The various race conditions that would appear in a multithreaded context would be a nightmare to debug because they would likely only occur sporadically and would be hard to reproduce. I would love to be wrong about this, but I see it as a major obstacle that will likely need to be sandboxed (as was done with lexical scoping).
For an example of how hard these kinds of issues could be to debug look no further than the incredible difficult that the undo-tree author had in debugging a race condition with the garbage collector. It took years.
In the context of this piece, having a buffer mutex makes certain use cases for multi-threading impossible, for example dividing the buffer up into chunk equal to the number of cores and running a regexp over each chunk.