Replace placeholder embeddings with real vector search

[BunsDev]

Problem

lib/opentrust/semantic.ts currently uses a simple character-code hashing function to generate "embeddings," which provides no meaningful semantic similarity. The sqlite-vec dependency is installed but underutilized because the vectors it stores aren't semantically meaningful.

Proposed Solution

Integrate a real embedding model for semantic search:

Option A: Local ONNX Model (Preferred)

• Use a small model like all-MiniLM-L6-v2 via onnxruntime-node
• ~80MB model file, runs locally with no API dependency
• Maintains the local-first philosophy

Option B: API-Based Embeddings

• Use OpenAI, Cohere, or Anthropic embeddings API
• Smaller bundle, but adds external dependency and latency
• Requires API key configuration

Acceptance Criteria

• embedText() produces semantically meaningful vectors
• Similar content produces similar vectors (verified with test cases)
• searchSemanticFallback() returns relevant results ranked by similarity
• Embedding dimension matches sqlite-vec column configuration
• scripts/index-semantic.ts rebuilds the index with real embeddings
• Performance: embedding a chunk takes < 100ms locally
• Documented setup instructions for the chosen approach

🤖 Generated with Claude Code