Allow AlgebraicMultigrid v2 in the AMG extension compat

[ChrisRackauckas-Claude]

Note

Please ignore until reviewed by @ChrisRackauckas. Opened as a draft.

What

One-line [compat] bump for the LinearSolveAlgebraicMultigridExt weak dependency:

-AlgebraicMultigrid = "1"
+AlgebraicMultigrid = "1, 2"

(plus a patch version bump 3.85.1 → 3.85.2 so it can be released).

Why

The AMG extension's weak-dependency compat was pinned to AlgebraicMultigrid = "1", which excludes AMG v2.0. Because the modern stack (e.g. OrdinaryDiffEq v7) requires LinearSolve >= 3.60 — exactly the LinearSolve versions that carry this weakdep — AMG 2.0 becomes unsatisfiable alongside it, even though AMG 2.0's own LinearSolve compat is "2 - 3" (which includes the current 3.85.x).

The Julia resolver hides the real edge, reporting things like:

AlgebraicMultigrid restricted by compatibility requirements with LinearSolve to versions: 1.0.0 - 1.2.0 — no versions left

with no transitive explanation. The give-away: replicating AMG 2.0's exact declared dependencies as a top-level project resolves fine with LinearSolve 3.85.1, while the AMG package itself does not — isolating the constraint to this weak-compat edge in LinearSolve.

This was surfaced while fixing the DiffEqDocs.jl build on the OrdinaryDiffEq v7 migration (SciML/DiffEqDocs.jl#869/#870), where AMG 2.0 could not be installed next to the v7 stack.

Extension code is unchanged / already compatible

ext/LinearSolveAlgebraicMultigridExt.jl uses only RugeStubenAMG(), SciMLBase.init(amg_alg, A, b), solve!(solver; maxiter, reltol) and solver.b — all identical between AMG 1.x and 2.0 (AMGSolver{T} still has a .b::Vector{T} field; AMG re-exports CommonSolve.{init, solve!}). No source change is required.

Verification (local, Julia 1.12.6, current General registry)

• OrdinaryDiffEq 7.0.0 + LinearSolve 3.85.x + AlgebraicMultigrid 2.0.0 + SciMLBase 3.21.0 now resolve together (fails without this bump).
• The existing AlgebraicMultigridJL test block in test/Core/basictests.jl passes verbatim against AMG 2.0.0 — default Ruge-Stuben, SmoothedAggregationAMG, tight-tolerance solve, and the rectangular-matrix AssertionError (4/4 pass).
• A 1600×1600 SPD Poisson solve via solve(LinearProblem(A,b), AlgebraicMultigridJL()) on AMG 2.0.0 converges (rel. residual 1.3e-8).

🤖 Generated with Claude Code