In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks from a few examples, making it promising for languages underrepresented in pre-training. Recent work on many-shot ICL suggests that modern LLMs can further benefit from more ICL examples enabled by their long context windows. However, the inference cost can be prohibitive for low-resource language communities, and it is unclear whether example selection can reduce the number of examples needed. In this paper, we present a systematic empirical study of many-shot ICL for machine translation from English into ten truly low-resource languages recently added to FLORES+. We analyze the effects of retrieving more informative examples, using out-of-domain data, and ordering examples. Our findings show that many-shot ICL becomes more effective as the number of examples increases. More importantly, we show that BM25 retrieval substantially improves data efficiency: 50 retrieved examples roughly match 250 many-shot examples, while 250 retrieved examples approach the performance of 1,000 many-shot examples. We further show that example ordering has little effect and that ICL provides additional gains on top of fine-tuning.