ci: add codes-ci-ubuntu24-mpich image (source-built MPICH)

.github/workflows/ubuntu24-mpich-image.yml

@@ -0,0 +1,74 @@
+name: Ubuntu 24.04 MPICH CI image
+
+# Builds the ubuntu-24.04 + source-MPICH image (ci/ubuntu24-mpich/Dockerfile) used by
+# build.yml's ubuntu-24.04 x mpich legs (build matrix, sanitizers, coverage) and
+# publishes it to GHCR. Ubuntu 24.04's apt mpich is broken on the runners
+# (ch4:ucx -> MPI_Init abort on the verbs device + singleton launches); this
+# image ships a source mpich built --with-device=ch4:ofi that forms a real
+# multi-rank world. See ci/ubuntu24-mpich/Dockerfile for the full rationale.
+#
+# Deliberately NOT triggered on every PR: the image only needs rebuilding when
+# its recipe changes. It runs on:
+#   * pushes to master that touch the Dockerfile or this workflow,
+#   * a weekly schedule (so ubuntu:24.04 base + libfabric security updates get
+#     picked up), and
+#   * manual dispatch — the BOOTSTRAP path: run this once (from any branch) to
+#     publish :latest, then make the GHCR package PUBLIC, before build.yml's
+#     ubuntu-24.04 mpich legs can pull it.
+#
+# Published as ghcr.io/codes-org/codes-ci-ubuntu24-mpich — :latest plus a :<sha> tag so
+# a build.yml change can pin a specific image if a Dockerfile bump regresses.
+on:
+  push:
+    branches:
+      - master
+    paths:
+      - 'ci/ubuntu24-mpich/Dockerfile'
+      - '.github/workflows/ubuntu24-mpich-image.yml'
+  schedule:
+    # Weekly, Monday 08:00 UTC — refresh the ubuntu:24.04 base + apt packages.
+    - cron: "0 8 * * 1"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  packages: write
+
+jobs:
+  build-and-push:
+    name: build + push codes-ci-ubuntu24-mpich
+    runs-on: ubuntu-24.04
+    steps:
+      - name: Checkout CODES
+        uses: actions/checkout@v4
+
+      - name: Log in to GHCR
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          # The Dockerfile COPYs nothing from the build context (mpich is fetched
+          # inside the build), so the context is just the directory holding it.
+          # `file` defaults to {context}/Dockerfile = ci/ubuntu24-mpich/Dockerfile.
+          context: ci/ubuntu24-mpich
+          # CI builds amd64 (the runner arch that build.yml's 24.04 legs run on).
+          platforms: linux/amd64
+          # Don't publish from PR builds (a PR can't be trusted with
+          # packages:write and shouldn't move :latest); push only on
+          # master / schedule / dispatch.
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+            ghcr.io/codes-org/codes-ci-ubuntu24-mpich:latest
+            ghcr.io/codes-org/codes-ci-ubuntu24-mpich:${{ github.sha }}
+          # Persist layer cache so an unchanged Dockerfile is a fast no-op and a
+          # version bump only rebuilds the mpich layer onward.
+          cache-from: type=gha
+          cache-to: type=gha,mode=max

ci/ubuntu24-mpich/Dockerfile

@@ -0,0 +1,58 @@
+# ubuntu-24.04 CI image with a source-built MPICH, for the build.yml jobs that
+# run ctest with MPICH on ubuntu-24.04: the build matrix's 24.04 x mpich legs,
+# the ASan/UBSan sanitizer lanes, and the coverage job.
+#
+# Why this image exists
+# ---------------------
+# Ubuntu 24.04's apt mpich (4.2.0-5build3) is built --with-device=ch4:ucx and is
+# broken on the GitHub-hosted runners in TWO independent ways:
+#   1. UCX probes the runner's InfiniBand-verbs device and tries the RDMA (rc)
+#      transport, aborting MPI_Init with "ibv_create_srq() failed: Operation not
+#      supported" -- every MPI-initializing test dies.
+#   2. Even past that, mpiexec starts every rank as its own size-1 singleton, so
+#      optimistic (--sync=3) tests silently run sequential.
+# There is no fixed mpich in the noble apt archive. So we build MPICH from source
+# with --with-device=ch4:ofi (libfabric, sockets/TCP -- no UCX, no verbs), which
+# forms a real multi-rank world on the runners. The mpi-multirank-sanity ctest is
+# the tripwire: it fails if the launcher ever regresses to singletons again.
+#
+# The base stays ubuntu:24.04 so these legs keep exercising the modern toolchain
+# (gcc-13 / clang) they exist for -- only the MPI is swapped for our own build.
+#
+# Published as ghcr.io/codes-org/codes-ci-ubuntu24-mpich by
+# .github/workflows/ubuntu24-mpich-image.yml. Like codes-ci-full, make the GHCR package
+# PUBLIC after the first publish so build.yml (including fork PRs) can pull it
+# anonymously.
+FROM ubuntu:24.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Toolchain + CODES build deps (cmake/ninja/pkg-config/flex/bison), BOTH compilers
+# the matrix drives (gcc via build-essential + clang), lcov for the coverage job,
+# and the tools to fetch/build MPICH. Deliberately NO apt mpich -- we build our
+# own below and put it ahead of everything on PATH.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        build-essential clang \
+        cmake ninja-build pkg-config \
+        flex bison \
+        lcov \
+        git ca-certificates wget \
+    && rm -rf /var/lib/apt/lists/*
+
+# --- MPICH from source, ch4:ofi device (embedded libfabric, sockets/TCP). No UCX
+#     and no verbs probing, reliable PMI wire-up -> mpiexec forms a real multi-
+#     rank world on the CI runners. Installed to /opt/mpich and put on PATH so
+#     find_package(MPI) discovers mpicc/mpicxx/mpiexec with no CODES/ROSS config
+#     change. Bump MPICH_VERSION to update -- that's the whole maintenance story.
+ARG MPICH_VERSION=4.2.3
+RUN wget -q "https://www.mpich.org/static/downloads/${MPICH_VERSION}/mpich-${MPICH_VERSION}.tar.gz" \
+    && tar xzf "mpich-${MPICH_VERSION}.tar.gz" \
+    && cd "mpich-${MPICH_VERSION}" \
+    && ./configure --prefix=/opt/mpich --with-device=ch4:ofi --disable-fortran \
+    && make -j"$(nproc)" \
+    && make install \
+    && cd / \
+    && rm -rf "mpich-${MPICH_VERSION}" "mpich-${MPICH_VERSION}.tar.gz"
+
+# Put our mpich first so find_package(MPI) picks it over anything else.
+ENV PATH=/opt/mpich/bin:$PATH