ci.yml

name: CI

# What this workflow asserts.
#
# A C++ standard library is not portable in the sense a program is: it is
# configured for one C library and compiled against that library's headers. The
# claim here is therefore not "it builds" but "a program above it does the
# things a C++ program does", and the one that settles it is an exception thrown
# across frames and caught --- because that is the path a build proves nothing
# about. Everything else this package could check, a runtime that was linked but
# never worked would also satisfy.

on:
  push:
    branches: [main]
  pull_request:
  workflow_dispatch:
    inputs:
      mcpp_ref:
        description: "Branch of mcpp-community/mcpp to build and test against (empty = the released pin)"
        required: false
        default: ""

env:
  MCPP_SOURCE_REF: ${{ github.event.inputs.mcpp_ref || vars.MCPP_SOURCE_REF }}
  MCPP_VERSION: 2026.8.27.1
  XLINGS_VERSION: v2026.8.17.2
  XLINGS_NON_INTERACTIVE: '1'

jobs:
  runtime:
    name: build the runtime and run what stands above it (${{ matrix.toolchain }})
    runs-on: ubuntu-24.04
    # ⚠️ Raised for the self-build step below, which is temporary. See the note
    # there: when mcpp#486 ships, that step goes and so does this.
    timeout-minutes: 90
    strategy:
      fail-fast: false
      matrix:
        toolchain: ['llvm@22.1.8']
    defaults:
      run:
        shell: bash
    steps:
      - uses: actions/checkout@v4

      - name: Install xlings and mcpp
        run: |
          curl -fsSL https://raw.githubusercontent.com/openxlings/xlings/main/tools/other/quick_install.sh \
            | bash -s "$XLINGS_VERSION"
          echo "$HOME/.xlings/subos/current/bin" >> "$GITHUB_PATH"

      - name: Install mcpp
        run: bash tools/install-mcpp.sh

      # ⭐ THE ENGINE EVERY STEP BELOW WILL USE, COMPARED AGAINST THE ONE BUILT.
      #
      # The step above appends a directory to GITHUB_PATH, which governs the
      # steps that follow it, so that step cannot observe its own effect. Whether
      # the appended spelling is one the runner accepts is a property of the
      # runner and differs between hosts. Left unasserted, a cross-validation run
      # on a host that ignores it builds this ecosystem with the released engine
      # and reports the result as though the change under review had been tested.
      - name: The engine on PATH is the one under review
        run: |
          set -euo pipefail
          if [ -z "${MCPP_UNDER_REVIEW:-}" ]; then
            echo "  no source reference: this run tests $(mcpp --version)"
            exit 0
          fi
          on_path=$(mcpp --version | awk '{print $2}')
          if [ "$on_path" != "$MCPP_UNDER_REVIEW" ]; then
            echo "::error::PATH resolves mcpp $on_path, and the build under review is $MCPP_UNDER_REVIEW"
            echo "        the directory appended to GITHUB_PATH did not take effect on this host"
            command -v mcpp
            exit 1
          fi
          echo "  every step below runs $on_path, built from $MCPP_SOURCE_REF"

      - name: Select the toolchain
        run: |
          spec='${{ matrix.toolchain }}'
          mcpp toolchain install "${spec%@*}" "${spec#*@}"
          mcpp toolchain default "$spec"

      # THE C LIBRARY THIS RUNTIME IS CONFIGURED FOR, AS WRITTEN ON THE BRANCH
      # UNDER TEST RATHER THAN AS PUBLISHED. The reasoning is in the script.
      - name: The stack, as written on this branch
        run: bash tools/branch-graph.sh '${{ github.head_ref || github.ref_name }}'

      - name: The runtime builds
        run: mcpp build

      # ⚠️ THE ONE OBSERVATION A BUILD CANNOT MAKE.
      #
      # An unwinder that cannot find the program's frame descriptions compiles,
      # links, and runs every path that does not throw. Measured while this
      # package was written: the whole example above printed its first line and
      # then `libc++abi: terminating due to uncaught exception', with
      # _Unwind_Backtrace walking zero frames --- and nothing named the cause.
      # So the throw is the check, and the destructor beside it is what says the
      # unwind was correct rather than merely non-fatal.
      # ⭐ THE UNWINDER, AND EVERY POSITION `__config_site` DECLARES.
      #
      # The throw and the destructor beside it are what say the runtime is
      # really there. The rest of the probe asks whether the CLAIMS this package
      # makes about the environment beneath are true: `std::filesystem` over
      # openkal.fs, `std::random_device` over openkal.random, and the two
      # operations openkal has no atom for --- which must be REFUSED rather than
      # ignored, and which a probe checking only the supported ones would pass
      # for a port that silently accepted them.
      - name: A C++ program above it throws, catches, and uses what is declared
        run: |
          cd examples/cxx && mcpp run 2>&1 | tee out.log
          grep -q -- '-- failures: 0 --' out.log
          ! grep -q '^FAIL:' out.log

      - name: import std above it
        run: |
          cd examples/import-std && mcpp run 2>&1 | tee out.log
          grep -q 'import std above openkal: 2 4 7' out.log

      # WHAT __config_site DECLARES AND WHAT THE PORT BENEATH PROVIDES ARE
      # RECONCILED HERE, RATHER THAN BEING CHECKED ONCE AND ASSUMED AFTERWARDS.
      #
      # `__config_site` is this package's statement about the environment it was
      # configured for, and the environment is openkal-musl. A statement that
      # drifts from what the port actually provides does not fail to build and
      # does not fail to link: it produces a program that takes a path the
      # environment cannot support, and reports nothing.
      #
      # `_LIBCPP_HAS_TERMINAL` is the position where that happened. It gates
      # `std::__is_posix_terminal`, which is `isatty(fileno(stream))` and nothing
      # else. The port answered `isatty` with TCGETS while musl asks with
      # TIOCGWINSZ, so every `isatty` returned 0 --- for a real terminal as
      # readily as for a pipe --- and `std::print` never took its terminal path.
      # Nothing failed; a program deciding on colour or on line buffering decided
      # wrongly and in silence.
      #
      # The remedy was to fix the port rather than to withdraw the declaration,
      # so the declaration is now true and this step is what keeps it true.
      #
      # THE CRITERION IS A RELATION AND NOT A VALUE. Asserting "a terminal is
      # detected" would need a terminal; asserting "a pipe is not" would pass
      # throughout the defect. What must hold is that the two DIFFER, and that
      # they differ the way the system's own C library says they do.
      - name: What the runtime declares is what the port provides
        run: |
          set -euo pipefail
          d="$(mktemp -d)"; mkdir -p "$d/src"
          cat > "$d/mcpp.toml" <<TOML
          [package]
          name    = "termdecl"
          version = "0.1.0"
          [dependencies]
          openkal-llvm-runtime = { path = "$PWD" }
          [targets.termdecl]
          kind = "bin"
          main = "src/main.cpp"
          TOML
          sed -i 's/^          //' "$d/mcpp.toml"
          cat > "$d/src/main.cpp" <<'CPP'
          #include <cstdio>
          #include <unistd.h>
          // The declaration this step exists to reconcile. If the package stops
          // claiming a terminal, the program says so rather than failing to
          // compile: the point is to compare the claim against the behaviour,
          // and a build error would compare nothing.
          int main() {
          #if defined(_LIBCPP_HAS_TERMINAL) && _LIBCPP_HAS_TERMINAL
              std::printf("declared=1 isatty=%d\n", isatty(1));
          #else
              std::printf("declared=0 isatty=%d\n", isatty(1));
          #endif
          }
          CPP
          sed -i 's/^          //' "$d/src/main.cpp"
          ( cd "$d" && mcpp build )
          bin="$(find "$d/target" -type f -name termdecl | head -1)"
          test -n "$bin" || { echo "::error::the probe did not build"; exit 1; }

          # The control: the system's own C library, through the same harness.
          # Without it a `script` that fails to allocate a pty would make the
          # runtime look wrong.
          printf '#include <stdio.h>\n#include <unistd.h>\nint main(void){ printf("%%d\\n", isatty(1)); return 0; }\n' > "$d/ctrl.c"
          cc "$d/ctrl.c" -o "$d/ctrl"
          ctrl_pipe="$("$d/ctrl" | cat | tr -d '\r')"
          ctrl_tty="$(script -qec "$d/ctrl" /dev/null | tr -d '\r' | head -1)"
          [ "$ctrl_pipe" = 0 ] && [ "$ctrl_tty" = 1 ] \
            || { echo "::error::the harness cannot tell a pty from a pipe (control gave $ctrl_pipe/$ctrl_tty)"
                 exit 1; }

          out_pipe="$("$bin" | cat | tr -d '\r')"
          out_tty="$(script -qec "$bin" /dev/null | tr -d '\r' | head -1)"
          echo "  control: pipe=$ctrl_pipe tty=$ctrl_tty"
          echo "  runtime: $out_pipe / $out_tty"

          declared="${out_pipe#declared=}"; declared="${declared%% *}"
          pipe_v="${out_pipe##*isatty=}"
          tty_v="${out_tty##*isatty=}"

          if [ "$declared" = 1 ]; then
            # The claim is that a terminal can be detected, so the two must
            # differ and must differ as the system's own library does.
            [ "$pipe_v" = "$ctrl_pipe" ] && [ "$tty_v" = "$ctrl_tty" ] \
              || { echo "::error::the runtime declares _LIBCPP_HAS_TERMINAL but the port answers $pipe_v/$tty_v where the system answers $ctrl_pipe/$ctrl_tty"
                   exit 1; }
            echo "  ok  the declaration holds: a terminal is distinguished from a pipe"
          else
            # The claim is that it cannot. Then it must not appear to: a
            # declaration of 0 beside a working isatty is also a drift, and the
            # remedy is to raise the declaration rather than leave it stale.
            [ "$tty_v" = "$ctrl_tty" ] \
              && { echo "::error::the runtime declares no terminal support while the port detects one; the declaration is stale"
                   exit 1; }
            echo "  ok  the declaration holds: no terminal support is claimed and none is present"
          fi

      # ⭐⭐ THE SAME PROGRAM ON A MACHINE WITH NO OPERATING SYSTEM.
      #
      # Everything above this step runs on a host, and a host has a C library, a
      # C++ runtime and an unwinder already installed. A program that reaches
      # one of them by mistake still works — so those steps can pass without
      # having exercised this package's own copies at all.
      #
      # There is nothing here to reach. The C library is openkal-musl, the
      # standard library and the unwinder are this package's, and beneath them
      # is firmware whose whole interface is `ecall`. ⇒ This is the step that
      # cannot go green by accident, which is why the design document makes it
      # the acceptance criterion rather than one more row.
      #
      # ⚠️ The assertion is on the OUTPUT and not on the exit status: firmware
      # that never reaches the payload exits zero, and so does a payload whose
      # console writes go nowhere.
      - name: Install the emulator
        if: matrix.toolchain == 'llvm@22.1.8'
        run: |
          xlings install xim:qemu-riscv -y
          XLINGS_HOME="$HOME/.mcpp/registry" xlings install xim:qemu-riscv -y

      - name: The same source on bare metal, with exceptions
        if: matrix.toolchain == 'llvm@22.1.8'
        run: |
          set -euo pipefail
          cd examples/same-source
          Q=$(ls -d "$HOME"/.mcpp/registry/data/xpkgs/xim-x-qemu-riscv/*/bin/qemu-system-riscv64 | head -1)
          # ⚠️ Anchored on the BARE NAME, so a checkout that already carries a
          # path is left alone rather than getting a path inside a path. The
          # manifest is required to carry the bare name; the note beside it says
          # why, and this is the step that relies on it.
          grep -q '"qemu-system-riscv64"' mcpp.toml \
            || { echo "::error::the manifest no longer carries the bare emulator name"; exit 1; }
          sed -i "s|\"qemu-system-riscv64\"|\"$Q\"|" mcpp.toml
          # ⚠️ `--target riscv64-none-elf`, AND THE FLAG IS THE WHOLE STEP.
          #
          # This manifest carries no `[build] target`, so a bare `mcpp run`
          # builds for the HOST and passes — the four lines below appear either
          # way, which is the manifest's own claim and not an accident. A step
          # named "on bare metal" that omits the flag therefore reports green
          # over a host build and has never once reached OpenSBI. Measured
          # 2026-08-24: the artefact was `target/x86_64-linux-gnu/…`.
          mcpp run --target riscv64-none-elf 2>&1 | tee out.log
          # The emulator's own banner, so a run that never left the host cannot
          # satisfy this step by printing the four application lines.
          grep -q 'Boot HART'       out.log   # firmware ran; this is OpenSBI
          grep -q 'sorted: 2 4 7'   out.log   # containers + algorithms + the allocator
          grep -q 'caught: 42'      out.log   # the unwinder found the handler
          grep -q 'unwound: true'   out.log   # ⭐ and ran a destructor on the way
          grep -q 'import std over openkal: ok' out.log

      # ⭐⭐ AND THE SAME SOURCE ON THIS MACHINE, WHICH IS WHAT MAKES THE STEP
      # ABOVE A DEMONSTRATION RATHER THAN AN ILLUSTRATION.
      #
      # Nothing is edited between the two commands — no `#if`, no second
      # directory, no second source. The only thing that differs is the memory
      # layout, and `build.mcpp` states it behind a condition on the target OS
      # because that is a statement about the machine.
      - name: The same source, on this machine, over openkal-linux
        if: matrix.toolchain == 'llvm@22.1.8'
        run: |
          set -euo pipefail
          cd examples/same-source && mcpp run 2>&1 | tee host.log
          grep -q 'import std over openkal: ok' host.log
          # The four lines are the same four lines.
          #
          # ⚠️ `tr -d '\r'` ON BOTH SIDES, AND IT IS NOT COSMETIC. The bare-metal
          # run reaches the console through an emulated 16550 UART, and a serial
          # console terminates lines with CRLF; the native run does not. Without
          # this the diff reports four differing lines whose visible text is
          # identical, which reads as a failure of the claim being tested rather
          # than of the transport carrying it. The assertion is about what the
          # program printed, not about how the bytes arrived.
          diff <(grep -E '^(sorted|caught|unwound|import std over openkal):' out.log  | tr -d '\r') \
               <(grep -E '^(sorted|caught|unwound|import std over openkal):' host.log | tr -d '\r')

      # ⭐⭐ AND THE SAME SOURCE FOR TWO MACHINES THIS ONE IS NOT.
      #
      # The two steps above prove the source does not know which machine it is
      # for. These prove the BUILD does not need to be on it: one Linux host
      # produces a PE and a Mach-O, and the jobs below run them on the real
      # thing with nothing installed.
      #
      # ⚠️ THE ARTEFACT IS THE ARGUMENT, WHICH IS WHY THOSE JOBS INSTALL NOTHING.
      # Not mcpp, not a compiler, not a C runtime — the program carries its C
      # library, its C++ runtime and its unwinder, and what remains is the
      # operating system it was built for. A run that needed a redistributable
      # installed first would be demonstrating something weaker.
      - name: The same source, built here for Windows and for macOS
        if: matrix.toolchain == 'llvm@22.1.8'
        run: |
          set -euo pipefail
          cd examples/same-source
          mkdir -p "$RUNNER_TEMP/cross"
          for t in x86_64-windows-gnu aarch64-macos; do
            rm -rf target
            mcpp build --target "$t"
            a=$(find target -type f \( -name 'openkal-same-source' -o -name '*.exe' \) | head -1)
            [ -n "$a" ] || { echo "::error::$t produced no artefact"; exit 1; }
            echo "$t → $(file -b "$a")"
            cp "$a" "$RUNNER_TEMP/cross/"
          done
          # ⚠️ The format is asserted here rather than left to the run jobs. A
          # run that fails tells you the program did not work; this tells you
          # what was produced, and the two failures need different fixes.
          file "$RUNNER_TEMP/cross/openkal-same-source.exe" | grep -q 'PE32+ executable'
          file "$RUNNER_TEMP/cross/openkal-same-source"     | grep -q 'Mach-O 64-bit arm64'

      - uses: actions/upload-artifact@v4
        if: matrix.toolchain == 'llvm@22.1.8'
        with:
          name: cross-artifacts
          path: ${{ runner.temp }}/cross/
          if-no-files-found: error

  # ---------------------------------------------------------------------------
  # ⭐⭐ THE ACCEPTANCE CRITERION FOR PORTABILITY OF THE ARTEFACT.
  #
  # A cross build that produces a well-formed file proves the compiler was told
  # the right target. It does not prove the program runs, and every difference
  # this ecosystem has had to find on these two platforms — the loader-bootstrapped
  # thread-local, the unwinder's search for its own tables, the personality
  # routine — links successfully and fails at run time.
  #
  # ⚠️ These jobs deliberately have NO toolchain steps. If one is ever added
  # because "the program needs it", that is the finding, not the fix.
  run-on-windows:
    name: the artefact built on Linux runs on Windows
    needs: runtime
    runs-on: windows-2022
    timeout-minutes: 10
    defaults:
      run:
        shell: bash
    steps:
      - uses: actions/download-artifact@v4
        with: { name: cross-artifacts, path: art }
      - name: It runs, and it unwinds
        run: |
          set -euo pipefail
          ./art/openkal-same-source.exe 2>&1 | tee out.log
          grep -q 'sorted: 2 4 7'   out.log
          grep -q 'caught: 42'      out.log
          # ⭐ The line a link cannot fake: a destructor ran during the unwind,
          # so libunwind found `.eh_frame` by reading the image rather than by
          # asking the operating system to enumerate modules.
          grep -q 'unwound: true'   out.log
          grep -q 'import std over openkal: ok' out.log

  run-on-macos:
    name: the artefact built on Linux runs on macOS
    needs: runtime
    runs-on: macos-14
    timeout-minutes: 10
    steps:
      - uses: actions/download-artifact@v4
        with: { name: cross-artifacts, path: art }
      - name: It runs, and it unwinds
        run: |
          set -euo pipefail
          # ⚠️ The executable bit does not survive an artefact upload.
          chmod +x art/openkal-same-source
          # ⚠️ AND THE SIGNATURE DOES. arm64 macOS refuses an unsigned image, so
          # this is asserted before the run: a failure here is "the linker did
          # not ad-hoc sign it", which is a different repair from "the program
          # crashed".
          codesign -dv art/openkal-same-source 2>&1 | grep -q 'adhoc\|Signature'
          # ⚠️ A CRASH IS DIAGNOSED HERE RATHER THAN GUESSED AT LATER. This is
          # the first time an artefact of this stack has run on this system, and
          # the interesting failures — the entry point's assumptions about what
          # the kernel hands it, the thread pointer, the two borrowed names —
          # all look identical from outside: `Segmentation fault: 11`.
          # One CI cycle that prints a backtrace is worth several that do not.
          if ! ./art/openkal-same-source > out.log 2>&1; then
            echo "--- how far the loader got ---"
            # ⚠️ To a file, then read. Piping into `tail` interleaves dyld's
            # output with the shell's own report of the signal, and the last
            # lines — the ones that say which initializer was running — are the
            # ones that get lost.
            DYLD_PRINT_INITIALIZERS=1 ./art/openkal-same-source > dyld.log 2>&1 || true
            tail -40 dyld.log
            echo "--- it did not run; what the debugger saw ---"
            # ⭐ `lr` IS THE DATUM. A jump to address 0 leaves no frame to
            # unwind, so `bt` says only "frame #0: 0x0" — which is the symptom
            # restated. The link register still holds the return address of
            # whoever made that call, and `image lookup` turns it into a name.
            # ⚠️ `-k` AND NOT `-o`. In `--batch` lldb abandons the remaining
            # `-o` commands after the first one that errors, and reading
            # register state at a PC of 0 errors — so the three commands that
            # would have said something never ran. `-k` is the list lldb
            # executes WHEN THE PROCESS CRASHES, which is the case at hand.
            lldb --batch \
                 -k 'register read pc lr sp fp x8 x9 x16 x17' \
                 -k 'image lookup --address $lr' \
                 -k 'thread backtrace all' \
                 -k 'image list -o -f' \
                 -o run \
                 -- ./art/openkal-same-source 2>&1 | tail -70 || true
            echo "--- the image's own dependencies ---"
            otool -L art/openkal-same-source || true
            otool -l art/openkal-same-source | grep -A4 LC_MAIN || true
            cat out.log
            exit 1
          fi
          cat out.log
          grep -q 'sorted: 2 4 7'   out.log
          grep -q 'caught: 42'      out.log
          grep -q 'unwound: true'   out.log
          grep -q 'import std over openkal: ok' out.log

  # ---------------------------------------------------------------------------
  # ⭐⭐ THE HOST DIMENSION — THE HALF OF THE CLAIM THE JOBS ABOVE DO NOT TOUCH.
  #
  # Everything above builds on Linux. That establishes "one host reaches every
  # target" and leaves open the thing an N×N matrix would otherwise have to
  # enumerate: whether the HOST matters. The scheme's answer is that it does not
  # — the target side is a set of packages and the compiler is a retargetable
  # clang, so N hosts × N targets collapses to N implementations plus one tool.
  #
  # ⚠️ THAT IS A CLAIM, AND CLAIMS OF THIS SHAPE HAVE BEEN WRONG HERE BEFORE:
  # the Linux host needed four separate repairs before it reached PE, and every
  # one of them was invisible until a build was actually run. Two more hosts
  # cost two jobs; asserting the collapse without running them costs a paragraph
  # and proves nothing.
  #
  # Each host builds all four targets and RUNS the one that is itself, which is
  # the same criterion the Linux job applies to itself.
  host-dimension:
    name: ${{ matrix.host }} host reaches every target
    needs: runtime
    runs-on: ${{ matrix.runner }}
    timeout-minutes: 120
    strategy:
      fail-fast: false
      matrix:
        include:
          - { host: macOS,   runner: macos-14,     native: aarch64-macos      }
          - { host: Windows, runner: windows-2022, native: x86_64-windows-gnu }
    defaults:
      run:
        shell: bash
    steps:
      - uses: actions/checkout@v4

      - name: Install xlings (Unix)
        if: runner.os != 'Windows'
        run: |
          curl -fsSL https://raw.githubusercontent.com/openxlings/xlings/main/tools/other/quick_install.sh \
            | bash -s "$XLINGS_VERSION"
          echo "$HOME/.xlings/subos/current/bin" >> "$GITHUB_PATH"

      - name: Install xlings (Windows)
        if: runner.os == 'Windows'
        shell: pwsh
        run: |
          irm https://d2learn.org/xlings-install.ps1.txt | iex
          "$env:USERPROFILE\.xlings\subos\current\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append

      - name: Install mcpp
        run: bash tools/install-mcpp.sh

      # ⭐ THE ENGINE EVERY STEP BELOW WILL USE, COMPARED AGAINST THE ONE BUILT.
      #
      # The step above appends a directory to GITHUB_PATH, which governs the
      # steps that follow it, so that step cannot observe its own effect. Whether
      # the appended spelling is one the runner accepts is a property of the
      # runner and differs between hosts. Left unasserted, a cross-validation run
      # on a host that ignores it builds this ecosystem with the released engine
      # and reports the result as though the change under review had been tested.
      - name: The engine on PATH is the one under review
        run: |
          set -euo pipefail
          if [ -z "${MCPP_UNDER_REVIEW:-}" ]; then
            echo "  no source reference: this run tests $(mcpp --version)"
            exit 0
          fi
          on_path=$(mcpp --version | awk '{print $2}')
          if [ "$on_path" != "$MCPP_UNDER_REVIEW" ]; then
            echo "::error::PATH resolves mcpp $on_path, and the build under review is $MCPP_UNDER_REVIEW"
            echo "        the directory appended to GITHUB_PATH did not take effect on this host"
            command -v mcpp
            exit 1
          fi
          echo "  every step below runs $on_path, built from $MCPP_SOURCE_REF"

      # ⭐ THE SAME ENGINE AND THE SAME STACK AS THE LINUX JOB, FROM A DIFFERENT
      # HOST. This job had neither: it installed the released engine and
      # resolved this ecosystem from the index, so a change spanning these
      # repositories was validated on one host of three and reported as
      # validated everywhere.
      - name: Select the toolchain
        run: |
          set -euo pipefail
          # ⚠️ INSTALL, THEN SELECT. `toolchain default` names a toolchain and
          # does not fetch one, so selecting an absent payload fails with
          # `llvm@22.1.8 is not installed` — measured on both rows of this job.
          mcpp toolchain install llvm 22.1.8
          mcpp toolchain default 'llvm@22.1.8'

      - name: The stack, as written on this branch
        run: bash tools/branch-graph.sh '${{ github.head_ref || github.ref_name }}'

      - name: Every target, from this host
        run: |
          set -euo pipefail
          cd examples/same-source
          for t in x86_64-linux-gnu x86_64-windows-gnu aarch64-macos riscv64-none-elf; do
            rm -rf target
            mcpp build --target "$t"
            a=$(find target -type f \( -name 'openkal-same-source' -o -name 'openkal-same-source.exe' \) | head -1)
            [ -n "$a" ] || { echo "::error::$t produced no artefact on this host"; exit 1; }
            echo "$t → $(file -b "$a" 2>/dev/null || echo built)"
          done

      # ⭐ And the one that is this machine, run rather than inspected — the same
      # criterion the Linux job holds itself to, applied from a different host.
      - name: The artefact for this host runs on it
        run: |
          set -euo pipefail
          cd examples/same-source
          rm -rf target
          mcpp run --target '${{ matrix.native }}' 2>&1 | tee out.log
          grep -q 'sorted: 2 4 7'   out.log
          grep -q 'caught: 42'      out.log
          grep -q 'unwound: true'   out.log
          grep -q 'import std over openkal: ok' out.log