ci: align sublibrary CI to OrdinaryDiffEq canonical translating dispatcher

[ChrisRackauckas-Claude]

Summary

This PR aligns the NonlinearSolve sublibrary CI to the exact OrdinaryDiffEq.jl canonical pattern: a translating root dispatcher plus clean-env sublibrary runtests, using the standard Core / QA / GPU test groups.

It keeps the centralized sublibrary CI form merged in #951 — .github/workflows/SublibraryCI.yml continues to call SciML/.github/.github/workflows/sublibrary-tests.yml@v1, and no bespoke CI_<sublib>.yml is restored. No workflow files are changed.

The canonical pattern (matching OrdinaryDiffEq)

The reusable sublibrary-tests.yml@v1 runs Pkg.test on the root package with env GROUP set to <sub> (Core) or <sub>_<Section> (e.g. <sub>_QA). The reference (SciML/OrdinaryDiffEq.jl test/runtests.jl) handles this by having the root runtests detect a sublibrary GROUP, activate lib/<sub>, translate the suffix to a clean section name, set it as a distinct env var, and Pkg.test the sublibrary. The sublibrary runtests reads only that clean env var (e.g. OrdinaryDiffEqRosenbrock reads just get(ENV, "ODEDIFFEQ_TEST_GROUP", "ALL")).

Earlier revisions of this PR instead put a per-package prefix-strip shim (_G/_SUB/startswith) in every sublibrary runtests and left the root runtests with no sublibrary dispatcher — so the sublibrary tests never actually ran under SublibraryCI. This revision adopts the OrdinaryDiffEq design exactly.

What changed

Root test/runtests.jl — canonical translating dispatcher prelude

Added at the top, reading the raw GROUP env before the main-package logic lowercases/parses it:

• _detect_sublibrary_group (copied verbatim from OrdinaryDiffEq): isdir(lib/<group>) → (group, "Core"), else scan underscores right-to-left for the longest lib/<prefix> match → (prefix, suffix), else (group, "Core").
• If base_group is a lib/<sub> directory: Pkg.activate(lib/<sub>), on Julia < 1.11 transitively Pkg.develop the activated project's [sources] path deps (verbatim OrdinaryDiffEq block, never developing the active project itself), then withenv("NLS_TEST_GROUP" => test_group) run Pkg.test(base_group; julia_args=["--check-bounds=auto","--depwarn=yes"], force_latest_compatible_version=false, allow_reresolve=true).
• Otherwise (all/downstream/nopre/bounds/cuda/wrappers/trim/…): control falls through to the existing main-package logic, preserved byte-for-byte in the else branch (its TEST_ARGS parsing, lowercase(GROUP), Enzyme/Mooncake/CUDA EXTRA_PKGS, worker config, ReTestItems.runtests, and trim activation are unchanged).

Each lib/<sub>/test/runtests.jl — clean-env reader, no shim

The _G/_SUB/startswith prefix-strip shim is replaced by:

const GROUP = get(ENV, "NLS_TEST_GROUP", "All")

The standard Core/QA(/GPU via cuda) dispatch in each file is otherwise unchanged — functional items run under Core/All, Aqua/JET/ExplicitImports/alloc under QA/All, CUDA under GPU/cuda. (SimpleNonlinearSolve's now-unused parse_test_args/get_from_test_args_or_env helpers, which only fed the GROUP read, are dropped in favor of the plain get(ENV, …).) This matches OrdinaryDiffEqRosenbrock/test/runtests.jl.

The pre-existing @itesitem → @testitem fix in SimpleNonlinearSolve/test/core/allocation_tests.jl (so the QA allocation test actually runs) is retained.

Standard groups

• Core — all functional/correctness tests (solver tests, adjoint, wrappers, basic, AD, …). Default-on group.
• QA — Aqua + ExplicitImports (+ JET / allocation where present).
• GPU — CUDA tests, handled by the separate unchanged GPU.yml workflow (which sets GROUP="cuda" on a self-hosted GPU runner).

No test_groups.toml in any sublibrary: with none present, the reusable compute_affected_sublibraries.jl falls back to its standard defaults — Core on [lts, 1.11, 1, pre] + QA on [1], all ubuntu-latest — emitting GROUP="<sub>" (Core) and GROUP="<sub>_QA" (QA). CUDA is covered by GPU.yml, so no [GPU] section is needed.

Verification (run locally, no heavy test runs)

• (a) _detect_sublibrary_group("<sub>") == ("<sub>", "Core") and _detect_sublibrary_group("<sub>_QA") == ("<sub>", "QA") for all 10 sublibraries — all OK.
• (b) For each sublibrary runtests, NLS_TEST_GROUP="Core" enters the Core block and NLS_TEST_GROUP="QA" enters the QA block (non-empty in each case; SciPy intentionally has no QA items, documented in-file) — all OK.
• (c) For the main package, GROUP values nopre/cuda/all/downstream/bounds/wrappers/trim are not detected as sublibraries (isdir(lib/<group>) is false) and fall through to the unchanged main-package logic — all OK.
• The main-package else branch was diffed against master's test/runtests.jl and is byte-identical (after accounting for the one extra level of else indentation).
• All 11 modified runtests.jl parse cleanly and are Runic-clean (Runic 1.5.1). No TOML files changed; no test_groups.toml present.

SublibraryCI.yml is unchanged and kept. No bespoke CI_<sublib>.yml is restored. Only Core / QA / GPU group names remain.

Please ignore until reviewed by @ChrisRackauckas.

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

Closing in favor of the project-model approach (per discussion with @ChrisRackauckas).

This PR added OrdinaryDiffEq's GROUP-dispatch _detect_sublibrary_group + transitive [sources]-develop into runtests.jl, but (a) it only touched runtests.jl with no workflow to drive the dispatcher, and (b) it duplicated the [sources]/registry workaround per repo (including ODE-specific comments that do not apply here).

Instead, the [sources] + registry-staleness handling is moving into the reusable workflow (tests.yml, SciML/.github#52), and sublibrary testing will be a matrixed tests.yml@v1 caller with project: lib/X — no per-repo runtests.jl dispatcher, no copy-pasted workaround. A replacement PR will follow once #52 is promoted to v1.