QA: ExplicitImports tracked @test_broken markers (qualified-accesses-public, no-implicit-imports)

[ChrisRackauckas-Claude]

ExplicitImports QA: tracked @test_broken markers

The QA suite (test/qa/qa.jl + lib/*/test/qa/qa.jl) now runs the full
ExplicitImports.jl check set via SciMLTesting.run_qa(...; explicit_imports = true).
Most checks pass or are handled with curated ignore-lists for non-public names
of stdlib / other SciML packages. A few checks have too many findings to ignore
name-by-name and are tracked as @test_broken (via ei_broken) until fixed:

Root LinearSolve — all_qualified_accesses_are_public

~90 qualified accesses (LinearSolve.x) of names that are not declared public:

• LinearSolve's own internals reached via LinearSolve.x from extension modules
  (e.g. defaultalg, do_factorization, init_sparse_reduction, init_cacheval,
  DefaultLinearSolver, QRFactorization, ...).
• stdlib / SciMLBase / LinearAlgebra internals (@max_methods, Experimental,
  BlasInt, build_linear_solution, checksquare, chkfinite, ...).

Fix path: declare the relevant LinearSolve internals public, and/or upstream
public declarations in SciMLBase / SciMLOperators (tracked separately). Once the
finding count is tractable, replace ei_broken with an ignore-list or drop it.

Sublibraries LinearSolvePyAMG, LinearSolveAutotune — no_implicit_imports

Both sublibraries using whole packages (LinearAlgebra, SparseArrays,
PythonCall, Plots, ...) and rely on names brought in implicitly. Making every
such name an explicit import is a source refactor; tracked as @test_broken until
done.

This is a QA-tracking issue opened alongside the run_qa v1.6 conversion PR; it can
be closed once the markers above are removed.

[ChrisRackauckas-Claude]

Surfaced during the run_qa v1.6 conversion (#1059).

Separately (not an EI marker), lib/LinearSolveAutotune's QA JET.test_package(...; target_defined_modules=true) has a pre-existing hard fail: real undefined-global / method reports in src/telemetry.jl (existing_files, split(::Nothing, "/")) and src/gpu_detection.jl (Base.libdl_name, parse(Int, ::Nothing)), plus GitHub.user not defined. The original qa.jl (Aqua.test_all + the same JET.test_package call) fails identically on the unmodified base, so #1059 does not introduce or mask it — it leaves JET hard (no jet_broken) so the report stays visible. Fixing those source bugs is a separate follow-up.

[ChrisRackauckas-Claude]

Added: LinearSolveAutotune JET jet_broken = true (pre-existing latent telemetry bugs)

During the run_qa v1.6 conversion (PR #1059) the lib/LinearSolveAutotune QA lane was red on JET. The JET check is not newly introduced by the conversion — the prior qa.jl already ran JET.test_package(LinearSolveAutotune; target_defined_modules = true), and #1033 (which first added this QA group) merged with the same lib/LinearSolveAutotune [QA] (julia 1) lane red. So these are pre-existing latent issues that the conversion merely preserves.

JET.report_package(LinearSolveAutotune; target_defined_modules = true) reports 8 issues, all in the telemetry / GPU-detection source (not touched by the conversion):

• src/gpu_detection.jl:660 — parse(::Type{Int64}, ::Nothing) (1/2 union split): mem_match.captures[1] is typed Union{Nothing, SubString} and JET can't narrow it after the mem_match !== nothing guard. Fix: bind the capture to a local and guard on it.
• src/telemetry.jl:518,519,527,532,550 — issue_number / target_repo "is not defined": both are assigned at the top of the try block in upload_to_github but used in the catch block; JET (correctly) treats them as possibly-undefined in catch. Fix: hoist their assignments above the try.
• src/telemetry.jl:621 — split(::Nothing, ::String) (2/2 union split): gist_url = gist.html_url where GitHub.jl types html_url as Union{Nothing, URI}. Fix: guard gist.html_url.
• src/telemetry.jl:667 — existing_files "is not defined": defined inside the try of upload_plots_to_gist (line 576) but used in the catch (the fallback call); genuine scope leak if the try throws before line 576. Fix: hoist existing_files above the try.

Marked jet_broken = true in lib/LinearSolveAutotune/test/qa/qa.jl (auto-flags an Unexpected Pass once the source bugs above are fixed). The real fixes are a small telemetry-source change, out of scope for the QA-conversion PR; tracked here.