Patterns Over Principles: The Fragility of Inductive Reasoning in LLMs under Noisy Observations