That's a good point about context sensitivity and the need for process-level controls needing a top level menu. I would say that part of the friction with window-level menus is moving targets. It's less cognitive effort to find and hit menu targets that are always in the same place on the screen.

Part of the discoverability of menus is learning what actions are modal (titles have the ellipsis) and learning what hotkeys and key chords do what in the app. There's nothing faster than hotkeys. Ideally, users train themselves to use hotkeys to get work done and forego the menu except to discover additional features.