Group one is the only part of this list with a real window, and it is also the one that has eaten the most of my time.
I built against a preview API in 2023 on exactly this logic, empty field, early mover, surface not crowded. Two things happened. The annotations changed shape twice, which is survivable, and then the discovery rules changed, which was not: the assistant stopped surfacing my functions for the query I had designed the whole product around. There was no deprecation notice because nothing was deprecated. The ranking just moved.
So the window is real but it is not a moat. What you are buying with early work is familiarity with a surface, not a position on it. Build the thing so it still stands up if the assistant never calls it once, and treat every invocation as upside. Idea 3 is the only one in that group that passes that test, because a notebook is still a notebook.