Feature: Support non-English annotations in AI insights

[texasbe2trill]

Summary

The current insight pipeline and NLP heuristics are optimized for English text. Users reading in other languages may get lower-quality or missing insights. This issue tracks improving multi-language support.

Areas to Investigate

• Audit services/insights.py for English-specific assumptions (e.g., keyword lists, regex patterns)
• Test the embedding model (sentence-transformers) with non-English text — some models are multilingual by default
• Ensure the chat system prompt in services/chat.py works with mixed-language libraries
• Add sample non-English annotations to test fixtures
• Document supported languages in the README

Considerations

• all-MiniLM-L6-v2 (the default model) has limited multilingual support — paraphrase-multilingual-MiniLM-L12-v2 may be a better default
• Insight classification heuristics may need language-aware variants
• This could be scoped incrementally — start with detection and graceful fallback, then add full support