feat: babel-memory integration — multilingual FTS & prompts

[AliceLJY]

Summary

Integrate babel-memory as an optional dependency to enhance UltraMemory's multilingual capabilities:

• BM25 search pre-tokenization: CJK queries are segmented before FTS (jieba for Chinese, kuromoji for Japanese), dramatically improving non-English recall
• Language-aware KG extraction: Uses bilingual prompts (EN/CJK) based on auto-detected text language
• Language-aware session distillation: Session summary prompts adapt to conversation language
• Language metadata on ingest: Detected language stored in memory metadata for downstream use

Design

All integration goes through a single language-hook.ts module that auto-detects babel-memory at startup. Zero breaking changes — when babel-memory is not installed, all functions gracefully degrade to current behavior (English passthrough).

Files changed

File	Change
language-hook.ts	New — core integration module with auto-detection
babel-memory.d.ts	New — type stub for optional dependency
store.ts	Pre-tokenize BM25 queries via tokenizeQuery()
kg-extractor.ts	Language-aware prompt selection via getLocalizedKgPrompt()
session-distiller.ts	Language-aware prompt selection via getLocalizedSessionPrompt()
ingestion-pipeline.ts	Language detection via detectLanguage(), stored in metadata
index.ts	Export language-hook public API
package.json	Add babel-memory as optional dependency

babel-memory capabilities

• detectLanguage() — 8 script systems (zh/ja/ko/th/ar/hi/ru/en)
• tokenizeForFts() — BM25 pre-tokenization for 27+ languages
• getKgPrompt() / getSessionPrompt() — bilingual LLM prompt templates

Test plan

• 18 tests covering both with/without babel-memory paths
• TypeScript typecheck passes (no new errors)
• Graceful degradation verified: all functions return safe defaults without babel-memory
• CJK language detection, tokenization, and prompt routing verified with babel-memory installed

🤖 Generated with Claude Code