It should be much easier than that. You should should be able to serially test if each edit decodes to a sane PDF structure, reducing the cost similar to how you can crack passwords when the server doesn't use a constant-time memcmp. Are PDFs typically compressed by default? If so that makes it even easier given built-in checksums. But it's just not something you can do by throwing data at existing tools. You'll need to build a testing harness with instrumentation deep in the bowels of the decoders. This kind of work is the polar opposite of what AI code generators or naive scripting can accomplish.

I wonder if you could leverage some of the fuzzing frameworks tools like Jepsen rely on. I’m sure there’s got to be one for PDF generation.

On the contrary, that kind of one-off tooling seems a great fit for AI. Just specify the desired inputs, outputs and behavior as accurately as possible.