v1.12.0

• AI now runs on gpt-oss-120b via OpenRouter, routed to the fastest host: explanations in ~0.2–0.6 s (was ~1 s), word checks ~1.2 s, no per-account request cap. gpt-6-luna at OpenAI takes over automatically if OpenRouter fails.
• Fixed: the word check could turn the secret word itself into the imposter's clue when the writer left the clue empty.
• Fixed: explanations sometimes came back in the wrong language.