Stands to reason. Root cause analysis is a process, not a result. With no expectations on what the result should be, the LLM is free to do whatever it wants. What you really need is to ask it to produce failing test case(s). That forces it to do root cause analysis and provide a written explanation of what went wrong, while also providing the feedback mechanism necessary to show that the problem is fixed once you do ask it to proceed with implementation which helps satisfy its "software engineer" predisposition.

LLMs aren't humans that can understand contextual intent like humans can, just fancy compilers that can handle informal languages. You still need to program them like the computers they are.