Thanks! I had noted this as a TODO item in #1972, but didn't yet get to it (and didn't expect our CI to fail for it, but so this is a good reminder ;))

Ideally we would honor the copy keyword if passed. I think in our case, this just means raising an error if copy is False, because the coordinates we return are always a copy?
But this is less critical, and could also be done later in a follow-up.

I had a play with various copy=True/False/None versions, and it all seems minor impact, mostly because CoordinateSequence objects aren't really used as separate objects. Each .coords property access creates a new sequence object by default.

A simple implementation would be:

    def __array__(self, dtype=None, copy=None):
        if copy:
            return self._coords.copy()
        return self._coords

then tested (somewhere):

def test_coords_array():
    coord_seq = point.coords
    assert np.array(coord_seq) is not np.array(coord_seq)
    assert np.array(coord_seq, copy=False) is np.array(coord_seq, copy=False)
    assert np.array(coord_seq, copy=True) is not np.array(coord_seq, copy=True)

Pull Request Test Coverage Report for Build 9087673717

Details

• 6 of 6 (100.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall first build on numpy-copy at 88.138%

Totals
Change from base Build 9076561797:	88.1%
Covered Lines:	2623
Relevant Lines:	2976

💛 - Coveralls

Hmm, yes, it depends on how you interpret the CoordinateSequence object. Is that just a wrapper around an already materialized numpy array self._coords (how it is implemented now in practice) or is it a logical abstraction around the geometry's coord seq (how it was actually also implemented in shapely 1.x)?

Because in the former case, you can indeed say that np.array(geom.coords, copy=False) is fine because getting the .coords object in geom.coords itself has already made a copy, and so the subsequent conversion to numpy does not require a copy.
But in the latter case you could argue that np.array(geom.coords, copy=False) should raise an error because it can never the the geom.coords without involving a copy of the actual values (like, np.array(geom.coords, copy=False) is np.array(geom.coords, copy=False) will never be true).

Now, since this CoordinateSequence class is a separate object from the Geometry objects, it might be that the former interpretation is the only viable one.

From The __array__() method docs "where copy=False should raise an exception if a copy is needed". And the basic dispatch example raises ValueError for this case. This PR is now updated, renamed as an enhancement.

Any further review/comment @jorisvandenbossche ?