QA: run_qa v1.6 form + ExplicitImports

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Converts the hand-rolled test/qa/qa.jl (raw Aqua.test_* + per-function JET.report_call) to the SciMLTesting 1.6 run_qa form and enables the ExplicitImports checks.

ExplicitImports findings (run locally vs released SciMLTesting 1.6.0)

Check	Outcome
no_stale_explicit_imports	fixed — removed stale ArrayInterface.allowed_getindex (imported, never used)
no_implicit_imports	broken (#231) — ~31 implicit names from using LinearAlgebra, SparseArrays, Printf, PrecompileTools; explicit refactor deferred
all_explicit_imports_via_owners	pass
all_qualified_accesses_via_owners	ignore — BlasFloat/chkstride1/libblastrampoline reached via LinearAlgebra.BLAS (owners: LinearAlgebra / libblastrampoline_jll)
all_qualified_accesses_are_public	ignore — Base / LinearAlgebra(.BLAS/.LAPACK) non-public names (@aliasscope, @assume_effects, @blasfunc, @propagate_inbounds, chkfinite, chklapackerror, gebal!, gesv!, rcswap!, stegr!, ...)
all_explicit_imports_are_public	ignore — non-public names imported from LinearAlgebra / ArrayInterface / Base (BlasInt, checksquare, ismutable, allowed_setindex!, typename) + Stegr submodule's @blasfunc/stegr!

All ignored names are other packages' internals that will become public as the base libs declare them; only own-package surface is otherwise touched.

Analyzability fix

ExplicitImports could not analyze the module because of the dynamic for i in 1:13; include("exp_generated/exp_$i.jl"); end in src/exp_noalloc.jl (UnanalyzableModuleException). Replaced it with 13 literal includes. Verified the Higham2005 matrix exponential still matches Base exp to ~6.7e-16 and expv is unchanged. This unblocked no_implicit_imports and no_stale_explicit_imports.

Deps

• test/qa/Project.toml: SciMLTesting compat → "1.6". Aqua + ExplicitImports are transitive via SciMLTesting; Aqua is kept a direct dep so the test_ambiguities child process can resolve it. JET kept for the JET check.
• Root Project.toml: SciMLTesting compat → "1.6".

Verification

QA group on Julia 1.10 (lts), released SciMLTesting 1.6.0 (Pkg.test, GROUP=QA, folder model):

Test Summary: | Pass  Broken  Total
QA/qa.jl      |   17       1     18
Testing ExponentialUtilities tests passed

(the one Broken is no_implicit_imports, tracked in #231).

Pre-existing JET-on-1.12 caveat

On Julia 1.12 the JET typo check reports pre-existing "local variable may be undefined" findings (kiops order/kest, Higham2005 ilo/ihi/scale/bal). master is already red on the QA(1)/1.12 lane today, and the source-level fixes live in the still-open draft #229. This PR does not bundle that numerical-source fix (out of scope, belongs in #229) and does not use jet_broken (it would turn the clean lts JET lane into an Unexpected-Pass error). The lts QA lane is fully green; the 1.12 JET red is pre-existing and tracked by #229.

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

QA(julia 1) red diagnosis + fix

Failing lane: tests / QA (julia 1, ubuntu-latest) — JET typo-mode reported 20 "local variable may be undefined" errors (JET.report_package(ExponentialUtilities; target_modules=(ExponentialUtilities,), mode=:typo)). Findings in src/kiops.jl (order, kest), src/exp_baseexp.jl (si), and src/exp_noalloc.jl (ilo/ihi/scale/bal).

Classification: conversion bug (run_qa's broader JET check surfaces genuine latent type-instabilities the old hand-rolled qa.jl never reached).

Two root causes:

1. Stale base. This branch was cut from v1.30.1 (pre-Fix master CI: expv zero-input NaN, JET-on-1.12 QA, GPU-in-All #229). Master Fix master CI: expv zero-input NaN, JET-on-1.12 QA, GPU-in-All #229 already fixed the kiops order/kest and exp_baseexp si undefined locals, and Release v1.30.2 #233 released v1.30.2. Rebased onto current master to pick those up (the only conflict was qa.jl, resolved in favor of the run_qa form).

2. exp_noalloc.jl ilo/ihi/scale/bal (not fixed by Fix master CI: expv zero-input NaN, JET-on-1.12 QA, GPU-in-All #229). The old qa.jl ran JET.report_call(exponential!, (Matrix{Float64},)). ExpMethodHigham2005(A::Matrix) sets do_balancing = (A isa StridedMatrix) as a constant, so JET constant-propagated do_balancing == true, both if method.do_balancing blocks folded to a single branch, and the balancing locals were always-defined — clean. run_qa's report_package analyzes exponential!(A, ::ExpMethodHigham2005, _cache) in isolation with an abstract ExpMethodHigham2005, so do_balancing is a runtime Bool, the two balancing blocks aren't provably correlated, and the undo block reads possibly-undefined locals. This was latent on master (its report_call never reached it) and is a real type-instability. Fixed by seeding ilo=1/ihi=n/scale=_scale and lifting the GenericSchur permutations into prow/pcol locals — behavior unchanged (seeds only live when do_balancing==false, where the undo block does not run).

Verified locally (released SciMLTesting 1.7.0):

• Julia 1.12.6: report_package typo mode 20 → 0 reports ("No errors detected").
• Julia 1.10.11 numerics unchanged: strided-BLAS balancing relerr 3.3e-16, GenericSchur/BigFloat balancing 1.1e-16, no-balancing 1.6e-16 vs reference exp.
• Runic check passes on exp_noalloc.jl.

Note — Julia 1.10 EI (not a CI lane): with explicit_imports=true, the no_stale_explicit_imports check throws UnanalyzableModuleException on Julia 1.10 because ExplicitImports' static parser won't recurse into the generated-code includes in exp_noalloc.jl ("Dynamic include found … not recursing"). This reproduces byte-identically on master's for i in 1:13 include(...) loop, so it's a pre-existing ExplicitImports-on-1.10 limitation, not introduced here, and it does not affect CI: the centralized grouped-tests workflow schedules QA only on julia-1 (≥1.11), where the module is analyzable and the full QA group passes (Aqua + EI 5 pass/1 broken + JET).

Ignore until reviewed by @ChrisRackauckas.