LLMs modify source code.
Comparing LLM-generated code with a deterministic compiler are two completely separate things. Wishing for feature/reliablity pairity isn't the same as provable pairity.
The Comment OP, imho, hits the nail on the head. No offense but exactly comments like yours enable people at the workplace to compare LLM-generated code to be equal to engineered solutions.
There is a gap and I won't aid any LLM to close said gap.
Anyway in such cases, it's the compiler (respectively the LLM) that "completes" the specification of the HLL (respectively natural language) so that the machine code (respectively HLL source code) is produced.
Although I agree that both humans and LLMs maintaining the HLL source code is a mess. And exactly based on the compiler analogy - simultaneously maintaining the same assembler source by hand and with a compiler (from an HLL) would be a mess too, we don't do it for a reason.