Align MuladdMacro downgrade floors with PureKLU

[ChrisRackauckas-Claude]

Summary

• raise the MuladdMacro floor from 0.2.1 to 0.2.4 in all 28 affected sublibraries coupled to OrdinaryDiffEqCore
• add a root QA regression that discovers the complete set of Core-coupled MuladdMacro projects, checks that there are 39, and requires every floor to be at least 0.2.4

Why

The sublibrary downgrade workflow selects PureKLU 1.1.1, whose registered
compat requires MuladdMacro >= 0.2.4. A floor of 0.2.1 therefore makes the
minimum-version graph unsatisfiable whenever a Core-coupled project reaches
PureKLU through the local path-source graph.

The first failing jobs exposed seven projects. Once recursive local path-source
handling advanced, another tranche reached the same conflict. This PR now fixes
the invariant rather than enumerating only the first failures: 28 projects are
raised, while the other 11 members of the 39-project set were already at
0.2.4. DiffEqBase remains at 0.2.1 because it is not coupled to
OrdinaryDiffEqCore and is outside this conflict.

The recursive path-source work itself is tracked separately in
julia-actions/julia-downgrade-compat#56.

Local validation

• exact action commit c01bc81127bdcec857160a41d815f0bd9fbd2dcc was run independently for every one of the 39 projects, using a fresh detached worktree, workspace-local temporary directory, and depot per project
  • 39/39 action invocations exited 0
  • 39/39 generated a manifest selecting MuladdMacro 0.2.4
  • all 14 manifests containing PureKLU selected PureKLU 1.1.1
• locked Julia 1.11 Core-group tests with those independent manifests:
  • OrdinaryDiffEqExplicitRK: 19/19 passed
  • OrdinaryDiffEqLowOrderRK: 70/70 passed across its four testsets
  • both tests left their generated Project and Manifest hashes unchanged
• earlier locked validation with action PR Update CI URLs to point to new caching infrastructure #56 passed the originally reported non-BDF projects:
  • Core: 107 passed
  • ExponentialRK: 134 passed / 4 broken
  • FunctionMap: 35 passed
  • Multirate: 88 passed
  • PDIRK: 4 passed
  • SDIRK: 282 passed / 2 broken
• BDF selected MuladdMacro 0.2.4, built successfully, compiled PureKLU 1.1.1, and reached the known inference-test baseline tracked by #3926; this is not reported as a full BDF pass
• root QA on the candidate: 87 passed / 1 failed, including 79/79 new assertions
• unchanged-base QA control: 8 passed / 1 failed; the sole failure in both runs is the same No unapproved public reexports dependency regression tracked by SciMLTesting#29
• full-tree Runic 1.7.0 --check --diff .: passed
• git diff --check and TOML audit: passed; exactly 39/39 Core-coupled floors are at least 0.2.4, and no generated manifest remains in the source worktree

The exact c01 action also exposes a separate locked-test preparation bug for
Core: it promotes SparseArrays from [weakdeps] to [deps] but leaves it in
[targets].test, which Julia's package manager rejects before tests execute.
A fresh standalone Core run reproduced that behavior with unchanged Project and
Manifest hashes. That action bug is separate from this repository's floor
changes; the contaminated combined-worktree attempt was discarded and is not
included in any result above.

Please ignore this PR until it has been reviewed by @ChrisRackauckas.

[ChrisRackauckas-Claude]

Local validation for 7dbe10cd7 on base 8fc48e69d:

• exact final floor/TOML assertions: 14/14 passed
• Runic 1.7.0 --check --diff .: passed repository-wide
• git diff --check: passed; no generated manifests remained
• BDF with the current downgrade action (fab1defb) selected MuladdMacro 0.2.4, built, compiled PureKLU 1.1.1, and passed multiple testsets before reaching the existing inference baseline in BDF coverage-mode inference regression is not isolated to the five NonlinearSolve/LinearSolve upgrades #3926; this is not a full-suite pass
• independent complete-graph validation with action Update CI URLs to point to new caching infrastructure #56 head 28fff453 built all six remaining projects, then passed their locked Core commands with unchanged manifests:
  • Core: 107/107
  • ExponentialRK: 134 passed / 4 broken
  • FunctionMap: 35/35
  • Multirate: 88/88
  • PDIRK: 4/4
  • SDIRK: 282 passed / 2 broken
• candidate root QA: 22 passed / 1 failed, including the new 14/14 assertions
• unchanged-base root QA control: 8 passed / 1 failed; the sole identical failure is the current SciMLTesting#29 reexport-audit regression

The triggering hosted failures classify as 7 PureKLU conflicts fixed here, 7 recursive path-source failures covered by julia-actions/julia-downgrade-compat#56, and 1 DiffEqDevTools hosted runner/log-loss or non-reproducing failure. An isolated exact-action DiffEqDevTools replay exited 0.

The initial combined six-project build reused generated state and is excluded. Every authoritative six-project result used a separate worktree, depot, temporary directory, and locked manifest.

[ChrisRackauckas-Claude]

Expanded the fix after the next downgrade tranche reached the same conflict.

Commit c6590bee41bf2e1bd9eb42aa21820913954c8e1c raises the remaining 21 affected floors, so the PR now changes all 28 Core-coupled projects that were below MuladdMacro 0.2.4. The QA regression now discovers and checks the complete 39-project set; 11 members were already compatible.

Fresh validation used exact action commit c01bc81127bdcec857160a41d815f0bd9fbd2dcc with one detached worktree, temporary directory, and depot per project:

• 39/39 action runs exited 0 and generated manifests with MuladdMacro 0.2.4
• all 14 PureKLU-containing manifests selected PureKLU 1.1.1
• locked ExplicitRK Core tests passed 19/19
• locked LowOrderRK Core tests passed 70/70
• Project and Manifest hashes stayed unchanged through both locked tests
• Runic 1.7.0 and git diff --check passed
• root QA passed all 79 new assertions; its sole remaining failure is the same inherited public-reexport failure reproduced on the unchanged base and tracked by SciMLTesting#29

The independent Core locked command reproduced a separate action defect before package tests run: c01 promotes SparseArrays from [weakdeps] to [deps] while retaining it in [targets].test. That result is documented as an action prerequisite, not attributed to this floor patch. The earlier shared-worktree experiment was discarded because sequential source rewriting contaminated later projects.