I'm salivating over the context switching wins, myself. I've got a simulation system that uses threads to pause and resume modeled tasks at the appropriate times relative to the simulation clock, and we're just passing a token back and forth with each step of a task. These threads aren't used for parallelism; they're just used to capture the task's continuation in a convenient way, allowing us to use a direct-style approach to modeling tasks.

The context switching sucks so much that an alternate approach, swapping threads for throwing an exception on yield and re-running the method and ignoring side-effects until we get back to the resumption point, saves us a significant amount of time.