Pin ModelingToolkit's public reexport surface

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Stacked on #4880 — review that one first; this branch contains its commit.

Problem

On unmodified master (32a955a), run_qa's No unapproved public reexports check fails with 487 findings. ModelingToolkit exposes all of these publicly and owns none of them:

owner	count	how it gets there
ModelingToolkitBase	339	@reexport using ModelingToolkitBase + @import_mtkbase
Symbolics	83	@reexport using Symbolics inside ModelingToolkitBase
SymbolicUtils (incl. Rewriters, Code)	32	transitively via Symbolics
SciMLBase	19	problem/function types re-exported by ModelingToolkitBase
TermInterface	4	transitively via SymbolicUtils
UnPack	3	@reexport using UnPack
CommonSolve / JumpProcesses / BipartiteGraphs / ModelingToolkitTearing / StateSelection	1 each	assorted

A blanket reexport cannot distinguish "we mean to expose this" from "a dependency happened to add a public binding", which is exactly what the audit is complaining about.

Change

1. REEXPORTED_API in test/qa/qa.jl records the approved set, grouped by owner, each group carrying the reason it is intentional. This is the value the check exists to protect: the set is now version-controlled, so a dependency adding a new public name fails this test rather than silently widening ModelingToolkit's API, and every addition to the surface becomes a deliberate decision. Removing an entry is a breaking change.

2. @reexport using Symbolics and @reexport using UnPack are dropped from src/ModelingToolkit.jl. ModelingToolkitBase already re-exports both and ModelingToolkit @reexports ModelingToolkitBase, so these two lines only duplicated a surface that already existed.

One entry is called out as an accidental leak rather than a deliberate export: find_solvables! is a StateSelection internal that StructuralTransformations exports but only ever calls qualified (StateSelection.find_solvables!). It is listed with a comment; dropping it is a breaking change and belongs in a major release, not a QA fix.

What this PR does not do

It does not shrink the surface. Trimming the names nobody uses (the sympy_* family, the Laplace/Groebner helpers, @makearray, print_ir, …) removes public API and therefore requires a major version bump, and I could not validate the removals without running the full ModelingToolkit + Catalyst + ModelingToolkitStandardLibrary suites. The pinned list makes that a tractable follow-up: the candidates are exactly the entries in the Symbolics/SymbolicUtils groups with no use site in docs/, test/, Catalyst.jl or ModelingToolkitStandardLibrary.jl.

Verification

Run against this branch on Julia 1.11, replicating run_qa's logic (public_reexports(pkg; allow = reexports_allow)), plus a stale-entry check that run_qa does not do:

REEXPORTED_API entries = 487
unique = 487
unapproved public reexports = 0
allow-list entries no longer on the surface = 0
Test Summary:                  | Pass  Total  Time
No unapproved public reexports |    2      2  0.0s

The reexport removal in src/ModelingToolkit.jl leaves the surface byte-identical — public_reexports(ModelingToolkit) was dumped before and after and diffed:

=== DIFF before/after ===
IDENTICAL SURFACE

Smoke test on this branch (Lorenz mtkcompile + solve(Rodas5P()), index-3 pendulum mtkcompile, linearization_function/linearize):

Test Summary: | Pass  Total   Time
smoke         |    3      3  22.3s

🤖 Generated with Claude Code

https://claude.ai/code/session_019ywCW8vbGoc9X3dUbmXyme