[antonsoo]

Vec2d.convert_to_basis currently computes two independent projections. These give basis coordinates only when the axes are perpendicular. For a common isometric basis:

from pymunk import Vec2d
print(Vec2d(32, 16).convert_to_basis((32, 16), (-32, 16)))
# before: Vec2d(1.0, -0.6)
# after:  Vec2d(1.0, 0.0)

Solve the two-coordinate system using its determinant. Add a nonperpendicular doctest, clarify that the axes must be independent, and cover skewed, reversed, scaled and ordinary perpendicular bases. Dependent axes raise ZeroDivisionError, as zero-length axes already do.

The native picking check projects world positions to screen coordinates and converts them back before Space.point_query_nearest. Across 30 isometric/skew/perpendicular view configurations, 4,278 wrong tile selections and 4,320 missed moving-body selections become zero. An independent NumPy dense solve agrees within 2.9e-15; native trajectories and direct-query controls are unchanged. These are reproducible constructed game scenes, not a claim about an affected released game or extreme/near-singular coordinate ranges.

Fresh Munk2D/CFFI source build, Python 3.12: python -m pymunk.tests runs 296 tests with two existing skips, all others passing (base: 293 tests, two skips). The new regression fails before the change. Mypy passes for vec2d.py; the new tests and changed method pass Black. Whole-file Black reports the same two pre-existing assert-layout differences on both sources.

[viblo]

Thanks!