[tae898]

What

A pull request now runs a small test matrix. Push to main, the nightly schedule, a manual dispatch and the release workflow run the full matrix exactly as before.

ArcadeData's org shares one runner cap with upstream ArcadeDB, so every macOS or Windows leg a pull request starts can delay upstream jobs. This builds on #38 (docs-only and draft PRs skip, superseded PR runs cancel); all of that stays.

Which legs run

Event	Unit tests (Test arcadedb)	Examples (Test Python Examples)
pull_request	linux/amd64 x 3.10-3.14, and linux/arm64, darwin/arm64, windows/amd64 x 3.10 and 3.14 = 11 legs	all four platforms x 3.14 = 4 legs
push to main, schedule, workflow_dispatch, workflow_call from release.yml	20 legs (4 platforms x 5 Pythons)	20 legs

The examples keep all four platforms on pull requests because they have platform-specific code: os.name == "nt" branches in examples 07 to 12, and per-platform workflow steps (CPU torch index on macOS, UTF-8 on Windows).

Check counts (counted from the YAML, old versus new, jobs with a matrix counted per leg)

Event	Before	After
Non-draft pull request	46	21
Push to main / nightly / dispatch / release	46	46

A pull request now runs 21 checks: 11 + 4 test legs, 2 x Download ArcadeDB JARs, 2 summary jobs, Bandit security scan and Dependency floor audit. The test legs go from 40 to 15, and the macOS and Windows legs from 20 to 6. A merge waiter should require at least 21 checks on a non-draft PR that touches more than docs; a docs-only or draft PR runs fewer, as before.

How

• scripts/select_matrix.py holds the platform and Python lists and the pull request selection, and prints the strategy.matrix JSON for the event. Any event other than pull_request (including one it does not know) gets the full matrix.
• The first job of each workflow, Download ArcadeDB JARs, runs it in a new first step and exposes the result as a job output. The test job uses matrix: ${{ fromJSON(needs.download-jars.outputs.matrix) }}. No extra job, so no extra runner, and the existing job names are unchanged.
• Each leg still carries platform, runs-on and python-version, so the check names stay Test arcadedb (<platform>, Python <version>) and Test Python Examples (<platform>, Python <version>), and their history stays comparable. main is not branch-protected, so dropping PR legs cannot block a merge.
• release.yml calls both workflows with workflow_call, where github.event_name is the caller's event (a tag push), so a release still builds all 20 wheels that verify-wheels counts. Only a comment there changed.
• There are no Python 3.15 legs in either workflow (3.15 appears only in two build comments about JPype), so there was nothing to move.

Tests and lint

• tests/test_select_matrix.py (21 cases, written first; the wiring cases failed before the workflow edit): the full matrix is written out independently and must equal 4 x 5 for every non-PR event and an unknown one; the PR matrices are exact and a subset of the full one; the Python list equals the pyproject.toml classifiers (release.yml counts wheels from them); both workflows take their matrix from the script and keep no copy of the lists.
• actionlint 1.7.12 (the pre-commit rev) is clean on all three workflows; black, isort and Bandit are clean on the new files.
• This PR's own run is the live check that the PR matrix expands as above.

Docs

New page development/ci.adoc (in the nav) describing what runs when and how to change the selection. The three "CI runs all five versions on every platform" sentences (README, index.adoc, installation.adoc) now say for main, the nightly run and each release, and CLAUDE.md's CI section has the new rule.

🤖 Generated with Claude Code