[Bug]: No CJK handling — English style rules conflict with Chinese/mixed-language conversations

## Bug Description Ponytail's style rules are designed for English-only conversations. When used in Chinese/Japanese/Korean (CJK) contexts, several issues arise: 1. **"Drop articles" instruction is misleading in CJK context**: Chinese has no article concept. When ponytail tells the model to drop articles, the model may incorrectly strip English articles embedded in Chinese text (e.g. "a LLM 模型" → "LLM 模型"), losing semantic specificity. 2. **"Short synonyms" rule assumes English**: The instruction "big not extensive, fix not implement a solution for" only works in English. In CJK, shorter synonyms don't exist in the same way — Chinese characters are already compact. 3. **Output style "fragments OK" is language-agnostic but the examples are English-only**: CJK users get English examples that don't map to their language patterns. ## Proposed Fix Add CJK detection (similar to caveman PR [#576](https://github.com/JuliusBrussee/caveman/pull/576)): 1. Detect CJK characters in user input/session context 2. When CJK is detected, modify the per-turn injection to: - Skip article-dropping instructions (Chinese has no articles) - Keep code-focused rules (stdlib first, YAGNI, shortest diff) — these are language-agnostic - Add CJK-specific brevity guidance (e.g. "简短回答，不用敬语") This is the same approach caveman uses: CJK detection → skip English-specific rules → keep language-agnostic rules. ## Related - Caveman PR #576: Same issue fixed for caveman's compression engine - #18, #68: Previous questions about i18n, closed as "works as designed"