[tae898]

The arcadedb-jvm wheels become py3-none-<platform>: one wheel per platform, 4 per release (was cp311..cp314 x 4 platforms = 16 wheels of about 65 MB, about 1 GB per release; PyPI's default project limit is 10 GB). 26.11.1.dev0 keeps its 16 files on PyPI; this lands for the next version.

Tag mechanism

setup.py keeps BinaryDistribution.has_ext_modules() -> True (platform wheel, platlib, bundled JRE) and adds a bdist_wheel subclass whose get_tag() takes the platform from the base class and returns ("py3", "none", plat). The platform part is unchanged: --plat-name manylinux_2_28_* in Dockerfile.build, _PYTHON_HOST_PLATFORM on macOS, win_amd64 on Windows. Only the interpreter and ABI tags go. pyproject.toml needs setuptools >= 70.1 for its own bdist_wheel.

Python 3.15

requires-python = ">=3.11,<3.15", with a comment. The wheel tag no longer keeps pip off 3.15, and jpype1 1.7.1 has no cp315 wheel (support is merged in jpype-project/jpype#1501, not released). Remove the cap, add the classifier and the matrix entry when JPype ships cp315 wheels. tests/test_select_matrix.py pins the cap to "minor after the newest classifier".

CI

• test-jvm.yml: a build job builds each platform once and uploads wheel-<platform>; the test job downloads it and installs it into one Python (no Java, Docker or build tools there). The Test arcadedb (<platform>, Python <version>) names are unchanged; new checks are Build arcadedb-jvm wheel (<platform>).
• Test legs now mirror test-native.yml: every Python (3.11 to 3.14) on linux/amd64, only 3.11 and 3.14 on linux/arm64, macOS arm64 and Windows. That is 10 legs on every event (it was 16 on main, nightly, dispatch and release; the extra legs tested the same file again). scripts/select_matrix.py holds the legs (JVM_LEGS); --wheels prints the build legs.
• release-jvm.yml verify: 4 wheels, one per platform, all py3-none in name and in the WHEEL file (scripts/verify_wheel_tag.py, also run by build.sh and the build job), sizes as before; downloads wheel-*.
• after_tag_verify.sh reads the previous release's cpXY files as the py3-none set, so the first release after this compares cleanly.
• Not changed: test-python-examples.yml still builds a wheel in each of its legs (the same py3-none file); a follow-up can build once per platform there too.

Docs and tests

development/ci.adoc (one wheel per platform, the 10 legs, 24 checks on a pull request and 36 elsewhere), release.adoc, installation.adoc, which-package.adoc, concurrency.adoc, CLAUDE.md. The JVM platform tags in the docs said manylinux_2_34; the wheels are manylinux_2_28 (as on PyPI for 26.11.1.dev0), fixed.

Tests: tests/test_wheel_tags.py (new: builds with setup.py for three platform names and checks name and WHEEL; installed-wheel check), test_select_matrix.py, test_release_name_guards.py (4 wheels accepted; the old 16, a cpXY wheel and a missing platform refused).

Behaviour change to know

Free-threaded 3.14t: the old cp314 tag kept the JVM wheel off 3.14t. py3-none does not, and JPype has Linux cp314t wheels, so pip can now install the package there. Nothing tests it; concurrency.adoc says so. Not claimed in the classifiers.

Checked locally

python -m build with --plat-name manylinux_2_28_x86_64, manylinux_2_28_aarch64, win_amd64 and _PYTHON_HOST_PLATFORM=macosx-11.0-arm64: Tag: py3-none-<plat>, Root-Is-Purelib: false, Requires-Python: <3.15,>=3.11. scripts/check_no_embedded_name.py --root . clean, bandit (as in CI) clean, the touched test files pass.

Also

• Native parity fix for JVM build: Java System.out/System.err go to a file the library owns when the console is off (Windows data corruption) #88's tests/test_log_redirect.py: the Java stream probes and the JEP 498 trigger call JPype, so they now run on the JVM build only (arcadedb.__backend__ == "jvm") and the redirect-hygiene assertions stay on both builds (21 passed on the native wheel and on the JVM wheel locally).

🤖 Generated with Claude Code