id: ADR-0028 title: Build Ipopt 3.14.20 with MUMPS and ASL from pinned sources in a container; probe HSL at runtime status: accepted date: 2026-09-13 deciders: [paul-heyse] level: decision principles: [DM-28, DM-43, DM-48, DM-45] blueprint: [§18.3, §6.13, §26] review: docs/design_review/reviews/design_review_arrow-native-idaes-core-blueprint-rev2_2026-09-13.md#7-principle-findings evidence: Proposed supersedes: [] superseded-by: null revisit: A platform’s recipe cannot be reproduced (macOS or Windows at phase 1b), or a wheel needs the native backend and the linkage question comes due verification: python / parity test_00_preflight (Ipopt 3.14.x, MUMPS, ASL, SolverFactory("ipopt").available(False)); solvers-image-rebuild-check (weekly, reproducibility of library checksums)

ADR-0028: Build Ipopt 3.14.20 with MUMPS and ASL from pinned sources in a container; probe HSL at runtime

Context

Blueprint §18.3 requires Ipopt 3.14 or newer for GetIpoptCurrentIterate/GetIpoptCurrentViolations, and §6.13 requires an explicit probe_host operation that records the available linear solvers. The first review’s F10 asked when the capability probe runs and where its result is persisted; the host’s system Ipopt is 3.11.9.

Scope

Binds how solvers are acquired for CI and development, and the fallback policy when a profile names an unavailable linear solver. Wheel linkage is deliberately left open (register row R-09, a phase-1 ADR).

Drivers

Iteration statistics as data require Ipopt 3.14; parity iteration counts require MUMPS ordering comparable to IDAES’s binaries; HSL cannot be redistributed; a capability claimed but unprobed is a lie (DM-43).

Options

idaes get-extensions in CI — rejected: it fetches IDAES’s binaries, which is neither clean-room nor pinned. conda-forge ipopt on Linux — rejected for CI: the recipe would not be ours to reproduce; it stays the documented Windows route. Bundle HSL — impossible: licensing.

Outcome

docker/solvers/ builds Ipopt releases/3.14.20, ThirdParty-Mumps releases/3.0.14 and ThirdParty-ASL releases/2.1.0 from checksum-verified tarballs on a digest-pinned ubuntu:24.04, with netlib reference BLAS/LAPACK and --without-metis so MUMPS ordering matches IDAES’s. HSL is never built; probe_host records ma27/ma57/ma86/ma97 availability at runtime.

Consequences

A profile naming ma57 on a host without HSL fails with capability.backend unless it declares fallback_linear_solver = mumps, in which case the fallback is a selected policy recorded in runs.resolved_options.

Compensating controls

runtime.host_capabilities persists the probe; runs.resolved_options records the options actually passed to CreateIpoptProblem, so a reproduction on a different host is either identical or visibly different.

Confirmation

python / parity runs in the container and fails the pre-flight rather than skipping; the weekly rebuild-without-cache job compares library checksums.

Pros and cons

Building from source costs a container image and a weekly job; using someone’s binaries costs reproducibility and the clean-room story.

More information

Blueprint §18.3 (in-process Ipopt), §6.13 (execution and evidence), §26 (residual risks); review finding F10; register rows R-08 and R-09.

Status history

• 2026-09-13 — accepted with the repository-seeding pull request (backfilled from blueprint revision 3).