When x must stay between two limits, the best point can be where the slope is zero or at one of the limits. A continuous function on a closed interval always has a highest and a lowest value, and they can only sit at a few candidates: the critical points inside, and the two end points. Evaluate f at each and compare.
\min_{a \le x \le b} f(x) = \min\bigl\{ f(a),\ f(b),\ f(c) : f'(c) = 0,\ a < c < b \bigr\} In words: list the end points and the interior critical points, evaluate f at each, and take the smallest (or the largest, for a maximum).
A box is folded from a 12 cm by 12 cm sheet by cutting a square of side x cm from each corner, so 0 \le x \le 6 and the volume is V(x) = x(12 - 2x)^2 cm^3. Then V'(x) = 12 \left(x - 6\right) \left(x - 2\right), zero at x = 2, 6. The candidates give V(0) = 0, V(2) = 128 and V(6) = 0, so the largest box has x = 2 cm and holds 128 cm^3.
Now an end point wins. On -3 \le x \le 3, f(x) = x^3 - 3x has critical points x = -1, 1 with values 2 and -2, but the end points give f(-3) = -18 and f(3) = 18. The minimum is -18 at x = -3, an end point where the slope f'(-3) = 24 is not zero.
In a model, the limits are bound constraints such as x \le 3. When the optimum sits at a limit, that constraint is active and has a nonzero multiplier, and ProvenDecisions reports it. A small convex example:
minimize (x - 5)^2
subject to
x >= 0
x <= 3
The kernel stops at x = 3, the upper end, and reports the multiplier of x \le 3 as -4, which equals the slope g'(3) = -4: raising the limit from 3 lowers the best value at that rate. The lower limit is slack and its multiplier is 0.