitnew/limitold leaking through the broad @reexport using ModelingToolkitBase.

This PR is the structural fix: it removes the wholesale MTKBase reexport and exports an explicit curated list, so the allowlist entry for limited is no longer needed. Note this drops ~300 transitive reexports (solve, Num, ODESystem, structural_simplify, ...) from using Catalyst — a user-visible breaking surface change, larger in scope than a CI fix typically warrants.

#1559 is the alternative minimal fix: it just approves the three names in LEGACY_DEPENDENCY_REEXPORTS and leaves the reexport surface untouched.

These are mutually exclusive — merge at most one. If this PR merges, #1559 should be closed unmerged.