Digital misinformation threatens democratic discourse, especially in low-resource languages like Brazilian Portuguese. We present a fact-checking pipeline comparing structured claim decomposition (Claimify) against full-text extraction across four retrieval-augmented classifiers: gemini-2.5-flash-lite and three local open-weight models (Llama-3.1-8B, Qwen2.5-7B, Gemma-2-9B). Robustness is assessed under characterlevel and synonym perturbations. On a pilot sample, the API-based model shows a preliminary macro-F1 advantage for decomposition, which the three local models fail to replicate consistently. We also observe a systematic bias toward the false label across models.