WRT 2: Working memory capacity is (according to one well-regarded theoretical model) actually a series of different components - the phonological loop (stores verbal/phonological/textual data, can hold just about the seven digits of a phone number), the visual/spatial sketchpad (can hold representations of images - not sure there's as straightforward of a measure of its capacity), the episodic buffer (not so clear it exists - separate component for integrating different components), and the executive system (controls attention, attending to specific bits of input). Many of these are confirmed theoretically by constructing experiments to overload one memory component, and then checking to see what other capabilities still have spare capacity.
While an individual's capacity in each of these components is pretty fixed, there is some variance in what component of working memory a task exercises.
In addition it might be that some tasks might lend themselves to good compression. For example, an even lightly experienced programmer thinking about a for loop might not require one "slot" of phonological memory each for the index, start, finish, and label, but could rather represent the entire structure as "for loop" in their phonological memory, this single unit acting as a reference to the long-term memory representation of iteration. In this way, while working memory in its raw form is rather constant over different types of tasks, the actual amount of semantic information it contains may vary widely with task domain.
Lastly, mnemonic strategies matter a lot. For example, it might make a big difference to a music learner whether, when sight-reading, they process a piece in their phonlogical loop ("the notes are A-G-G-A") or in their visual-spatial sketchpad ("This bunch of dots is the shape of the tune on the page").