This is why the real-tests should be entirely in pseudo-code and about the logical structure.

Any optimization should happen after a basic algorithm that works and might go along the lines of (telemetry suggests our platform has a bottleneck here) how would you speed things up if there's an issue in this section?