fix(tracking): ADR-028 RC4 -- unoriented SkillCorner frames + paired GkCompletionModel retrain -- silly-kicks 4.73.0 (PR-S141, ADR-051) #501
Workflow file for this run
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: CI | |
| on: | |
| push: | |
| branches: [main] | |
| pull_request: | |
| branches: [main] | |
| permissions: read-all | |
| # A newer push/PR update supersedes an in-flight run for the same ref instead of | |
| # queueing, so CI resources aren't spent on stale commits. | |
| concurrency: | |
| group: ci-${{ github.workflow }}-${{ github.ref }} | |
| cancel-in-progress: true | |
| jobs: | |
| lint: | |
| runs-on: ubuntu-latest | |
| steps: | |
| - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1 | |
| - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0 | |
| with: | |
| python-version: "3.12" | |
| - run: pip install ruff==0.15.7 pyright==1.1.409 pandas-stubs==2.3.3.260113 | |
| - run: ruff check silly_kicks/ tests/ scripts/ | |
| - run: ruff format --check silly_kicks/ tests/ scripts/ | |
| - run: pip install -e ".[test]" | |
| # config-driven include: silly_kicks + tests + scripts (pyproject [tool.pyright]) | |
| - run: pyright | |
| test: | |
| runs-on: ${{ matrix.os }} | |
| strategy: | |
| matrix: | |
| os: [ubuntu-latest, windows-latest] | |
| python-version: ["3.10", "3.11", "3.12"] | |
| exclude: | |
| - os: windows-latest | |
| python-version: "3.10" | |
| - os: windows-latest | |
| python-version: "3.11" | |
| include: | |
| # Augments the existing ubuntu-3.12 leg (GitHub merges a matching include rather | |
| # than creating a new leg) with primary: true -- the single source of truth for | |
| # "which leg runs the platform-invariant @slow set + the benchmark step". See ADR-023. | |
| - os: ubuntu-latest | |
| python-version: "3.12" | |
| primary: true | |
| env: | |
| # Strict mode: promote the in-converter direction-of-play invariant | |
| # assertion in `silly_kicks.spadl.utils._finalize_output` (PR-S22 / | |
| # ADR-006) from default-warn to raise, AND promote the input-convention | |
| # validator's `on_mismatch` resolution from "warn" to "raise". Catches | |
| # home_team_id-heuristic regressions in test fixtures and any future | |
| # loader-shape drift in consumer pipelines that pin against silly-kicks | |
| # main. See feedback_home_team_id_heuristic_fragile.md. | |
| SILLY_KICKS_ASSERT_INVARIANTS: "1" | |
| steps: | |
| - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1 | |
| - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0 | |
| with: | |
| python-version: ${{ matrix.python-version }} | |
| # `das` (accessible-space) is installed on EVERY leg, not just the primary one: the | |
| # TF-28 DAS suites are all `importorskip`-guarded, so without the extra they SKIP | |
| # rather than fail, and they had therefore never run in CI at all. gkdv's DAS arm | |
| # (TF-19 PR-3) is a second consumer of that subsystem, whose correctness depends on a | |
| # direction-inference subtlety inside it -- a guard that never runs is not a guard. | |
| # The arm's own STRUCTURAL direction-pinning guard needs no extra and runs regardless. | |
| - run: pip install -e ".[kloppy,xgboost,das,test]" | |
| # Bulk suite (serial). xdist `-n auto` was reverted: on the 4-core/7GB CI runners it | |
| # regressed py3.12 from pass to a memory/JIT-pressure kill (4x heavy ML imports + 4x numba | |
| # compiles + the ghost-GK suite), the opposite of the 16-core local speedup. The dominant | |
| # ghost-GK cost is instead cut at the source (exact cpu-numba KDE backend + shrunk golden). | |
| # Perf budgets are now deterministic structural guards in the bulk (no wall-clock), so this | |
| # step is stable; the pure benchmark *measurements* run single-threaded below for clean timing. | |
| # | |
| # Slow-test gating (ADR-023): the platform-/interpreter-INVARIANT heavy tests (train-script | |
| # smokes, same-run internal-consistency/parity, calibration cache-equivalence) carry | |
| # @pytest.mark.slow and run ONCE on the primary leg (ubuntu-3.12); every other leg runs the | |
| # full fast/contract suite with `not slow`. Version-sensitive tests (golden/snapshot/absolute) | |
| # are deliberately NOT marked slow -- they stay on all legs (OS + interpreter axes). The | |
| # partition is structural (matrix.primary) and guarded by tests/test_ci_slow_gating_wired.py. | |
| # Mutually-exclusive `if:`s keep every leg at exactly ONE bulk pytest process. | |
| - if: ${{ !matrix.primary }} | |
| run: pytest tests/ -m "not e2e and not slow" --benchmark-skip --tb=short --durations=25 | |
| - if: ${{ matrix.primary }} | |
| run: pytest tests/ -m "not e2e" --benchmark-skip --tb=short --durations=25 | |
| # Benchmark measurements (no hard timing asserts after the structural-guard conversion), run | |
| # single-threaded on the primary leg only so the recorded numbers are clean trend data. | |
| - if: ${{ matrix.primary }} | |
| run: pytest tests/ -m "not e2e" --benchmark-only --tb=short | |
| # Public-API doctest enforcement (PR-S124). Initial scope: the PUBLIC surface only -- | |
| # `--ignore-glob="*/_[!_]*.py"` skips single-underscore private modules while KEEPING dunder | |
| # `__init__.py`, so runnable examples on non-underscore modules are executed and stay honest. | |
| # Private-module examples are kept CORRECT (the whole package sweep is clean) but are | |
| # deliberately NOT executed here, to bound CI wall-clock. Runs on EVERY leg because doctest | |
| # output (repr) is version-sensitive; verified green on py3.10 + py3.14. No `addopts` in | |
| # pyproject, and `testpaths` is ignored when a path is passed, so no `-o` override is needed. | |
| - run: pytest --doctest-modules silly_kicks/ --ignore-glob="*/_[!_]*.py" --tb=short |