Below The Crossover

You are sorting batches. Production batches hold 20 to 60 items and always will, because the upstream API caps a page at 64. At $n = 48$ you measure insertion sort at 1.8 microseconds and your merge sort at 6.4 microseconds. What do you do?

Discard the measurement. Asymptotic analysis deliberately ignores constant factors, so a 3.5x difference at a single input size carries no information.

Insertion sort below a measured cutoff, merge sort above it, with the cutoff and the measurement written down. This is what production sorts do: CPython's own list.sort never reaches its merge machinery below MAX_MINRUN, which is 64, and binary-insertion-sorts the whole list instead.

Ship merge sort. $O(n \log n)$ beats $O(n^2)$ , and that is what the notation means.

Ship insertion sort on its own. The batches are small, so asymptotics never enter the picture.