fix(instructions): add CJK context detection for Chinese/Japanese/Korean conversations

[lg320531124]

Problem

Ponytail's style rules are designed for English-only conversations. In CJK (Chinese/Japanese/Korean) contexts:

1. "Drop articles" is misleading — Chinese has no articles, so the instruction can cause the model to strip English articles embedded in Chinese text (e.g. "a LLM 模型" → "LLM 模型"), losing semantic specificity
2. "Short synonyms" only works in English — CJK characters are already compact
3. No CJK brevity guidance — Chinese users get English examples that don't map to their language patterns

Fix

Add CJK detection to getPonytailInstructions(). When CJK characters are detected in user input, append a CJK Context section:

## CJK Context
- Chinese has no articles — do not strip English articles in mixed text
- "Short synonyms" rule applies to English only; CJK characters are already compact
- Output brevity: 简短回答，不用敬语，不用客套话
- Code rules (YAGNI, stdlib first, shortest diff) are language-agnostic — apply fully

This mirrors the approach in caveman PR #576 which adds CJK detection to the compression engine.

The inputText parameter to getPonytailInstructions() is optional — when not provided, behavior is unchanged (backward compatible).

Fixes #333

[lg320531124]

Gentle ping — open ~2 days. Small, focused fix (+22/-6): adds CJK context detection so Chinese/Japanese/Korean conversations aren't mangled by compression heuristics tuned for Latin text. No CI gating; ready when you have a moment.