Automating compliance analysis against catalogs such as NIST SP 800-53, which defines over a thousand hierarchical security and privacy controls, is attractive for resource-constrained organizations, yet sending sensitive infrastructure documents to cloud large language models is an unacceptable confidentiality risk. The natural alternative – local retrieval-augmented generation (RAG) on commodity hardware – raises a deployment question: given a privacy-mandated local RAG stack for NIST SP 800-53, which platform should an organization choose? Using a controlled, fully local protocol – three open-weight models (Llama 3.1 8B, Qwen 2.5 7B, Gemma 2 9B, all Q4) pinned to identical weights, 20 questions across four difficulty levels under three retrieval conditions, scored with RAGAS and a local judge for 180 measurements per platform per run, repeated three times – we report two findings. First, performance diverges sharply by whether a model fits in VRAM: the NVIDIA RTX 3060 is about 1.8× faster on the 7–8B models that fit, but for the 9B model 8 GB is no longer enough, part of the model spills to the CPU, and throughput collapses to 9.1 tok/s – nearly three times slower than on Apple Silicon (M1 Pro, 16 GB unified memory), which runs all three models at ~24–26 tok/s. Second, answer quality is formally equivalent across platforms under all conditions (TOST with pre-specified margin Δ = 0.05; p_TOST < 0.001 in every condition), so the platform decision reduces to a speed-versus-capacity trade-off rather than an accuracy one. The proposed hybrid retriever (regex identifier matching, metadata hard-filtering, and semantic search) with context injection yields a large, significant gain precisely on enhancement-level questions (level 4): ragas_avg +0.226 on Apple and +0.264 on NVIDIA, both p < 0.001, while leaving easier levels statistically unchanged.