feat: replace Ollama with node-llama-cpp for embeddings

[dandaka]

Summary

• Add node-llama-cpp as primary embedding backend (Qwen3-Embedding-0.6B, ~639MB auto-download on first use)
• Route embed(), embedQuery(), embedBatch() through in-process llama.ts with automatic Ollama HTTP fallback
• Add embedQuery() for asymmetric query embedding with Qwen3 instruction prefix
• Update traul search to use embedQuery() for better semantic search quality

Architecture

Three layers: llama.ts (node-llama-cpp singleton, model-specific formatting) → embeddings.ts (public API, routing + Ollama fallback) → callers (search.ts, embed.ts)

Key changes

• src/lib/llama.ts — new: singleton wrapper, lazy model loading, idle unload (5min), Qwen3 formatting
• src/lib/embeddings.ts — rewritten: llama-first routing, Ollama fallback with separate model name
• src/commands/search.ts — uses embedQuery() instead of embed()
• skill.md — updated embedding backend docs

Test plan

• 192 tests pass (0 failures)
• llama.ts formatting helpers (6 tests)
• llama.ts wrapper with mocked node-llama-cpp (8 tests)
• embeddings.ts llama routing + Ollama fallback (8 tests)
• No mock.module cross-file conflicts
• Manual: traul search "test" with model download on first run

🤖 Generated with Claude Code