As LLM translation becomes a common way for English-language readers to read non- Western literature, a class of harm emerges for which NLP currently lacks vocabulary: the source author can be stigmatised through accumulated translation choices that are not detected by standard MT metrics. Humanities scholarship documents this process for Chinese danmei author Mo Xiang Tong Xiu (MXTX), whose novels and fan-translated public statements circulate widely in Anglophone fandom. Under China’s censorship of queer content, MXTX writes pseudonymously and hedges her public statements (Hu et al., 2024). Anglophone anti-fans read this caution as suspicious evasion rather than a survival strategy (Ng, 2024; Springman, 2024). We ask whether LLM-based translations reproduce the translation choices underlying this stigmatisation, operationalising three measurable hazards: voice collapse, connotation flattening, and hedging erasure. We introduce DANMEI-EPV, a 412- item benchmark double-annotated by professional literary translators, and evaluate eight LLMs in four prompting conditions against a 2,000-sentence GuoFeng-Webnovel reference. Every model exhibits all hazards at non-trivial rates, 5 to 6× higher than the GuoFeng reference. Further, newer high-capacity models optimised for fluent English flatten connotation more than smaller older models, suggesting that optimising for English fluency can trade off against source rhetoricity (Spivak, 1993). Finally, a lightweight inference-time intervention, “rhetorical-silence prompting”, reduces voice collapse by 27.4% and connotation flattening by 19.6% without retraining.