ci.yml

# ─────────────────────────────────────────────────────────────────────────────
# CI — the gate every change to main passes through.
#
# Four tiers, cheapest first, all required:
#   build     the parser and the engine driver compile and typecheck
#   hygiene   repo invariants that need no solver and catch silent breakage
#   engine    the java / typescript / python / javascript / csharp regression suites, in
#             parallel, each compiling its own engine from the rules
#   engines   every language's engine builds on every platform (the reusable
#             build-engines workflow — the same artifacts publish-npm ships)
#
# NO WORKFLOW-LEVEL PATH FILTERS, deliberately. A required check that is skipped
# by a path filter never reports, and a pull request waiting on a check that will
# never report can never merge. Instead the workflow always starts, the `changes`
# job classifies the diff, and the expensive jobs skip THEMSELVES: a docs-only pull
# request runs build and hygiene (which carries the version gate) and nothing else,
# and the platform engine build runs only for main, and only when what it compiles
# changed. The `CI` job reports either way.
#
# dev is the default branch: every pull request lands there and gets build, hygiene
# and the five suites. The every-platform engine build is the slow part and no test
# uses its output, so it runs on the way INTO main (a promotion pull request, a push
# to main) and in the nightly, not on every change to dev.
# ─────────────────────────────────────────────────────────────────────────────
name: CI

on:
  push:
    # dev takes direct pushes, so they get a result too. A release branch (0.1.6, ...)
    # takes merges from pull requests: the run on the merge is what saves the compiled
    # engines where the NEXT pull request into it can restore them — a cache a pull
    # request saves is visible to that pull request alone.
    branches: [main, dev, '0.*']
  pull_request:
  merge_group:
  workflow_dispatch:
  # nightly.yml calls this with fresh=true: no restored engine cache, so every engine
  # is compiled from the rules as they are, and the platform build always runs.
  workflow_call:
    inputs:
      fresh:
        type: boolean
        default: false

# One run per ref. A new push to a pull request cancels the previous run, but a
# run on main is always allowed to finish — main's history is the record.
concurrency:
  group: ci-${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

permissions:
  contents: read

env:
  # bin/axiomcode and the parser need Node ≥ 22.5.
  NODE_VERSION: '22'
  SOUFFLE_VERSION: '2.5'
  SOUFFLE_SHA512: '6b86e554f6aa5abf8a8b55d8312ae37c0957c5bd6c9edeea89246db9406f645ec5e600b84fe6636b1c163da556f0da6c3d2dad46c1083413f2fcf4f95b9ac62c'

jobs:
  changes:
    name: what changed
    runs-on: ubuntu-24.04
    timeout-minutes: 5
    outputs:
      code: ${{ steps.c.outputs.code }}
      engines: ${{ steps.c.outputs.engines }}
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - id: c
        env:
          BASE: ${{ github.event.pull_request.base.sha }}
        run: |
          set -euo pipefail
          if [ "${{ inputs.fresh }}" = true ]; then
            echo "code=true" >> "$GITHUB_OUTPUT"; echo "engines=true" >> "$GITHUB_OUTPUT"
            echo "nightly: everything runs"; exit 0
          fi
          if [ "${{ github.event_name }}" = push ] && [ "${{ github.ref }}" = refs/heads/dev ]; then
            echo "code=true" >> "$GITHUB_OUTPUT"; echo "engines=false" >> "$GITHUB_OUTPUT"
            echo "push to dev: suites run, platform engines wait for main"; exit 0
          fi
          if [ "${{ github.event_name }}" = push ] && [[ "${{ github.ref }}" == refs/heads/0.* ]]; then
            echo "code=true" >> "$GITHUB_OUTPUT"; echo "engines=false" >> "$GITHUB_OUTPUT"
            echo "push to a release branch: suites run (and save its engines for pull requests into it), platform engines wait for main"; exit 0
          fi
          # A push to main is a release when its version has no tag yet: that commit, and only that one,
          # builds every platform and runs the five-platform e2e, and release.yml drafts from it once
          # it is green. A push whose version is already tagged released nothing new and builds nothing.
          if [ "${{ github.event_name }}" = push ] && [ "${{ github.ref }}" = refs/heads/main ]; then
            v="$(node .github/scripts/version.mjs get)"
            if git ls-remote --exit-code --tags origin "refs/tags/v$v" >/dev/null; then
              echo "code=true" >> "$GITHUB_OUTPUT"; echo "engines=false" >> "$GITHUB_OUTPUT"
              echo "push to main at v$v, already tagged: suites run, nothing to release"
            else
              echo "code=true" >> "$GITHUB_OUTPUT"; echo "engines=true" >> "$GITHUB_OUTPUT"
              echo "push to main at v$v, not yet tagged: the release build runs on every platform"
            fi
            exit 0
          fi
          if [ "${{ github.event_name }}" != pull_request ]; then
            echo "code=true" >> "$GITHUB_OUTPUT"; echo "engines=true" >> "$GITHUB_OUTPUT"
            echo "${{ github.event_name }} on ${{ github.ref }}: everything runs"; exit 0
          fi
          files="$(git diff --name-only "$BASE"...HEAD)"
          printf '%s\n' "$files" | sed 's/^/  /'
          # DOCS: prose nothing executes. Markdown under graph/ or parser/ is NOT
          # docs: graph/bundle/SCHEMA.md is generated, and a suite checks it is current.
          code="$(printf '%s\n' "$files" | grep -vE \
            -e '^$' \
            -e '^(graph|parser)/' -e '^[^/]+\.md$' -e '^docs/' -e '^paper/' \
            -e '^\.github/(ISSUE_TEMPLATE/|pull_request_template\.md$|CODEOWNERS$|RELEASING\.md$)' \
            -e '^LICENSE' -e '\.(png|jpe?g|gif|svg)$' || true)"
          # graph/ and parser/ are code even when the file is markdown
          code="$code$(printf '%s\n' "$files" | grep -E '^(graph|parser)/' || true)"
          # ENGINES: what the platform build compiles or packages.
          engines="$(printf '%s\n' "$files" | grep -E \
            -e '\.dl$' -e '^graph/pipeline/' -e '^packaging/' -e 'scripts/dl_program\.py$' -e '^\.github/scripts/query-smoke\.sh$' \
            -e '^\.github/workflows/build-engines\.yml$' -e '^package\.json$' -e '^\.github/scripts/e2e-' -e '^\.github/e2e/' || true)"
          [ -n "$code" ] && echo "code=true" >> "$GITHUB_OUTPUT" || echo "code=false" >> "$GITHUB_OUTPUT"
          # No pull request builds the platform engines: a release builds them once, on the push that
          # lands it on main, and publishes exactly that build (#1350).
          engines=""
          [ -n "$engines" ] && echo "engines=true" >> "$GITHUB_OUTPUT" || echo "engines=false" >> "$GITHUB_OUTPUT"
          echo "suites: $([ -n "$code" ] && echo run || echo skip)   platform engines: $([ -n "$engines" ] && echo run || echo skip)"

  build:
    name: build & typecheck
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: ${{ env.NODE_VERSION }}
      # No lock file is committed (#476), so this is `npm install`, not `npm ci`,
      # and setup-node's npm cache (keyed on a lock file) is not used. The install
      # runs the `prepare` script, which builds the parser workspace and the
      # driver. Keeping the explicit build step anyway means a prepare-script
      # change cannot silently stop compiling this repo.
      - run: npm install
      - run: npm run typecheck
      - run: npm run build
      - name: the parser actually built
        run: |
          test -f parser/dist/index.js \
            || { echo "::error::parser/dist/index.js is missing after build"; exit 1; }

  hygiene:
    name: repo invariants
    runs-on: ubuntu-24.04
    timeout-minutes: 5
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0    # the version gate needs the merge base with the target branch

      # A fixture input that .gitignore matches passes on the machine that wrote
      # it and fails on every clone. The suites run this too; running it here as
      # well means the answer arrives in seconds rather than after the engine.
      - name: every fixture input is tracked by git
        run: bash graph/test/tools/no-ignored-fixtures.sh

      # A relation staged for the client but not for libraries is EMPTY on every
      # run and nothing errors — no golden can see it. This is the only check
      # that can.
      - name: IR staging maps are consistent
        run: |
          fail=0
          for lang in java typescript python javascript csharp; do
            echo "── $lang"
            python3 graph/test/tools/check_staging.py --lang "$lang" || fail=1
          done
          exit $fail

      # The engine's base declarations are a COPY of the parser's generated
      # schema (graph/<lang>/souffle/decls_base.dl ← parser/src/schema/<lang>/).
      # A column appended on the parser side and not here is an arity error at
      # solve time in every suite at once, with the cause two directories away.
      # Compared on the `.decl` lines only: the copies carry their own preambles.
      - name: engine declarations match the parser schema
        run: |
          fail=0
          for pair in typescript:decls_base_ts.dl python:decls_base_py.dl javascript:decls_base_js.dl csharp:decls_base_cs.dl; do
            lang="${pair%%:*}"; file="${pair#*:}"
            if ! diff <(grep '^\.decl' "parser/src/schema/$lang/$file") \
                      <(grep '^\.decl' "graph/$lang/souffle/decls_base.dl"); then
              echo "::error::graph/$lang/souffle/decls_base.dl has drifted from parser/src/schema/$lang/$file"
              fail=1
            fi
          done
          exit $fail

      # The client->library half is carried by a smaller set of fixtures than the
      # client->client half, and it is the half that disappears silently: delete a
      # golden and the case still runs, still passes, and simply stops claiming
      # anything. A suite can only check the assertions it still has.
      - name: client->library coverage has not shrunk
        run: bash graph/test/tools/lib-coverage.sh

      - name: shell scripts parse
        run: |
          fail=0
          while IFS= read -r f; do
            bash -n "$f" || { echo "::error file=$f::does not parse"; fail=1; }
          done < <(git ls-files '*.sh')
          exit $fail

      # npm will not republish a version, so anything inside the tarball reaches
      # nobody unless the version moves — README.md included, since `files` names
      # it. The inverse is the error worth avoiding too: a bump demanded for a CI
      # tweak or a test fixture teaches people to bump without asking why, and a
      # version that moves for reasons users cannot observe stops meaning
      # anything. The gate reads package.json's own `files` declaration to tell
      # the two apart, and prints the files that decided it.
      - uses: actions/setup-node@v4
        with:
          node-version: ${{ env.NODE_VERSION }}

      # package.json's version is repeated in the engine pins, the parser and every
      # plugin manifest. Checked on every event, not only pull requests, so main
      # can never hold a tree whose manifests disagree about what it is.
      - name: every manifest carries the same version
        run: node .github/scripts/version.mjs check

      - name: a change that reaches a user has a version
        if: github.event_name == 'pull_request'
        run: bash .github/scripts/version-gate.sh "origin/${{ github.base_ref }}"

      # A push to main builds and releases only when its version is new (#1350). The gate above already
      # refuses a pull request that changes what users get without a new version; a CI or docs change
      # may keep main's version and then builds nothing when it lands. What is left to refuse here is a
      # new version that was already released: its tag exists, so the push would build nothing and the
      # change would never ship.
      - name: a pull request into main does not reuse a released version
        if: github.event_name == 'pull_request' && github.base_ref == 'main'
        run: |
          set -euo pipefail
          head="$(node .github/scripts/version.mjs get)"
          base="$(git show origin/main:package.json | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>console.log(JSON.parse(s).version))')"
          if [ "$head" = "$base" ]; then
            echo "main stays at $base: nothing in this change is released, and landing it builds nothing"; exit 0
          fi
          if git ls-remote --exit-code --tags origin "refs/tags/v$head" >/dev/null; then
            echo "::error::v$head is already released — pick the next version"; exit 1
          fi
          echo "main $base -> $head: landing this builds every platform and drafts v$head"

  engine:
    name: engine (${{ matrix.lang }})
    needs: [changes]
    if: needs.changes.outputs.code == 'true'
    runs-on: ubuntu-24.04
    timeout-minutes: 45
    env:
      # the engine is compiled for any x86-64 runner, so a cached one can be restored
      # on whichever runner this job lands (see the engine cache step below)
      AXIOM_ENGINE_MARCH: portable
    strategy:
      fail-fast: false
      matrix:
        include:
          # --oracle scores the engine against GROUND TRUTH, not only against the
          # goldens. A golden says "the same as last time", which a wrong answer
          # satisfies perfectly well as long as it was wrong last time too.
          #
          # The ground truth is built with the toolchain that defines the language:
          # javac and javap for Java, the TypeScript compiler for TypeScript and —
          # over allowJs/checkJs — for JavaScript. No third-party analyzer, and no
          # third-party library is downloaded to do it. A case whose ground truth would need an external
          # classpath reports itself unscored rather than pulling one in.
          #
          # --no-torture for java ONLY. Those families call java.util.List, Map and
          # the functional interfaces, so they need the JVM platform IR staged as a
          # library. That IR is 1.8 GB and is built from a JDK source checkout, so it
          # cannot live in a repository or a cache. Without it those receivers are
          # unresolvable BY CONSTRUCTION — the census goes from 10 missing edges to
          # 23 and recall to 0.847 — which measures the staging, not the rules, and
          # the resulting red would read as a regression in whatever PR met it.
          #
          # Java client->library resolution is still covered here: six cases ship
          # their own stub library in lib-src/ and are solved with it as --library.
          # The torture families remain a local gate until the platform IR can be
          # produced reproducibly; the suite prints EXCLUDED so it is never mistaken
          # for a family that passed.
          - lang: java
            oracle: '--oracle --no-torture'
          - lang: typescript
            oracle: '--oracle'
          # Python's ground truth is frozen CPython output, authored by a separate
          # harness checkout ($AXIOM_PY_ORACLE) that CI cannot reach yet, so this leg
          # is goldens-only for now. That separation is deliberate —
          # graph/test/python/run-tests.sh explains why the ability to re-bless
          # ground truth must not sit beside the code under test — but it does mean
          # the python leg is a weaker check than the other three until the harness
          # is reachable from here.
          - lang: python
            oracle: ''
          # JavaScript: 19 cases; the library case ships its dependency under
          # src/node_modules and is solved twice. The execution oracles (torture/,
          # realapp/) are separate harnesses and stay a local gate — realapp needs
          # network for its own npm install.
          - lang: javascript
            oracle: '--oracle'
          # C# has no goldens: every case is scored against the Roslyn oracle
          # (graph/test/csharp/ground-truth), which the step below builds with the
          # .NET 8 SDK. No flag — scoring against the compiler is all it does.
          - lang: csharp
            oracle: ''
    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-node@v4
        with:
          node-version: ${{ env.NODE_VERSION }}

      # Builds the in-repo parser (parser/dist) through the prepare script, and
      # supplies the TypeScript compiler the typescript and javascript oracles run.
      # `npm install`, not `npm ci`: no lock file is committed (#476).
      - run: npm install

      - name: the parser actually built
        run: |
          test -f parser/dist/index.js \
            || { echo "::error::parser/dist/index.js is missing after npm install"; exit 1; }

      # TWO interpreters, because the python suite needs two different things and
      # they cannot be the same version.
      #
      #   3.10 — the tier-1 attribution preflight reads CPython OPCODES, whose
      #          shapes are not stable across minor versions. It resolves
      #          `python3.10` by name, so this only has to exist on PATH.
      #   3.12 — the torture fixtures are SOURCE that has to import: one of them
      #          uses `typing.Self`, which is 3.11+ (PEP 673), so on 3.10 the
      #          tracer dies at import and the family scores nothing.
      #
      # The later setup-python wins for plain `python3`, so 3.12 must come second.
      - uses: actions/setup-python@v5
        with:
          python-version: '3.10'

      - uses: actions/setup-python@v5
        with:
          python-version: '3.12'

      - name: both interpreters are on PATH
        run: |
          set -euo pipefail
          echo "python3    -> $(python3 --version)"
          echo "python3.10 -> $(python3.10 --version)"

      # JDK 24, not the runner's default. The java torture harness reads class files
      # with java.lang.classfile, which is not final before 24 — on an older JDK it
      # exits 77 and the whole ten-family oracle silently does not run.
      - uses: actions/setup-java@v4
        if: matrix.lang == 'java'
        with:
          distribution: temurin
          java-version: '24'

      - uses: actions/setup-dotnet@v4
        if: matrix.lang == 'csharp'
        with:
          dotnet-version: '8.0.x'

      # Without the oracle binary the suite exits 77, which run-suite.sh turns into
      # a failure — so a missing build is loud, never a silent skip.
      - name: build the Roslyn oracle
        if: matrix.lang == 'csharp'
        run: dotnet build -c Release graph/test/csharp/ground-truth/AxiomCsOracle

      - name: cache the Soufflé package
        uses: actions/cache@v4
        with:
          path: ~/souffle-pkg
          key: souffle-deb-${{ env.SOUFFLE_VERSION }}-ubuntu-2404

      # Soufflé is the solver, not a library under test: the engine is compiled
      # from .dl to C++ and linked against Soufflé's headers, so a build of it has
      # to be present the way a compiler has to be present.
      #
      # It is pinned to an exact version AND verified against the checksum upstream
      # published for that release, so what CI links against is decided in this
      # file rather than by whatever the archive happens to serve today. The pin
      # the driver reads is graph/pipeline/engine.conf; this must agree with it.
      - name: install Soufflé ${{ env.SOUFFLE_VERSION }}
        run: |
          set -euo pipefail
          deb="x86_64-ubuntu-2404-souffle-${SOUFFLE_VERSION}-Linux.deb"
          dir="$HOME/souffle-pkg"; mkdir -p "$dir"
          if [ ! -f "$dir/$deb" ]; then
            curl -fsSL --retry 3 -o "$dir/$deb" \
              "https://github.com/souffle-lang/souffle/releases/download/${SOUFFLE_VERSION}/${deb}"
          fi
          echo "${SOUFFLE_SHA512}  ${dir}/${deb}" | sha512sum -c -
          sudo apt-get update -qq
          sudo apt-get install -y --no-install-recommends "$dir/$deb"
          souffle --version | head -2

      # ── the compiled engine ────────────────────────────────────────────────
      # run-souffle.sh caches the compiled solver under a content hash of the
      # .dl program text. Mirroring that key here skips a multi-minute C++ build
      # on every run whose rules did not change.
      # Keyed by this language's ENGINE_ID — the hash the driver itself names its
      # compiled binary by (its rules and the Soufflé version) — so a rule change in
      # one language recompiles that language only, and an unchanged one never does.
      # Two things decide the binary that ENGINE_ID does not cover, so they are in the
      # key too: the driver that compiles it (run-souffle.sh holds the compiler flags),
      # and the target it is compiled for.
      #
      # THE TARGET IS PORTABLE, NOT THE RUNNER'S CPU. The driver's default is
      # -march=native, and a binary built on one hosted runner died with SIGILL on
      # another (runners that report the same model name do not all expose the same
      # instruction set). The key used to carry a hash of the runner's CPU flags to
      # keep such a binary off other CPUs — and so it missed on almost every run,
      # since which CPU a job lands on is luck: a three-minute compile per language,
      # per run, for rules nothing had changed. AXIOM_ENGINE_MARCH=portable compiles
      # for the compiler's baseline x86-64 target instead, as the published engines
      # are, so any runner can run any runner's binary and the CPU leaves the key.
      #
      # A cache saved by a pull request is visible to that pull request only. The
      # ones every pull request can restore are the base branch's and the default
      # branch's, which is why a push to a release branch runs the suites too (see
      # `on.push` and the `changes` job): it leaves the engine for its rules where
      # every pull request into that branch finds it.
      - name: this language's engine id
        id: eid
        run: |
          echo "id=$(bash graph/pipeline/run-souffle.sh --language ${{ matrix.lang }} --print-engine-id)" >> "$GITHUB_OUTPUT"

      - name: restore the compiled Soufflé engine
        id: engine-cache
        if: ${{ !inputs.fresh }}
        uses: actions/cache/restore@v4
        with:
          path: .souffle-cache
          key: souffle-engine-${{ matrix.lang }}-${{ steps.eid.outputs.id }}-${{ hashFiles('graph/pipeline/run-souffle.sh') }}-march-${{ env.AXIOM_ENGINE_MARCH }}

      - name: ${{ matrix.lang }} regression suite
        env:
          AXIOM_PARSER: ${{ github.workspace }}/parser/dist/index.js
          # The driver's default cache is ~/.cache/axiomcode/souffle; point it at the
          # directory the cache step above saves and restores, or it never hits.
          AXIOM_SOUFFLE_CACHE: ${{ github.workspace }}/.souffle-cache
        # The cases run concurrently, up to one per CPU (graph/test/tools/case-pool.sh);
        # AXIOM_SUITE_JOBS=1 here would run them one at a time, as they used to.
        run: bash .github/scripts/run-suite.sh ${{ matrix.lang }} ${{ matrix.oracle }}

      # Saved whether or not the suite passed. The binary does not depend on the verdict:
      # run-souffle.sh publishes it only whole and verified (a temp name, then a rename),
      # so a red run's engine is as good as a green one's, and the run that most needs the
      # next push to be fast is the red one. Not when cancelled, and not unless this
      # language's engine is actually there: a key saved without it would stay taken, and
      # every later run would restore the gap and could never save over it.
      - name: save the compiled Soufflé engine
        if: >-
          ${{ !cancelled() && !inputs.fresh && steps.engine-cache.outputs.cache-hit != 'true'
              && hashFiles(format('.souffle-cache/souffle-engine-{0}-{1}', matrix.lang, steps.eid.outputs.id)) != '' }}
        uses: actions/cache/save@v4
        with:
          path: .souffle-cache
          key: souffle-engine-${{ matrix.lang }}-${{ steps.eid.outputs.id }}-${{ hashFiles('graph/pipeline/run-souffle.sh') }}-march-${{ env.AXIOM_ENGINE_MARCH }}

  # Every language's engine, every platform, built once per release: on the push that
  # lands a new version on main (and in the nightly). publish-npm ships these very
  # artifacts, so what the e2e below tested is what users install (#1350).
  engines:
    name: engines build on every platform
    needs: [build, changes]
    if: needs.changes.outputs.engines == 'true'
    uses: ./.github/workflows/build-engines.yml
    with:
      fresh: ${{ inputs.fresh == true }}

  # THE RELEASE GATE: what a user installs, on every platform, answers every verb. The engines
  # job only proves each binary compiles and starts; the suites run from the checkout. Neither
  # installs the packages, so a missing `files` entry, an engine package the CLI does not find,
  # a query program that needs Soufflé, or a verb that only breaks on Windows reached users.
  # Runs wherever the platform engines are built: on the push that lands a release on main, and in the nightly.
  pack:
    name: pack @axiomcode/code-graph
    needs: [build, changes]
    if: needs.changes.outputs.engines == 'true'
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: '22'
      - run: npm install --no-audit --no-fund
      # --ignore-scripts: `prepare` already built it; the tarball carries what the build produced
      - run: mkdir -p tgz && npm pack --ignore-scripts --pack-destination tgz && ls -la tgz
      # kept as long as the engines: publish-npm ships this exact tarball, whenever the draft is published
      - uses: actions/upload-artifact@v4
        with: { name: code-graph-tgz, path: tgz, retention-days: 90, if-no-files-found: error }

  e2e:
    name: e2e on ${{ matrix.target.platform }}
    needs: [changes, engines, pack]
    if: needs.changes.outputs.engines == 'true'
    strategy:
      fail-fast: false
      matrix:
        target:
          - { os: ubuntu-24.04,     platform: linux-x64 }
          - { os: ubuntu-24.04-arm, platform: linux-arm64 }
          - { os: windows-2025,     platform: win32-x64 }
          - { os: macos-15,         platform: darwin-arm64 }
          - { os: macos-15-intel,   platform: darwin-x64 }
    runs-on: ${{ matrix.target.os }}
    timeout-minutes: 45
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: '22'
      - uses: actions/setup-python@v5
        with:
          python-version: '3.12'
      - uses: actions/download-artifact@v4
        with: { name: 'engines-${{ matrix.target.platform }}', path: artifacts/engines }
      - uses: actions/download-artifact@v4
        with: { name: code-graph-tgz, path: artifacts/tgz }
      - name: installed from the tarballs, no Soufflé, every language, every verb
        shell: bash
        env:
          PLATFORM: ${{ matrix.target.platform }}
        run: |
          set -euo pipefail
          command -v souffle && { echo "::error::this runner has souffle; the gate would not prove anything"; exit 1; }
          export CODEGRAPH_TGZ="$(ls "$PWD"/artifacts/tgz/*.tgz)"
          for lang in java typescript python javascript csharp; do
            # A glob, not ls | head: under pipefail, head exiting early SIGPIPEs ls and fails the step.
            cases=(graph/test/"$lang"/cases/*/src)
            case_dir="${cases[0]}"
            [ -d "$case_dir" ] || { echo "::error::no $lang case under graph/test/$lang/cases"; exit 1; }
            echo "::group::$lang ($case_dir)"
            bash .github/scripts/e2e-install.sh artifacts/engines "$PLATFORM" "$lang" "$case_dir"
            echo "::endgroup::"
          done

  # A single job the branch ruleset can require. Without it, every new matrix
  # entry has to be added to the protection rules by hand, and a matrix job that
  # fails to start reports nothing at all — which a ruleset reads as "not
  # failing" rather than as "did not run".
  ci:
    name: CI
    runs-on: ubuntu-24.04
    needs: [changes, build, hygiene, engine, engines, pack, e2e]
    if: always()
    steps:
      - name: every required job succeeded
        run: |
          # The expression quotes its separator with SINGLE quotes because that is
          # the only string delimiter a GitHub expression has. A double quote there
          # is a lex error that invalidates the entire workflow file, and the run
          # then fails in zero seconds with no job having started.
          # changes, build and hygiene always run and must succeed. engine and
          # engines may be SKIPPED, but only when `changes` said so; any other
          # skip (a job that never started) is a failure.
          always="${{ needs.changes.result }} ${{ needs.build.result }} ${{ needs.hygiene.result }}"
          echo "always-run jobs: $always"
          for r in $always; do
            [ "$r" = "success" ] || { echo "::error::a required job reported '$r'"; exit 1; }
          done
          check() {  # name result expected-to-run
            if [ "$3" = true ]; then
              [ "$2" = success ] || { echo "::error::$1 reported '$2'"; exit 1; }
            else
              [ "$2" = skipped ] || { echo "::error::$1 reported '$2' though nothing it tests changed"; exit 1; }
            fi
            echo "$1: $2"
          }
          check "engine suites" "${{ needs.engine.result }}" "${{ needs.changes.outputs.code }}"
          check "platform engines" "${{ needs.engines.result }}" "${{ needs.changes.outputs.engines }}"
          check "pack" "${{ needs.pack.result }}" "${{ needs.changes.outputs.engines }}"
          check "e2e on every platform" "${{ needs.e2e.result }}" "${{ needs.changes.outputs.engines }}"
          echo "all required jobs passed"