tests.yml

name: tests

on:
  push:
    branches: [main]
  pull_request:

jobs:
  test:
    runs-on: ubuntu-latest
    strategy:
      # One failing version must not cancel the others. Twice now a single
      # transient `git clone` failure inside the clean-clone check has taken
      # the whole matrix down, and on the first occurrence a job that had
      # already reported "1257 passed" was marked failed because it was
      # cancelled mid-cleanup — which made one flake read as three. Letting
      # every version finish is what distinguishes an environment failure
      # from a real one, and that distinction is the reason to run three.
      fail-fast: false
      matrix:
        # The floor, and the two the developer actually works on. 3.10 is what
        # the clean-clone check runs, deliberately older than the working
        # environment so an incompatibility is found before CI rather than by
        # it; 3.14 is the interpreter every local suite run uses, and until
        # 2026-09-09 nothing in CI covered it or the release below it. The
        # versions between the floor and the ceiling are dropped rather than
        # kept: a defect present in one of them and in neither end is rarer
        # than a defect in the version the developer runs all day and CI never
        # saw.
        python-version: ['3.10', '3.13', '3.14']
    steps:
      - uses: actions/checkout@v4
        with:
          # Full history, not the default shallow clone. story-053 converted
          # every module that resolved this repository's history for an *input*
          # onto committed fixtures under tests/history-fixtures/, so this is no
          # longer load-bearing for those. It still is for the modules whose
          # subject genuinely is what a past commit did — the entries
          # tests/test_baseline_honesty.py declares as history readers rather
          # than pending conversion, currently test_baseline_honesty.py,
          # test_baseline_resolution_is_single.py, test_git_history_loading_retired.py
          # and test_validation_module_naming.py. A one-commit checkout has no
          # revision older than the change under test, so those go red.
          fetch-depth: 0
      - uses: actions/setup-python@v5
        with:
          python-version: ${{ matrix.python-version }}
      # Install from the tracked declaration rather than a hand-listed set, so
      # the plugin the configured test_command needs cannot be missing here
      # while present locally.
      - run: pip install -r requirements-dev.txt
      # One retry, of the failures only, and only here. It was added while the
      # clean-clone and revert checks still cloned with `--no-hardlinks`, which
      # copies .git/objects as a directory tree: git enumerates the tree, then
      # copies it, and three CI runs failed because a file vanished in between.
      # One named objects/pack/multi-pack-index.lock, which exists only while
      # something is writing a multi-pack index in the *source* repository.
      #
      # story-033 removed that cause. Both checks now clone over git's normal
      # transport (`--no-local`), which negotiates a pack over a pipe and never
      # reads the source's object files at all, so a concurrent writer in the
      # source cannot reach the clone. The failure this step was added for is
      # impossible rather than unlikely, and no call site is outstanding.
      #
      # The step stays anyway, for the reason it always named as its second
      # purpose: the next unknown. It is not evidence of a known flake, and a
      # green job that used it still says so below.
      #
      # The retry lives here rather than in orchestration/ deliberately. A
      # harness that retries its own clone behaves differently depending on
      # where it runs, and a developer running the suite locally should see
      # the failure rather than a green run that quietly needed two attempts.
      # So the harness is unchanged and the environment that produces the
      # flake is the one that absorbs it.
      #
      # `--last-failed` re-runs only what failed, so the log names exactly
      # what needed a second attempt, and the warning puts it in the run
      # summary. A job that is green *because* of this step says so.
      - name: tests (one retry of failures on a second attempt)
        run: |
          set -o pipefail
          if python -m pytest tests/ -q; then exit 0; fi
          echo "::warning title=Suite retried::The first attempt failed; re-running only the failed tests once. If this job is green, it needed two attempts — check the list below before trusting it."
          python -m pytest tests/ -q --last-failed