Agreed. Using keyboard keys to emulate a mouse cursor seems like it ought to be a last resort for graphical applications that lack proper accessibility affordances.
Contrast that with command palettes, accessibility controls, syntax tree navigation, and other approaches that rely on the names, content, and document structure that users already know rather than a special mode that displays two letter codes that must be read each time or memorized. Many of these other approaches also allow users to activate buttons, menu items, and links that are outside the current viewport or hidden in menus which reduces the overall number of "clicks" required to perform those actions. The downside is that they can take longer to type than a two-letter code. Still, my guess is that for most people it would be overall more efficient to optimize for cognitive load than pure speed.
(Though in the long run, I suspect that improvements in eye-tracking will lead to hybrid systems that are both lower cognitive load and faster than any of these.)