docs: add canonical monorepo structure spec (Monorepo.md)

[ChrisRackauckas-Claude]

Adds Monorepo.md, a comprehensive reference for the canonical SciML monorepo standard — the layout future monorepos and contributors should follow — and links it from README.md under a new Monorepo structure section.

The reference implementation throughout is SciML/OrdinaryDiffEq.jl. Every concrete value, file shape, and code snippet was verified against the live @v1 reusable workflows in this repo and OrdinaryDiffEq.jl master before being written down.

What it documents

1. Layout — umbrella root package + lib/<Name>/ full packages (name/uuid/version/authors/[deps]/[sources]/[compat]/[extras]/[targets], src/, test/, LICENSE.md, component-template README.md).
2. [sources] graph — sublibs list in-repo deps by relative path; root lists tested sublibs as {path = "lib/<Name>"}; true dependency direction; never a [sources] entry without a matching [deps]; how compute_affected_sublibraries.jl builds the affected + reverse-dep matrix from [deps].
3. Test structure — every test in exactly one group, each in its own test/<Group>/ folder; dependency-driven per-group Project.toml (no extra deps → no Project.toml, runs in main env, part of All; adds deps → isolated env via Pkg.activate, excluded from All); QA in test/qa; light main [extras]/[targets].test; no top-level test/Project.toml.
4. Group names + env — capitalized/Title-case, default All, never lowercase(get(ENV, ...)); shared <REPO>_TEST_GROUP threaded by group-env-name; root runtests.jl _detect_sublibrary_group dispatcher.
5. test_groups.toml — required versions, optional runner/timeout/num_threads/local_only; version convention Core ["lts","1.11","1","pre"], QA ["1"], GPU ["1"]; GPU is a dep-adding group on a self-hosted runner — no bespoke GPU.yml.
6. Workflows — thin @v1 callers with secrets: "inherit" (SublibraryCI, DowngradeSublibraries, CI, Downgrade, Documentation, Downstream, FormatCheck/RunicSuggestions, SpellCheck, TagBot root + TagBot-Subpackages matrix, DependabotAutoMerge, DocPreviewCleanup, benchmark); downgrade is strict fleet-wide (allow_reresolve: false hardcoded centrally, no input).
7. Repo files — .codecov.yml (comment: false), .typos.toml, .gitignore, per-package LICENSE.md.
8. Formatting — Runic default; Catalyst/JumpProcesses are the JuliaFormatter exceptions; CompatHelper removed fleet-wide; Invalidations not standard.

Notes for review

• Documented against the live @v1 (which equals origin/master / v1.6.0): sublibrary-downgrade.yml / downgrade.yml hardcode allow_reresolve: false with no allow-reresolve input, and documentation.yml has a runner input for GPU-queue docs. The README's older sublibrary-downgrade.yml table still lists an allow-reresolve input; the new doc reflects the current strict-no-input behavior (flagging the discrepancy here for a possible README follow-up).
• A couple of values in the original task brief were corrected to match the live reference: Core/QA version sets are ["lts","1.11","1","pre"] / ["1"] (per ODE's test_groups.toml and the detection script defaults), and Downgrade* callers run on Julia 1.11 in ODE today; the skip input defaults to Pkg,TOML and the caller lists stdlibs + in-repo sublibs (it is not auto-populated centrally).

typos is clean on the new file.

Ignore until reviewed by @ChrisRackauckas.

🤖 Generated with Claude Code