Fact Recall, Heuristics or Pure Guesswork? Precise Interpretations of Language Models for Fact Completion