We present a systematic study of machine translation for Chakma, an endangered and extremely low-resource Indo-Aryan language. We introduce a large Chakma–Bangla machine translation resource comprising 15,021 parallel translation pairs, 42,783 Chakma monolingual sentences, and a trilingual evaluation benchmark. To address data scarcity and the script mismatch between Chakma and Bangla, we develop a character-level transliteration framework that enables transfer from Bangla and multilingual pretrained models. We evaluate from-scratch machine translation systems, fine-tuned pretrained models, and large language models using in-context learning. Our results show that transliteration is crucial for the tested pretrained models and that performance is strongly direction-dependent: in-context learning performs best for Chakma-to-Bangla translation, while Bangla-to-Chakma remains considerably more challenging, with the strongest approach depending on the evaluation metric.