I have heard that this is not possible because Racket needs to make a stricgter separation of compile time and run time... but I am not sure whether this is actually true.
But, though I don't know current Racket internals, I'm not sure those need to be a barrier to monkeypatching that approaches CL.
Other than optimizations that make some debugging info and dynamic changes difficult-to-impossible, if the system had a mode to disable these optimizations (or to maintain some information in parallel that permits a cascading unraveling of optimizations affected by a change), I'm not aware of a fundamental reason the system couldn't support more loose poking at things.
Though, one implementation complication that Schemes do have is first-class continuations.