LLMs don't actually semantically parse the error messages. They will generate the most likely sequence resulting from the error message based on their training data, so you're back to the training data argument.

Perhaps the training data about what compiler diagnostics mean is particularly semantically rich training data.