The proper form of your analogy is that you want to chip away at your code, not at a runtime image. The REPL is the advantage here (in Smalltalk), not the fact that there is an image.
Interlisp did not work like that; editing took place on in-memory lists and the definitions would be serialized into files on request, which were sort of treated like mini databases. Image-based development was used a lot on Symbolics machines, there are advantages to it over always loading from flat text files.
The image just makes the lively computation persistent.
Do authors write books sequentially, first-chapter to last?
Is code used sequentially, only in the same order those procedures are defined in a text-file?
Does the text-file order of procedure definitions help us understand the call sequence?
Because we lack the wits to archive source-code elsewhere?
Save the source-code in files (or some other repository) and when desired build a working image — duplicate the distro image and fileIn your own source-code files (hint: that’s scriptable) — fileOut changes as convenient.
Talk about... I'm about to go Frank Grimes here. Images are bad because you want a visible and complete description of your program, not some image with details squirreled away in nooks and crannies. It's basic engineering to have an exhaustive and explicit description of what your program is doing.
Smalltalk's main shortcoming is that it offers nothing useful that other languages lack. And it lacks stuff other languages offer. Other languages let you look at the code and see that certain things cannot happen, while looking at a small fraction of the code. Smalltalk doesn't offer this. That's a serious deficiency.