Knowledge graph reasoning (KGR) composes multi-hop facts over KGs to support complex knowledge-intensive tasks, providing reliable external factual support for large language models (LLMs) and helping mitigate hallucinations. However, building compact, faithful, and high-quality contexts from large-scale KGs remains challenging. We observe a non-monotonic relationship between retrieval scale and reasoning quality: insufficient retrieval misses critical paths, while excessive retrieval introduces redundant and noisy paths that burden LLMs reasoning. Motivated by this observation, we propose super-path ranking and constraint verification (SPaRC), an efficient framework that constructs, ranks, and verifies super-paths before LLMs reasoning. SPaRC compresses redundant paths into super-paths, ranks them with contrastive learning, verifies top-ranked paths under query semantic constraints, and provides the verified context to the LLMs for answer reasoning. Experiments on WebQSP and CWQ show that SPaRC achieves state-of-the-art performance, improving Hit@1 by up to 12.0 points and F1 by up to 5.9 points on CWQ. On challenging settings covering deeper hop reasoning, multiple topic entities, and multi-answer questions, SPaRC further improves F1 on two benchmarks by up to 17.6% and 42.1%, 26.7% and 5.9%, and 10.4% and 79.4%, respectively, while reducing reasoning tokens by 29.7% and retrieval time by 25%.