Prune the sdist and prepare 0.5.0 release

[wrignj08]

Two commits: the sdist fix, then the 0.5.0 changelog promotion. Merging this gives a main commit that is ready to tag v0.5.0 per RELEASING.md.

---

1. Keep notebooks and README artwork out of the sdist

setuptools-scm's file finder hands the sdist every git-tracked file, so the example notebooks and the sample rasters they read were swept into every release automatically.

Before

multiclean-0.4.1.tar.gz                       7.2 MB
├── notebooks/Cloud example.ipynb             1.1 MB
├── notebooks/Land use example.ipynb          172 KB
├── notebooks/Land use example (Colab).ipynb  172 KB
├── notebooks/data/*.tif                      3.5 MB
├── assets/multiclean-logo.svg                112 KB
├── assets/Make README graphic.ipynb           32 KB
├── assets/land_use_before_after.png           44 KB
├── .vscode/settings.json
└── ...

After

3.9 MB. notebooks/, assets/ and .vscode/ are pruned via a new MANIFEST.in.

The wheel was never affected — it is built from packages alone — and is byte-for-byte unchanged at 12 KB.

What is kept, and why

Path	Size
tests/	4.9 MB	An unpacked sdist should be verifiable. Confirmed: unpacked and ran pytest, 573 passed.
.github/workflows/	8 KB	Lets a packager see how the release is built.
RELEASING.md, CHANGELOG.md	12 KB	Release process and history.
.gitignore, .pre-commit-config.yaml	12 KB	Left alone; pruning ~12 KB is churn for no gain.

assets/ is safe to drop because the README references its images by absolute raw.githubusercontent.com URL, not by relative path, so nothing in the sdist reads them. The notebooks stay on GitHub, where the README's tutorial badge links them.

Difference from OmniWatermask 0.6.1

This is the same fix that project shipped, but its MANIFEST.in keeps tests/ on the grounds that "downstream packagers (conda-forge) run the suite against the sdist." That reason does not hold here — MultiClean's feedstock recipe only does imports: plus a pip check and a version assertion; it never invokes pytest.

tests/ is kept anyway, deliberately, so that an unpacked sdist stays verifiable. Its 4.9 MB of fixture rasters is now most of what remains in the tarball; the MANIFEST.in comment records that as the only thing left to reconsider if size ever becomes a concern again.

---

2. Prepare 0.5.0

Promotes [Unreleased] to [0.5.0] - 2026-08-14 and adds the comparison link.

A minor bump rather than a patch, because the even-kernel fix in #5 changes output for even smooth_edge_size and the default is 2 — anyone running on defaults gets different results from the same input. The changelog entry says so explicitly and tells them to regenerate cached results.

0.5.0 contains

• Fix one-pixel shift when smooth_edge_size is even #5 — edge smoothing no longer shifts the output one pixel down and right for even smooth_edge_size.
• This PR — the sdist no longer ships notebooks, their rasters, or README artwork.

After merging

git checkout main && git pull
git tag v0.5.0
git push origin v0.5.0

Then approve the pypi environment deployment in Actions.

🤖 Generated with Claude Code