ci.yml

name: CI

# What this workflow asserts about the specification.
#
# A specification is a claim about programs, so the claims are tested by
# building and running programs rather than by reading the text. There are four:
#
#   declarations  the two forms compile, under three compiler families and on
#                 three systems, and declare the same entities
#   substitution  a program's source is invariant under a change of
#                 implementation, and the check that says so fails when it should
#   conformance   the suite in this repository runs against the implementation
#                 for each system and every observation holds
#   composability an implementation that provides three interfaces is examined
#                 for three, rather than failing to link
#
# The three compiler families are covered because the specification is a
# contract and a contract that holds only under the compiler its author used is
# a description of that compiler.

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 }}
  # A version verified to build these packages, not a measured minimum. The pin
  # exists for reproducibility rather than because an older mcpp is known to
  # fail.
  MCPP_VERSION: 2026.8.27.1
  XLINGS_VERSION: v2026.8.17.2
  XLINGS_NON_INTERACTIVE: '1'

jobs:
  # ---------------------------------------------------------------------------
  # The declarations compile, everywhere, in both forms.
  declarations:
    name: declarations (${{ matrix.os }}, ${{ matrix.toolchain }})
    runs-on: ${{ matrix.os }}
    timeout-minutes: 45
    strategy:
      fail-fast: false
      matrix:
        include:
          - { os: ubuntu-24.04, toolchain: 'gcc@16.1.0'  }
          - { os: ubuntu-24.04, toolchain: 'llvm@22.1.8' }
          - { os: macos-14,     toolchain: 'llvm@20.1.7' }
          - { os: windows-2022, toolchain: 'llvm@20.1.7' }
          - { os: windows-2022, toolchain: 'msvc@system' }
    defaults:
      run:
        shell: bash
    steps:
      - uses: actions/checkout@v4

      - name: Install xlings and mcpp (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 and mcpp (Windows)
        if: runner.os == 'Windows'
        shell: pwsh
        run: |
          irm https://d2learn.org/xlings-install.ps1.txt | iex
          # The installer amends the user's environment; a later step in this
          # job reads none of it, so the directory is named here.
          "$env:USERPROFILE\.xlings\subos\current\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append

      - name: Install mcpp
        run: |
          # ⚠️ A LOOP, BECAUSE ONE `xlings update` CAN RETURN A STALE INDEX
          # WITHOUT SAYING SO.
          #
          # The index is published as an artifact behind a pointer, and that
          # pointer propagates asynchronously after a version bump is merged.
          # Measured on release day: an update run four minutes after the merge
          # printed `index updated`, and the install then failed with
          #
          #     package 'mcpp@<ver>' not found in the synced index
          #     (xim@artifact:<an older sha>, ...), synced 0 seconds ago
          #
          # Nothing had gone wrong. The update fetched the PREVIOUS artifact,
          # and "synced 0 seconds ago" describes when it was fetched rather than
          # what it contains — which is why the message reads as freshness.
          #
          # So this is not a retry around flakiness; it is the wait that a
          # single update does not perform. A pin naming a version that was
          # never published still fails, after the last attempt, and says which
          # of the two situations it is.
          for attempt in 1 2 3 4 5 6; do
            xlings update > /dev/null 2>&1 || true
            if # ⚠️ THE PIN MAY NAME THE RELEASE THIS RUN IS VALIDATING, which does
          # not exist yet — that is the whole point of MCPP_SOURCE_REF. Bootstrap
          # from whatever the index has; the step below replaces it with the
          # build under review, and the pin is what an ordinary run tests.
          if [ -n "${MCPP_SOURCE_REF:-}" ]; then
            xlings install mcpp -y -g
          else
            xlings install "mcpp@$MCPP_VERSION" -y -g
          fi; then break; fi
            if [ "$attempt" = 6 ]; then
              echo "::error::mcpp@$MCPP_VERSION never appeared in the index (6 attempts over 5 minutes). If it was just released, the pointer has not propagated; if the pin names a version that was never published, it never will."
              exit 1
            fi
            echo "the index has not caught up yet (attempt $attempt of 6); waiting 60s"
            sleep 60
          done
          mcpp --version
          mcpp self config --mirror GLOBAL
          # ⭐⭐ CROSS-VALIDATION: BUILD THE mcpp UNDER REVIEW AND USE THAT ONE.
          #
          # Empty in the ordinary run, so this job keeps testing the RELEASED
          # mcpp the pin above names. Set it — `workflow_dispatch` input, or the
          # repository variable — and the same job runs against that source.
          #
          # ⚠️ THIS EXISTS BECAUSE THE ORDER USED TO BE WRONG. Several mcpp
          # releases went out green and only then turned this ecosystem red: the
          # engine's own CI cannot see a defect that appears only in a real
          # dependency graph, and this repository could not see the engine until
          # it had been published. Validating before the release closes that gap.
          #
          # The released mcpp installed just above is the bootstrap that compiles
          # it; mcpp builds itself and there is no other compiler for it here.
          if [ -n "${MCPP_SOURCE_REF:-}" ]; then
            src="$RUNNER_TEMP/mcpp-src"
            [ -d "$src" ] || git clone --quiet --depth 1 \
              --branch "$MCPP_SOURCE_REF" \
              https://github.com/mcpp-community/mcpp.git "$src"
            # ⚠️ THE CLONE'S OWN WORKSPACE PIN MUST NOT DECIDE WHICH mcpp
            # BUILDS IT. `.xlings.json` at mcpp's root pins the mcpp that
            # compiles mcpp, and that pin does not move when mcpp is released —
            # so a build inside the checkout obeys it and tries to install a
            # version the index may no longer carry:
            #
            #     [error] xlings: version '2026.8.17.1' not found for 'mcpp'
            #       available: 2026.8.25.1
            #
            # What is wanted here is the source compiled by the mcpp installed
            # above, which is exactly what removing the file leaves.
            rm -f "$src/.xlings.json"
            ( cd "$src" && mcpp build --release )
            # ⚠️ BOTH SPELLINGS, AND NO `-perm`. The matrix reaches Windows and
            # macOS runners too: on Windows the artefact is `mcpp.exe`, and
            # `-perm -u+x` is not a question that filesystem answers the way this
            # expects. Measured: `Finished release [optimized] in 173.44s`
            # followed by "mcpp did not build" — the build had succeeded and the
            # search was looking for the wrong name.
            #
            # `$src` is a FRESH clone each run, so `target/` holds exactly what
            # this step just built; `-printf` would be the safer form on a cached
            # tree and is a GNU extension this must not use.
            built=$(find "$src/target" -type f \
                    \( -name mcpp -o -name mcpp.exe \) | head -1)
            [ -n "$built" ] || { echo "::error::mcpp did not build from $MCPP_SOURCE_REF"; exit 1; }
            echo "$(cd "$(dirname "$built")" && pwd)" >> "$GITHUB_PATH"
            # ⚠️ Reported, because a PATH entry that does not win looks exactly
            # like one that does until something built with the wrong engine.
            echo "under review: $("$built" --version)  (from $MCPP_SOURCE_REF)"
          fi

      # The compiler family and version for this row. mcpp keeps its toolchains
      # in a sandbox of its own, so this selects rather than installs into the
      # system, and `mcpp test' and `mcpp run' have no flag for it --- which is
      # why it is set once here rather than passed to each command.
      - name: Select the toolchain
        run: |
          spec='${{ matrix.toolchain }}'
          case "$spec" in
            msvc*) mcpp toolchain default msvc ;;
            *)     mcpp toolchain install "${spec%@*}" "${spec#*@}"
                   mcpp toolchain default "$spec" ;;
          esac
          mcpp toolchain list

      # The C++ form. The modules are the artefact a C++ consumer imports, and
      # building them is what proves the compiler accepts them.
      # ⚠️ A 32-BIT TARGET, AND NOTHING ELSE IN THIS REPOSITORY IS ONE.
      #
      # Every row of every job here is hosted, and every hosted target this
      # specification serves is 64-bit. That is a structural blind spot rather
      # than an omission: the C ABI this repository freezes is instantiated at
      # two pointer widths, and only one of them was ever compiled.
      #
      # Measured 2026-08-20: `kal_node_info`'s frozen-layout assertion read
      # `offsetof(modified_ns) == sizeof(kal_uintptr)`, which is true at 64 bits
      # and false at 32 — the timestamp is a naturally-aligned `kal_u64`, so a
      # 32-bit `size` is followed by four bytes of padding and the offset stays
      # at eight. The failure surfaced in a CONSUMER (`mcpplibs/riscv-virt-rt`,
      # whose rv32 leg activates the openkal feature), which is the only place
      # the 32-bit layout existed at all.
      #
      # Compile-only: there is no 32-bit implementation to run against, and the
      # claim under test is about layout rather than behaviour.
      # ⚠️ A 32-BIT TARGET, AND NOTHING ELSE IN THIS REPOSITORY IS ONE.
      #
      # Every other row here is hosted, and every hosted target this
      # specification serves is 64-bit. That is a structural blind spot: the C
      # ABI this repository freezes is instantiated at two pointer widths, and
      # only one of them was ever compiled.
      #
      # Measured 2026-08-20: `kal_node_info`'s frozen-layout assertion read
      # `offsetof(modified_ns) == sizeof(kal_uintptr)`, true at 64 bits and
      # false at 32 — the timestamp is a naturally-aligned `kal_u64`, so a
      # 32-bit `size` is followed by four bytes of padding and the offset stays
      # at eight. It surfaced in a CONSUMER (`mcpplibs/riscv-virt-rt`'s rv32
      # leg), the only place the 32-bit layout existed at all.
      #
      # ⚠️ THROUGH `mcpp`, NOT THROUGH A BORROWED clang. The first version of
      # this step globbed the LLVM payload for `clang++` — which is absent on
      # the gcc and msvc rows, and is named `clang++.exe` on Windows, so the
      # command expanded to nothing and the step died with `: command not
      # found` on four rows. mcpp resolves the toolchain for a target itself;
      # asking it is the only spelling that holds on every row.
      # ⚠️ A 32-BIT TARGET, AND NOTHING ELSE IN THIS REPOSITORY IS ONE.
      #
      # Every other row here is hosted, and every hosted target this
      # specification serves is 64-bit. That is a structural blind spot: the C
      # ABI this repository freezes is instantiated at two pointer widths, and
      # only one of them was ever compiled.
      #
      # Measured 2026-08-20: `kal_node_info`'s frozen-layout assertion read
      # `offsetof(modified_ns) == sizeof(kal_uintptr)`, true at 64 bits and
      # false at 32 — the timestamp is a naturally-aligned `kal_u64`, so a
      # 32-bit `size` is followed by four bytes of padding and the offset stays
      # at eight. It surfaced in a CONSUMER (`mcpplibs/riscv-virt-rt`'s rv32
      # leg), the only place the 32-bit layout existed at all.
      #
      # ⚠️ THE PROBE LIVES INSIDE THE CHECKOUT AND USES A RELATIVE PATH. Two
      # earlier spellings failed for reasons that had nothing to do with the
      # ABI: globbing the LLVM payload for `clang++` died with `: command not
      # found` on the rows whose toolchain is gcc or msvc, and interpolating
      # `$GITHUB_WORKSPACE` into a C++ string literal produced
      # `"D:\a\openkal\openkal/include"`, whose backslashes the compiler read
      # as escape sequences. A relative directory has neither problem, and
      # `include_dir` resolves it against the package root.
      - name: The C ABI holds at thirty-two bits as well as sixty-four
        run: |
          set -euo pipefail
          rm -rf .abi-probe && mkdir -p .abi-probe/src
          cat > .abi-probe/mcpp.toml <<'EOF'
          [package]
          name    = "abi"
          version = "0.1.0"
          EOF
          cat > .abi-probe/build.mcpp <<'EOF'
          import mcpp;
          int main() { mcpp::include_dir("../include"); return 0; }
          EOF
          cat > .abi-probe/src/main.cpp <<'EOF'
          #include <openkal/types.h>
          #include <openkal/fs.h>
          static_assert(__builtin_offsetof(kal_node_info, size) == 0);
          static_assert(__builtin_offsetof(kal_node_info, modified_ns) == 8);
          static_assert(sizeof(kal_node_info) == 24);
          static_assert(sizeof(kal_dir)  == sizeof(kal_uintptr));
          static_assert(sizeof(kal_file) == sizeof(kal_uintptr));
          extern "C" void _start() { for (;;) {} }
          EOF
          for t in riscv64-none-elf riscv32-none-elf; do
            ( cd .abi-probe && rm -rf target && mcpp build --target "$t" ) \
              || { echo "the frozen layout does not hold for $t"; exit 1; }
            echo "$t: layout holds"
          done
          rm -rf .abi-probe

      - name: The module form compiles
        run: mcpp build

      # The C form, with the environment's own headers excluded --- because the
      # consumer this form exists for, a C library being ported onto openkal, is
      # compiled that way. The tool needs a driver it can pass -nostdinc to; the
      # toolchain that has no such spelling compiles the same declarations in
      # the translation unit the conformance suite carries, which every row of
      # the conformance job below builds.
      - name: The C form compiles without the environment's headers
        if: runner.os != 'Windows'
        run: |
          CC=cc bash tools/check-declarations.sh
          command -v clang >/dev/null && CC=clang bash tools/check-declarations.sh || true

      # ⭐⭐ AND NO DECLARATION NAMES A BACKEND TYPE.
      #
      # The job above asserts the declarations COMPILE. This asserts what they
      # are written WITH, which is a different freedom and the one that decides
      # whether a `kal_*` signature means the same thing on four object formats.
      # `sizeof(long)` is 8 on x86_64 Linux and 4 on x86_64 Windows; a `long` in
      # an interface would be one signature with two meanings.
      #
      # ⚠️ It runs on every row of this matrix rather than once, because a type
      # is a claim about a TARGET and this matrix is where the targets are.
      #
      # ⚠️ Needs a clang-family driver (`-Xclang -ast-print`); rows whose
      # toolchain has no such spelling are covered by the other rows, and the
      # step says so rather than passing silently.
      - name: No declaration names a backend type
        if: runner.os != 'Windows'
        run: |
          cc="$(command -v clang || true)"
          if [ -z "$cc" ]; then
            echo "no clang on this row; the type check runs on the rows that have one"
            exit 0
          fi
          bash tools/check-types.sh "$cc" include

      # ⚠️ AND THE CHECK DETECTS WHAT IT EXISTS TO DETECT.
      #
      # A check that succeeds by finding nothing succeeds identically when it
      # reads nothing, and every false green in this ecosystem has had that
      # shape. `check-types.sh` proves it parsed the headers before reporting;
      # this proves it would have failed.
      - name: The type check detects a leak
        if: runner.os != 'Windows'
        run: |
          cc="$(command -v clang || true)"
          [ -n "$cc" ] || exit 0
          cp include/openkal/env.h /tmp/env.h.orig
          printf '\nint kal_env_probe_leak(unsigned long n);\n' >> include/openkal/env.h
          if bash tools/check-types.sh "$cc" include; then
            cp /tmp/env.h.orig include/openkal/env.h
            echo "the type check accepted 'unsigned long' in an interface" >&2
            exit 1
          fi
          cp /tmp/env.h.orig include/openkal/env.h
          echo "the type check rejects a backend type, as it must"

  # ---------------------------------------------------------------------------
  # The property the specification exists for.
  substitution:
    name: substitution holds (${{ matrix.toolchain }})
    runs-on: ubuntu-24.04
    timeout-minutes: 40
    strategy:
      fail-fast: false
      matrix:
        toolchain: ['gcc@16.1.0', 'llvm@22.1.8']
    steps:
      - uses: actions/checkout@v4

      - name: Install xlings
        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: |
          # ⚠️ A LOOP, BECAUSE ONE `xlings update` CAN RETURN A STALE INDEX
          # WITHOUT SAYING SO.
          #
          # The index is published as an artifact behind a pointer, and that
          # pointer propagates asynchronously after a version bump is merged.
          # Measured on release day: an update run four minutes after the merge
          # printed `index updated`, and the install then failed with
          #
          #     package 'mcpp@<ver>' not found in the synced index
          #     (xim@artifact:<an older sha>, ...), synced 0 seconds ago
          #
          # Nothing had gone wrong. The update fetched the PREVIOUS artifact,
          # and "synced 0 seconds ago" describes when it was fetched rather than
          # what it contains — which is why the message reads as freshness.
          #
          # So this is not a retry around flakiness; it is the wait that a
          # single update does not perform. A pin naming a version that was
          # never published still fails, after the last attempt, and says which
          # of the two situations it is.
          for attempt in 1 2 3 4 5 6; do
            xlings update > /dev/null 2>&1 || true
            if # ⚠️ THE PIN MAY NAME THE RELEASE THIS RUN IS VALIDATING, which does
          # not exist yet — that is the whole point of MCPP_SOURCE_REF. Bootstrap
          # from whatever the index has; the step below replaces it with the
          # build under review, and the pin is what an ordinary run tests.
          if [ -n "${MCPP_SOURCE_REF:-}" ]; then
            xlings install mcpp -y -g
          else
            xlings install "mcpp@$MCPP_VERSION" -y -g
          fi; then break; fi
            if [ "$attempt" = 6 ]; then
              echo "::error::mcpp@$MCPP_VERSION never appeared in the index (6 attempts over 5 minutes). If it was just released, the pointer has not propagated; if the pin names a version that was never published, it never will."
              exit 1
            fi
            echo "the index has not caught up yet (attempt $attempt of 6); waiting 60s"
            sleep 60
          done
          mcpp self config --mirror GLOBAL
          # ⭐⭐ CROSS-VALIDATION: BUILD THE mcpp UNDER REVIEW AND USE THAT ONE.
          #
          # Empty in the ordinary run, so this job keeps testing the RELEASED
          # mcpp the pin above names. Set it — `workflow_dispatch` input, or the
          # repository variable — and the same job runs against that source.
          #
          # ⚠️ THIS EXISTS BECAUSE THE ORDER USED TO BE WRONG. Several mcpp
          # releases went out green and only then turned this ecosystem red: the
          # engine's own CI cannot see a defect that appears only in a real
          # dependency graph, and this repository could not see the engine until
          # it had been published. Validating before the release closes that gap.
          #
          # The released mcpp installed just above is the bootstrap that compiles
          # it; mcpp builds itself and there is no other compiler for it here.
          if [ -n "${MCPP_SOURCE_REF:-}" ]; then
            src="$RUNNER_TEMP/mcpp-src"
            [ -d "$src" ] || git clone --quiet --depth 1 \
              --branch "$MCPP_SOURCE_REF" \
              https://github.com/mcpp-community/mcpp.git "$src"
            # ⚠️ THE CLONE'S OWN WORKSPACE PIN MUST NOT DECIDE WHICH mcpp
            # BUILDS IT. `.xlings.json` at mcpp's root pins the mcpp that
            # compiles mcpp, and that pin does not move when mcpp is released —
            # so a build inside the checkout obeys it and tries to install a
            # version the index may no longer carry:
            #
            #     [error] xlings: version '2026.8.17.1' not found for 'mcpp'
            #       available: 2026.8.25.1
            #
            # What is wanted here is the source compiled by the mcpp installed
            # above, which is exactly what removing the file leaves.
            rm -f "$src/.xlings.json"
            ( cd "$src" && mcpp build --release )
            # ⚠️ BOTH SPELLINGS, AND NO `-perm`. The matrix reaches Windows and
            # macOS runners too: on Windows the artefact is `mcpp.exe`, and
            # `-perm -u+x` is not a question that filesystem answers the way this
            # expects. Measured: `Finished release [optimized] in 173.44s`
            # followed by "mcpp did not build" — the build had succeeded and the
            # search was looking for the wrong name.
            #
            # `$src` is a FRESH clone each run, so `target/` holds exactly what
            # this step just built; `-printf` would be the safer form on a cached
            # tree and is a GNU extension this must not use.
            built=$(find "$src/target" -type f \
                    \( -name mcpp -o -name mcpp.exe \) | head -1)
            [ -n "$built" ] || { echo "::error::mcpp did not build from $MCPP_SOURCE_REF"; exit 1; }
            echo "$(cd "$(dirname "$built")" && pwd)" >> "$GITHUB_PATH"
            # ⚠️ Reported, because a PATH entry that does not win looks exactly
            # like one that does until something built with the wrong engine.
            echo "under review: $("$built" --version)  (from $MCPP_SOURCE_REF)"
          fi

      # The compiler family and version for this row. mcpp keeps its toolchains
      # in a sandbox of its own, so this selects rather than installs into the
      # system, and `mcpp test' and `mcpp run' have no flag for it --- which is
      # why it is set once here rather than passed to each command.
      - name: Select the toolchain
        run: |
          spec='${{ matrix.toolchain }}'
          case "$spec" in
            msvc*) mcpp toolchain default msvc ;;
            *)     mcpp toolchain install "${spec%@*}" "${spec#*@}"
                   mcpp toolchain default "$spec" ;;
          esac
          mcpp toolchain list

      - name: Substitution holds
        working-directory: examples/substitution/app
        run: |
          # The property under test is that the source is invariant. The
          # checksum is taken before the first build and compared after the
          # second, so that a change made by either build would be detected.
          before="$(sha256sum src/main.cpp | cut -d' ' -f1)"

          mcpp build > /dev/null
          ./target/*/*/bin/app > with-fd.log 2>&1
          grep -q 'the application produced this line' with-fd.log

          sed -i 's|openkal-fd = { path = "../impl-fd" }|openkal-discard = { path = "../impl-discard" }|' mcpp.toml
          rm -rf target
          mcpp build > /dev/null
          ./target/*/*/bin/app > with-discard.log 2>&1
          if grep -q 'the application produced this line' with-discard.log; then
            echo "the discarding implementation produced output"; exit 1
          fi

          after="$(sha256sum src/main.cpp | cut -d' ' -f1)"
          [ "$before" = "$after" ] || { echo "the source changed between builds"; exit 1; }

      - name: The surface checker detects an addition
        working-directory: examples/substitution/impl-fd
        run: |
          # The checker is only useful if it fails when it should. An
          # unspecified name is introduced, the checker is required to reject
          # it, and the name is then removed.
          mcpp build > /dev/null
          bash ../../../tools/check-surface.sh ../../../SURFACE.txt \
            $(find target -name '*.o' | tr '\n' ' ')

          printf 'extern "C" void kal_vendor_extension(void) {}\n' > src/extra.cpp
          mcpp build > /dev/null
          if bash ../../../tools/check-surface.sh ../../../SURFACE.txt \
               $(find target -name '*.o' | tr '\n' ' '); then
            echo "the checker accepted an unspecified name"; exit 1
          fi
          rm -f src/extra.cpp

  # ---------------------------------------------------------------------------
  # The suite, against the implementation for each system.
  #
  # The implementations are checked out at the branch under test where they have
  # one and at their default branch otherwise, so that this job asserts what it
  # is for: that the specification as written here and the implementations as
  # written there agree today.
  conformance:
    name: conformance (${{ matrix.implementation }}, ${{ matrix.toolchain }})
    runs-on: ${{ matrix.os }}
    timeout-minutes: 60
    strategy:
      fail-fast: false
      matrix:
        include:
          # `features' names the optional interfaces the implementation in that
          # row provides, on top of `full'. It is per-row rather than global
          # because clause 6.1 makes a missing optional interface not a
          # deviation: a set that demanded every one of them from every
          # implementation would report a link failure where the specification
          # says there is nothing to report.
          - { os: ubuntu-24.04, toolchain: 'gcc@16.1.0',  implementation: openkal-linux,   features: 'full,optional' }
          - { os: ubuntu-24.04, toolchain: 'llvm@22.1.8', implementation: openkal-linux,   features: 'full,optional' }
          - { os: macos-14,     toolchain: 'llvm@20.1.7', implementation: openkal-macos,   features: 'full' }
          - { os: windows-2022, toolchain: 'llvm@20.1.7', implementation: openkal-windows, features: 'full' }
          - { os: windows-2022, toolchain: 'msvc@system', implementation: openkal-windows, features: 'full' }
    defaults:
      run:
        shell: bash
    steps:
      - uses: actions/checkout@v4

      - name: The implementation
        run: |
          git clone --quiet https://github.com/mcpplibs/${{ matrix.implementation }}.git .impl
          branch='${{ github.head_ref || github.ref_name }}'
          if git -C .impl rev-parse --verify --quiet "origin/$branch" > /dev/null; then
            git -C .impl checkout --quiet "origin/$branch"
            echo "the implementation is at $branch $(git -C .impl rev-parse --short HEAD)"
          else
            echo "the implementation has no $branch; its default branch is used" \
                 "($(git -C .impl rev-parse --short HEAD))"
          fi

          # Whether the two are describing the same version of the contract.
          #
          # This job asserts that the specification as written here and the
          # implementation as written there agree today, and that assertion is
          # only meaningful when the implementation is written against this
          # version. During a release across five repositories they are out of
          # step for the seconds between one merge and the next, and the
          # compiler's report on that is a hundred errors about names a header
          # no longer has --- which names the symptom and not the cause.
          # An implementation states which openkal it is written against in one
          # of two forms, and the question the check asks is different in each.
          #
          #   openkal = "0.5.2"            a released version. The question is
          #                                whether it is this one.
          #   openkal = { git = …, branch = … }
          #                                a development line. The version is
          #                                not stated and cannot be; what makes
          #                                the two in step is that the branch is
          #                                this branch, and that is what to ask.
          #
          # The second form was not anticipated when this check was written, and
          # the check then extracted nothing and reported a mismatch against the
          # empty string --- which named neither the cause nor the symptom.
          here="$(sed -n 's/^version[[:space:]]*=[[:space:]]*"\([^"]*\)".*/\1/p' mcpp.toml | head -1)"
          there="$(sed -n 's/^openkal[[:space:]]*=[[:space:]]*"\([^"]*\)".*/\1/p' .impl/mcpp.toml | head -1)"
          if [ -n "$there" ]; then
            if [ "$here" != "$there" ]; then
              echo "::error::this is openkal $here and ${{ matrix.implementation }} is written against openkal $there. Nothing is wrong with either; they are not in step. Re-run after the implementation's branch has the matching version." >&2
              exit 1
            fi
            echo "both describe openkal $here"
          else
            impl_branch="$(sed -n 's/^openkal[[:space:]]*=[[:space:]]*{.*branch[[:space:]]*=[[:space:]]*"\([^"]*\)".*/\1/p' .impl/mcpp.toml | head -1)"
            if [ -z "$impl_branch" ]; then
              echo "::error::${{ matrix.implementation }} states neither a version nor a branch for openkal, so there is nothing to be in step with." >&2
              exit 1
            fi
            if [ "$impl_branch" != "$branch" ]; then
              echo "::error::this is $branch and ${{ matrix.implementation }} follows $impl_branch. They are not in step." >&2
              exit 1
            fi
            echo "both follow $branch; this is openkal $here"
          fi

      - name: Install xlings and mcpp (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 and mcpp (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: |
          # ⚠️ A LOOP, BECAUSE ONE `xlings update` CAN RETURN A STALE INDEX
          # WITHOUT SAYING SO.
          #
          # The index is published as an artifact behind a pointer, and that
          # pointer propagates asynchronously after a version bump is merged.
          # Measured on release day: an update run four minutes after the merge
          # printed `index updated`, and the install then failed with
          #
          #     package 'mcpp@<ver>' not found in the synced index
          #     (xim@artifact:<an older sha>, ...), synced 0 seconds ago
          #
          # Nothing had gone wrong. The update fetched the PREVIOUS artifact,
          # and "synced 0 seconds ago" describes when it was fetched rather than
          # what it contains — which is why the message reads as freshness.
          #
          # So this is not a retry around flakiness; it is the wait that a
          # single update does not perform. A pin naming a version that was
          # never published still fails, after the last attempt, and says which
          # of the two situations it is.
          for attempt in 1 2 3 4 5 6; do
            xlings update > /dev/null 2>&1 || true
            if # ⚠️ THE PIN MAY NAME THE RELEASE THIS RUN IS VALIDATING, which does
          # not exist yet — that is the whole point of MCPP_SOURCE_REF. Bootstrap
          # from whatever the index has; the step below replaces it with the
          # build under review, and the pin is what an ordinary run tests.
          if [ -n "${MCPP_SOURCE_REF:-}" ]; then
            xlings install mcpp -y -g
          else
            xlings install "mcpp@$MCPP_VERSION" -y -g
          fi; then break; fi
            if [ "$attempt" = 6 ]; then
              echo "::error::mcpp@$MCPP_VERSION never appeared in the index (6 attempts over 5 minutes). If it was just released, the pointer has not propagated; if the pin names a version that was never published, it never will."
              exit 1
            fi
            echo "the index has not caught up yet (attempt $attempt of 6); waiting 60s"
            sleep 60
          done
          mcpp self config --mirror GLOBAL
          # ⭐⭐ CROSS-VALIDATION: BUILD THE mcpp UNDER REVIEW AND USE THAT ONE.
          #
          # Empty in the ordinary run, so this job keeps testing the RELEASED
          # mcpp the pin above names. Set it — `workflow_dispatch` input, or the
          # repository variable — and the same job runs against that source.
          #
          # ⚠️ THIS EXISTS BECAUSE THE ORDER USED TO BE WRONG. Several mcpp
          # releases went out green and only then turned this ecosystem red: the
          # engine's own CI cannot see a defect that appears only in a real
          # dependency graph, and this repository could not see the engine until
          # it had been published. Validating before the release closes that gap.
          #
          # The released mcpp installed just above is the bootstrap that compiles
          # it; mcpp builds itself and there is no other compiler for it here.
          if [ -n "${MCPP_SOURCE_REF:-}" ]; then
            src="$RUNNER_TEMP/mcpp-src"
            [ -d "$src" ] || git clone --quiet --depth 1 \
              --branch "$MCPP_SOURCE_REF" \
              https://github.com/mcpp-community/mcpp.git "$src"
            # ⚠️ THE CLONE'S OWN WORKSPACE PIN MUST NOT DECIDE WHICH mcpp
            # BUILDS IT. `.xlings.json` at mcpp's root pins the mcpp that
            # compiles mcpp, and that pin does not move when mcpp is released —
            # so a build inside the checkout obeys it and tries to install a
            # version the index may no longer carry:
            #
            #     [error] xlings: version '2026.8.17.1' not found for 'mcpp'
            #       available: 2026.8.25.1
            #
            # What is wanted here is the source compiled by the mcpp installed
            # above, which is exactly what removing the file leaves.
            rm -f "$src/.xlings.json"
            ( cd "$src" && mcpp build --release )
            # ⚠️ BOTH SPELLINGS, AND NO `-perm`. The matrix reaches Windows and
            # macOS runners too: on Windows the artefact is `mcpp.exe`, and
            # `-perm -u+x` is not a question that filesystem answers the way this
            # expects. Measured: `Finished release [optimized] in 173.44s`
            # followed by "mcpp did not build" — the build had succeeded and the
            # search was looking for the wrong name.
            #
            # `$src` is a FRESH clone each run, so `target/` holds exactly what
            # this step just built; `-printf` would be the safer form on a cached
            # tree and is a GNU extension this must not use.
            built=$(find "$src/target" -type f \
                    \( -name mcpp -o -name mcpp.exe \) | head -1)
            [ -n "$built" ] || { echo "::error::mcpp did not build from $MCPP_SOURCE_REF"; exit 1; }
            echo "$(cd "$(dirname "$built")" && pwd)" >> "$GITHUB_PATH"
            # ⚠️ Reported, because a PATH entry that does not win looks exactly
            # like one that does until something built with the wrong engine.
            echo "under review: $("$built" --version)  (from $MCPP_SOURCE_REF)"
          fi

      # The compiler family and version for this row. mcpp keeps its toolchains
      # in a sandbox of its own, so this selects rather than installs into the
      # system, and `mcpp test' and `mcpp run' have no flag for it --- which is
      # why it is set once here rather than passed to each command.
      - name: Select the toolchain
        run: |
          spec='${{ matrix.toolchain }}'
          case "$spec" in
            msvc*) mcpp toolchain default msvc ;;
            *)     mcpp toolchain install "${spec%@*}" "${spec#*@}"
                   mcpp toolchain default "$spec" ;;
          esac
          mcpp toolchain list

      # Every interface and every kind of examination. The exit status is the
      # verdict: 0 when every observation held, 1 when one did not, and 2 when
      # nothing was observed --- the last being the outcome a run that selected
      # no interface would otherwise pass silently.
      - name: Every interface, every kind of examination
        run: |
          bash tools/run-conformance.sh '${{ matrix.implementation }}' .impl '${{ matrix.features }}'

      # The suite is composable because openkal is: an implementation provides
      # an interface in whole or not at all, and a suite that examined all eight
      # unconditionally would fail to link against a conforming implementation
      # of three. Selecting three is therefore asserted to produce a report
      # rather than a link failure.
      # openkal-kit, which is in this repository and is not this specification.
      #
      # The kit composes facilities from openkal's interfaces --- awaiting a
      # channel, writing an endpoint down and reading one back --- that the
      # specification declines to admit because they can be composed. Running its
      # tests here is what keeps the composition working as the interfaces it
      # rests upon change.
      #
      # THE SURFACE CHECK ABOVE IS WHAT KEEPS THE TWO APART, and it does so by
      # the shape of what is exported rather than by a declaration. The kit is
      # C++ modules in `namespace kal::kit` and exports no name beginning with
      # `kal_`; a C++ name is mangled to `_ZN3kal3kit…`, so an implementation
      # that linked the kit still passes `check-surface.sh --complete`.
      - name: The kit, composed from the interfaces above
        if: matrix.implementation == 'openkal-linux'
        run: |
          bash tools/run-kit-tests.sh '${{ matrix.implementation }}' .impl

      # ⭐⭐ THE EXAMPLE A CONSUMER READS, BUILT AND RUN.
      #
      # examples/portable is the program the specification points at as the one
      # every implementation must satisfy, and NOTHING BUILT IT. Its README
      # stated that "an implementation's continuous integration fetches this
      # source ... builds it against itself, and asserts the lines below";
      # measured across all six implementation repositories, the word `portable`
      # appeared in none of their workflows. A promise with nothing executing it.
      #
      # ⚠️ AND THE DRIFT THAT FOLLOWS FROM THAT IS VISIBLE IN ITS OWN MANIFEST:
      # it named `openkal = "0.5.0"` --- three specification versions behind ---
      # while claiming to demonstrate the current contract to a reader who would
      # copy it. Nothing could have noticed.
      #
      # The manifest names published versions, because that is what a consumer
      # writes. Those do not exist while a change spanning these repositories is
      # under review, so the working trees stand in, exactly as
      # tools/run-conformance.sh does for the suite.
      #
      # ⚠️ RELATIVE PATHS, NOT `pwd`. This job runs on Windows too, where MSYS
      # reports `/d/a/...`, which the build tool does not resolve. Relative
      # paths also cannot put a local directory into a file that might be
      # committed.
      #
      # ⚠️ NO `standalone` FEATURE HERE. That feature makes the implementation
      # supply `_start`, which collides with the C library's crt1.o on a hosted
      # toolchain --- measured: "multiple definition of `_start`". The
      # conformance runner defaults to no features for the same reason.
      - name: The portable program builds and runs over this implementation
        working-directory: examples/portable
        run: |
          set -euo pipefail
          subst() { sed -E "$2" "$1" > "$1.next" && mv "$1.next" "$1"; }
          subst mcpp.toml 's|^openkal = .*$|openkal = { path = "../.." }|'
          subst mcpp.toml 's|^${{ matrix.implementation }} = .*$|${{ matrix.implementation }} = { path = "../../.impl" }|'
          grep -q 'path = "../../.impl"' mcpp.toml \
            || { echo "::error::the implementation was not substituted"; exit 1; }

          # ⚠️ THE IMPLEMENTATION REACHES THE SPECIFICATION TOO, AND BY ITS OWN
          # FORM. Substituting only this manifest leaves two routes to one
          # package and the build refuses, correctly:
          #
          #   error: dependency 'mcpplibs.openkal' is requested as both a path
          #          dep (by 'portable') and a git dep (by 'openkal-linux@path')
          #
          # Both must name the same directory, which is what run-conformance.sh
          # does for the suite. From `.impl/` the specification is `..`; the two
          # spellings resolve to one canonical path and therefore to one node.
          subst ../../.impl/mcpp.toml 's|^openkal = .*$|openkal = { path = ".." }|'
          grep -q 'openkal = { path = ".." }' ../../.impl/mcpp.toml \
            || { echo "::error::the implementation still reaches openkal its own way"; exit 1; }

          mcpp run 2>&1 | tee portable.log

          # ⭐ THE COUNT IS THE CRITERION, AND SO IS THE DENOMINATOR. A program
          # that failed to start would print none of these lines, and a check
          # for the absence of failures would hold over an empty log.
          lines=$(grep -c '^openkal: ' portable.log)
          [ "$lines" -ge 10 ] \
            || { echo "::error::the program printed $lines lines; it reports one per interface"; exit 1; }
          grep -q 'openkal: observations that did not hold: 0' portable.log \
            || { echo "::error::the portable program reported observations that did not hold"; exit 1; }
          echo "  the portable program held every observation over ${{ matrix.implementation }}"

      - name: A selection of three interfaces is examined, not refused
        run: |
          rm -rf conformance/target
          bash tools/run-conformance.sh '${{ matrix.implementation }}' .impl core,fs,task \
            | tee selected.log

          # A report was produced, and nothing in it failed.
          grep -qE 'observations: [0-9]+ held, 0 did not hold' selected.log

          # An interface that was not selected is reported as not examined and
          # carries the reason, rather than being absent --- a report that
          # omitted it could not be distinguished from a report on an
          # implementation that provides it.
          grep -q 'openkal.process --- the interface was not selected' selected.log

          # And an interface that was selected was examined.
          grep -qE 'held +\[behaviour\].*' selected.log