The largest single cause of "code passes my eye-test but fails the grader" on bigcodebench is returning the wrong TYPE. The algorithm produces the right data but wraps it in the wrong object: `plt.scatter()` (a `PathCollection`) instead of `plt.Axes`; `np.array(...)` instead of a Python `list`; a `PathCollection` from `ax.scatter()` instead of `ax` itself.
