From Guess2Graph: When and How Can Unreliable Experts Safely Boost Causal Discovery in Finite Samples?