feat: upgrade to qwen3-embed v2.0.0 — tiny local models + offline reranking

[n24q02m]

Summary

Upstream qwen3-embed is releasing v2.0.0 with two major changes that affect this project:

1. Smaller embedding model: tiny-embed-v1 (~60MB ONNX INT8) replaces Qwen3-Embedding-0.6B (573MB) — ~10x smaller, faster inference, multilingual 50+ languages + code support
2. New local reranking: tiny-reranker-v1 (~70-90MB ONNX INT8) — cross-encoder reranker that runs entirely locally via ONNX Runtime

Impact on mnemo-mcp

Embedding (existing)

• Model download size drops from ~573MB to ~60MB
• Memory search quality maintained (target >= 90% of Qwen3-Embedding-8B teacher)
• Output dimension remains 768 — backward compatible with existing memory indices
• First-launch and cold-start times significantly reduced

Reranking (new capability)

• Memory retrieval can now be reranked locally before returning results
• Cross-encoder attention between query and memory content provides better relevance scoring
• Particularly useful for distinguishing between semantically similar but contextually different memories
• Zero network dependency for the entire memory search flow

What needs to change

1. Bump qwen3-embed dependency to >=2.0.0
2. Integrate reranking into memory search (after vector retrieval, before returning results)
3. Test with multilingual memory content and code-related memories
4. Verify embedding quality hasn't regressed
5. Consider if reranking should be opt-in or default (latency trade-off: ~5-10ms per rerank call for 10 documents)

Technical details

Both models share the same backbone (mDeBERTa-v3-base pruned 6L, vocabulary pruned 250K → 64K tokens). Distilled from Qwen3-Embedding-8B (embed) and Qwen3-Reranker-8B (reranker). Apache-2.0 licensed, auto-downloaded from HuggingFace Hub.

Timeline

Blocked on qwen3-embed v2.0.0 release.