Drop pose-variant absolute coordinates from native-3d descriptors

[roshan2004]

Closes the descriptors review. The native-3d preset emitted six absolute-coordinate columns — centroid_x/y/z and center_of_mass_x/y/z — recorded in the input file's coordinate frame. They are translation- and rotation-variant: two identical molecules placed differently in their files get different feature values, letting a model learn the arbitrary input frame. In a table whose stated purpose is fixed-size ML features, that is a footgun. Size and shape are already captured frame-independently by radius_of_gyration, dim_x/y/z, principal_moments, and the gyration-tensor shape scalars.

Changes

• Remove the six columns from the descriptor emission, the empty-structure path, and the native-3d name list — so preset=None and preset="native-3d" stay consistent. Width drops 80 → 74.
• native-basic and rdkit-basic are unaffected.
• CHANGELOG entry under Changed (Breaking); a regression test pins the columns' absence (named preset, descriptor_feature_names, and the default preset=None output).

Verification

This came out of an empirical review of the shape descriptors. I also checked the related claim that asphericity/acylindricity are redundant with relative_shape_anisotropy — across 23 shape-diverse molecules, leave-one-out R² against the principal-moments block was 0.65 / 0.93 / negative respectively (all below the 0.96 redundancy bar, and mutually uncorrelated), so those were kept. Only the pose-variant coordinates are removed here.

Note

This is a breaking change to the native-3d schema. It is recorded under [Unreleased]; it warrants a 0.15.0 → 0.16.0 bump whenever the next release is cut (left out of this PR).

Full suite: 867 passed, 21 skipped. ruff check clean.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!