Test on every platform we publish a wheel for (today: build 4, test 1)

[n1ckyb]

The gap

publish.yml builds wheels on ubuntu-latest, macos-latest, windows-latest and windows-11-arm.
ci.yml runs the test suite on windows-latest, Python 3.12, only.

So three of the four artefacts we publish have never had a test run against them on the
platform they target. They are compiled and uploaded on the strength of a Windows test run.

This is the same shape of gap that produced the 0.0.1 pull: a check that proves the code
builds, standing in for a check that proves the artefact works.

Why it is not theoretical

The parts most likely to diverge by platform are exactly the parts we lean on:

• ctypes loading of the core: .dll / .so / .dylib resolution differs per platform
• path handling, which Windows-only testing systematically under-exercises
• the wasm component cache directory and its permissions
• console-script shims, which differ between Scripts/ and bin/

What to do

1. Widen ci.yml to os: [ubuntu-latest, macos-latest, windows-latest] x
   python-version: ["3.10", "3.11", "3.12", "3.13"]
2. Run scripts/smoke_published_wheel.py --wheel <built> on each matrix leg, against the
   wheel built by that leg — not against a source checkout
3. Make the smoke run a required check before any tag

Definition of done

No wheel is published for a platform on which the full suite and the artefact smoke have not
both passed. If we cannot test a platform, we do not ship a wheel for it — an untested
artefact is a claim we have not earned.

[n1ckyb]

Correction to the matrix above: pyproject.toml sets requires-python = ">=3.12", so the version axis is 3.12 and 3.13, not 3.10–3.13 as originally written.

Worth noting separately, because it is a likely cause of user-reported "pip install didn't work": on Python ≤3.11 pip reports Could not find a version that satisfies the requirement intentumdiff-python, which reads as a missing or broken package rather than an unsupported interpreter. Now documented at https://buchochelliq-labs.github.io/intentumdiff-docs/troubleshooting/

[n1ckyb]

A related gap worth folding in: tests that exist but never run

The extension turned out to be running 17 of its 246 unit tests — a deleted test's stale
compiled output kept failing, and the fix at the time excluded it by name, which silently
excluded sixteen healthy files too (details and fix: buchochelliq-labs/intentumdiff-vscode#19).

The tests all passed; nothing was watching them.

This issue is about breadth of platforms, but it shares a root question with that one: does
the suite we believe is running actually run? Widening to a 3-OS matrix is worth much less if
a subset is silently skipped on some legs.

Suggested additions here:

• Record the collected test count per matrix leg and fail if it drops below a floor.
  Platform-conditional skips are legitimate; a silent collapse is not
• Report skips explicitly per leg, so "skipped on macOS" is a decision someone made rather
  than something nobody noticed