Switch to opencv-python-headless, declare numpy explicitly

[wrignj08]

Summary

MultiClean's runtime uses exactly three OpenCV functions — morphologyEx, connectedComponentsWithStats and distanceTransformWithLabels. No highgui, no video I/O. The GUI build of OpenCV was therefore pulling in shared libraries we never call, which slim container and CI images do not ship, producing the classic:

ImportError: libGL.so.1: cannot open shared object file

Given the remote-sensing / batch-processing audience, that failure mode is squarely in the path of normal use.

This PR:

• Switches the runtime dependency to opencv-python-headless.
• Declares numpy>=1.21 explicitly. It is imported directly in multiclean/utils.py and multiclean/multiclean.py and appears in the public type signatures (np.ndarray), but was only ever installed transitively via OpenCV — we were relying on another package's dependency graph for something we import ourselves.
• Raises the OpenCV floor to >=4.10. Builds before 4.10.0.84 are compiled against NumPy 1.x and fail at import when paired with NumPy 2.x, which the previous >=4.0 floor allowed a resolver to select.

 dependencies = [
-    "opencv-python>=4.0",
+    "opencv-python-headless>=4.10",
+    "numpy>=1.21",
 ]

Verification

Full CI check set, run locally against a freshly synced environment:

Check	Result
ruff check .	All checks passed
mypy multiclean/	Success, no issues in 3 source files
pytest tests/	322 passed in 5.50s

Confirmed only opencv-python-headless is installed — no duplicate cv2 from both distributions.

Reviewer notes

Two things worth a deliberate decision rather than a rubber stamp:

The resolver picked OpenCV 5.0.0, not 4.x. The old opencv-python>=4.0 pin was holding us on the 4.x line in practice; changing the distribution name let the resolver cross a major version. All 322 tests pass on 5.0.0, so the three functions we use are stable across the boundary — but downstream users will get the same jump. If we'd rather not ship a major OpenCV bump alongside the headless switch, add a <5 cap to this PR.

Both distributions provide the same cv2 module. If a downstream environment also needs the full opencv-python build for other work, installing both means one clobbers the other's files. This is noted in the changelog entry for users. For a batch-processing library headless is the conventional default, but it is a downstream-visible change.

uv.lock is gitignored in this repo, so it is intentionally not part of this diff.

Release

This is a downstream-visible dependency change, so it warrants a minor version bump rather than a patch when tagged. Version is setuptools-scm-driven, so per RELEASING.md that happens at tag time — no version edit in this PR.

🤖 Generated with Claude Code