Rather than doing binary translations, though, wouldn't it be easier to just establish a primitive in the smalltalk VM that constructs, as an atomic operation, a graph of interrelated object references, then hands out the IDs to be populated with data?
Then you make the VM larger. ObjectStudio (the old Enfin) Smalltalk has "primitive classes." Classes largely implemented in the VM. With that Smalltalk, you can build the image from source files. The strange loops are resolved this way.
