Skip to content

Latest commit

 

History

History
140 lines (115 loc) · 7.64 KB

gate.md

File metadata and controls

140 lines (115 loc) · 7.64 KB

Gate and CI

One entry point runs every local gate: ../scripts/release_check.py. Each arm is a separately runnable selection, and CI invokes the same arms instead of keeping a second copy of any command.

uv run python scripts/release_check.py              # fast arms (default)
uv run python scripts/release_check.py --arm all --gnn-runtime /path/to/gnn-runtime/bin/python
uv run python scripts/release_check.py --list       # arms and their commands

Run uv sync --locked --extra dev before the gate. Its child commands set UV_NO_SYNC=1 to reuse that prepared project environment instead of repeating dependency synchronization. Release selections also require the independently reviewed GNN_BOOTSTRAP and GNN_SOURCE_MANIFEST inputs described in gnn.md.

Captured command failures retain their original stdout and stderr before the nonzero exit is recorded. A failed pytest/coverage arm cannot hide its diagnostics or become a passing receipt.

Measured wall times and test counts live in the generated _generated/gate-metrics.md, which shows the recorded before/after arm times and labels baseline workloads that are not equivalent to their arms. These are one-run, host-specific receipts, not a controlled benchmark or a claimed speedup. The offline package baseline failed because its required package was absent from the local cache; preserve that failure rather than comparing it as a successful workload. Do not hand-edit measurements.

Arms

Arm Command In all
lint uv run ruff check blockference tests GRTs scripts yes
types uv run mypy blockference (configured at the requires-python floor) yes
test uv run pytest -m "not slow" no
test-full uv run pytest --cov=blockference --cov-report=term --cov-report=xml --cov-fail-under=90 yes
docs scripts/check_docs_links.py, scripts/gate_metrics.py --check yes
manuscript scripts/manuscript_variables.py --check, scripts/validate_manuscript.py, scripts/build_manuscript.py, scripts/build_manuscript.py --check; gnn_* metadata comes from checked producer inspection, not a qualification verdict yes
gnn scripts/export_gnn_spec.py --check, then scripts/validate_gnn_spec.py --verify-tag-remote, which owns both guarded deep conformance and the shared syntax/pointer/workflow/job check; remote tag verification needs network yes
smoke pipeline, validation, and standalone run JSON smoke checks in a temporary directory yes
package invocation-owned uv build --wheel --out-dir, locked dependency constraints, and an isolated installed-wheel import outside the source tree yes
render scripts/build_manuscript.py --render pdf (Pandoc + XeLaTeX) no

--arm is repeatable and arms always run in the canonical order above. fast is the default because it is the selection used while editing; all is the release gate. The slow marker (declared in pyproject.toml) keeps notebook execution, real-pipeline integration, CLI smoke, and figure-byte checks out of the fast arm. They stay in test-full.

GNN conformance

The canonical pin is ../specs/gnn-pin.json with schema gnn-importer-pin/v1. It pins ActiveInferenceInstitute/Generalized_Notation_Notation at v5.1.0 (75c834bd8519dc7b5adc4d24f16bf393cfd1ecdf) and records SHA-256 digests for the five parser files under src/gnn/. The pointer ../.github/GNN.md records exactly what the conformance job parses.

The conformance job's key is gnn-conformance and its display name is gnn conformance (pinned).

The gnn arm checks exporter drift, then invokes the project validator with the selected external runtime and remote-tag request. The validator owns deep conformance and the producer's shared syntax/pointer/workflow/job check in separate isolated, qualified lifetimes. No trailing unchecked python -m gnn.importers command bypasses admission.

Select the external interpreter and independently reviewed bootstrap/full manifest as described in gnn.md. Admission binds complete content and actual module ownership through verdict exit; five parser hashes alone are insufficient. Installed mode additionally needs real canonical PEP610 Git provenance. Source-artifact proof is separately labelled and never becomes an installed-origin or scientific verdict. Missing resources, origin, runtime, or failed revalidation fails the arm and therefore all, never silently skips.

The package arm accepts exactly one fresh regular, non-symlink wheel from its own build directory. It neither installs nor deletes stale repository dist contents. Protocol unit controls are not an actual build/install smoke.

GitHub Actions

../.github/workflows/ci.yml calls these arms:

Job Arms
test (matrix: Python 3.11 only) --arm test-full
gnn-conformance (Python 3.12; display name gnn conformance (pinned)) --arm all, with the provisioned GNN_RUNTIME_PYTHON and checksum-verified GNN_BOOTSTRAP / GNN_SOURCE_MANIFEST

The full suite therefore runs on both supported CI Python versions. The remaining release arms run in gnn-conformance on Python 3.12, alongside the real pinned conformance check; there is no separate gate job. The pinned GNN runtime remains separate from the locked library environment because their NumPy requirements differ. PDF rendering is an explicit local --arm render selection, not part of all or CI.

Before provisioning the parser runtime, the workflow downloads the reviewed bootstrap and full v5.1.0 manifest from immutable producer resource commit 66d491f86895443cf588011cc6f40c37b8bae8fd, verifies both source-owned SHA-256 digests, and forwards their paths to all. That delivery revision is not a parser repin. Acquisition or digest failure stops the job before execution. The bootstrap's known strict-origin limitation and the separate producer qualification prerequisite are retained in gnn.md.

The workflow installs uv directly with astral-sh/setup-uv@68559e99cc807c1eb7e9238fd43a15bba460eec2 (v10.3.3, pinned by commit because setup-uv publishes no floating v10 tag). It does not use the Institute composite action ActiveInferenceInstitute/.github/.github/actions/setup-python-uv@7159e67315330aac8e414aebae0ad1e999a41448; the workflow installs the locked environment directly.

Recording measurements

release_check.py --timings PATH writes the actual commands, statuses, and elapsed times for the selected run, including a failed arm. Use a fresh receipt path for each run; selecting render separately records the explicit local PDF build without adding it to CI. A passing render arm also records artifact: the repository-relative PDF path, exact byte count, SHA-256, and source commit (git rev-parse HEAD). Commit source changes before recording a release render so that commit identifies the rendered source. The PDF remains ignored; the receipt identifies the local bytes without publishing the binary.

Import receipts before regenerating the checked metrics page:

uv run python scripts/gate_metrics.py \
  --record-baseline /path/to/before-receipt.json \
  --record-baseline /path/to/before-render-receipt.json \
  --record-timings /path/to/after-receipt.json
uv run python scripts/gate_metrics.py --check

Replace these paths with receipts produced by real gate runs. --record-baseline is repeatable for disjoint selections; --record-timings accepts one combined --arm all --arm test --arm render receipt. Do not synthesize timings or turn failed, missing, or non-equivalent measurements into passes.