Use box midpoint to choose finite-difference direction in jacobian_fd

[Nas01010101]

Problem

In minimize.py, jacobian_fd chooses each coordinate's finite-difference direction so the perturbation steps away from the nearer bound. It compared x against

mid = 0.5 * (bounds[1] - bounds[0])
eps_vec = np.where(x > mid, -eps, eps)

but 0.5 * (bounds[1] - bounds[0]) is half the box width, not the box midpoint. For bounds that are not centered on the origin this picks the wrong direction. Concretely, with bounds = [10, 20] the comparison value is 5.0 (true midpoint is 15.0), so a point at the lower bound x = 10 satisfies x > 5 and steps −eps, i.e. to 9.9999998 — outside the box.

Fix

Compare against the actual midpoint:

mid = 0.5 * (bounds[0] + bounds[1])

At the lower bound the step is now +eps (inward); at the upper bound −eps (inward).

Tests

Adds test_jacobian_fd_respects_off_center_bounds, which records every point jacobian_fd evaluates with x at the lower bound of the off-center box [10, 20] and asserts they all stay within [lo, hi]. It fails before this change (evaluates 9.9999998) and passes after; the rest of minimize_test.py remains green.