[tae898]

Since #92 the JVM wheel is py3-none-<platform>, the same file for every Python, and test-jvm.yml builds it once for each platform. test-python-examples.yml still built it in each of its legs. This PR gives it the same build job and a per-leg download.

Build count

Event	Wheel builds before	Wheel builds after
push to main, nightly, dispatch, release	16 (all py3-none, 12 of them identical repeats)	4 (one for each platform)
pull request	4	4 (one new Build wheel for examples (<platform>) job in front of each leg)

On a pull request the number of builds does not change, because the examples already run one Python for each platform there. What changes on a pull request is that the build moves into its own job, so there are 4 more jobs; each leg no longer sets up Java or Docker Buildx, downloads the jars or installs build tools. If the extra job per platform costs more queueing than it saves on pull requests, say so and I will build inline for that event.

What changes

• download-jars also outputs wheel-matrix (scripts/select_matrix.py examples --event ... --wheels, which already existed).
• New build job, one leg for each platform: it runs scripts/build.sh, which checks size, licences and tag, then scripts/verify_wheel_tag.py again (as in test-jvm.yml, so the tag shows in the log and summary), and uploads the wheel as examples-wheel-<platform> (1 day). The scripts are called, not copied.
• Test Python Examples (<platform>, Python <version>) downloads the wheel of its platform and installs it. Leg names, the examples matrix (newest Python on each platform on a pull request, 16 legs on the other events) and the examples that run are unchanged.
• Examples Test Summary also needs build.
• The artifact is named examples-wheel-*, not wheel-*: the release calls both workflows in one run, release-jvm.yml downloads wheel-* and expects exactly four wheels, and two uploads of the same name would collide.
• ci.adoc (one sentence, plus the check counts), the same counts in CLAUDE.md: a pull request runs 28 checks (was 24), the other events 40 (was 36). The counts come from the matrices, not from memory.
• tests/test_select_matrix.py gets a test for the examples workflow like the one for test-jvm.yml: the --wheels matrix, no Python in the build job, build.sh only there, the wheel download in the leg, and no wheel-* artifact name.

Not done

test-jvm.yml and this workflow each build the same wheel for a commit, so a push to main builds 8 wheels, not 4. Sharing one build between the two workflows would need a reusable workflow or a combined one; I left that out of this change.

Checked

• Workflow YAML parses; the job graph is download-jars -> build -> test-examples -> test-examples-summary.
• python3 scripts/check_no_embedded_name.py --root .: ok.
• tests/test_select_matrix.py, test_check_version.py, test_release_name_guards.py, test_select_native_matrix.py, test_no_embedded_name.py: 109 passed.
• The workflow file itself changed, so this PR's own run exercises the new path on the four platforms.

🤖 Generated with Claude Code