Insights into LLM Long-Context Failures: When Transformers Know but Don't Tell