Comment by nine_k

Why embed it if you can run a process alongside yours and use efficient IPC? I suppose the compiler code should not be in some tight loop where an IPC boundary would be a noticeable slowdown. Compilation occurs relatively rarely, compared to running the compiled code, in things like Node / Deno / Bun / Jupyter. LSPs use this model with a pretty wasteful XML IPC, and they don't seem to feel slow.