[cmyui]

Summary

Fixes three bugs in the Catmull curve implementation that caused it to produce incorrect curve shapes compared to both osu-stable and osu-lazer:

• np.roll without axis=0: Flattens the 2D points array before rolling, scrambling x/y coordinates when computing tangents. All Catmull curves with 3+ points had incorrect tangent values.
• Swapped roll directions: np.roll(points, 1) (shift right) was assigned to p_aheads instead of p_behinds, producing negated tangents.
• Wrong segment at t=0: math.ceil(0 * N) - 1 = -1 indexed the last curve segment instead of the first. Also, the global parameter t was used as the Hermite parameter s instead of being remapped to the local segment range [0, 1].
• Wrong endpoint boundary condition: The last control point's tangent was clamped (duplicated), but osu-stable and osu-lazer both extrapolate (v4 = v3 + (v3 - v2)), producing a tangent twice as large.

Verified against osu-stable (SliderOsu.cs + Vector2.CatmullRom) and osu-lazer (PathApproximator.CatmullToPiecewiseLinear + catmullFindPoint), which use the same 4-point formula and boundary logic.

Test plan

• Added slider/tests/test_curve.py with tests that verify against a reference implementation of osu's exact Catmull-Rom formula and boundary handling
• Tests cover 1-point (degenerate), 2-point, 3-point, and 5-point cases
• Each test evaluates multiple t values per segment and compares against osu's 4-point formula
• All 27 tests pass (20 existing + 7 new)

🤖 Generated with Claude Code