Idea: Match each algorithm's best-case behaviour to a nearly sorted input.
Insertion sort has a best-case cost that appears exactly when the data is already close to order:
\[T_{best} = O(n)\]
Each key needs only one or two comparisons before it settles, so the total work is nearly linear. By contrast, selection sort forces $n$ full scans no matter the arrangement, staying $O(n^2)$; bubble sort still swaps and passes repeatedly; and quick sort's pivoting gains nothing from pre-sortedness.
The one algorithm whose cost collapses toward linear on nearly sorted data is insertion sort.
\[\boxed{\text{Insertion sort}}\]