v1.7.0 — Vector-rot defenses, research + chat fixes, eval honesty

Self-healing HNSW vector-index rebuilds, deep-research and chat regression fixes, self-improvement loop closure, and honest eval reporting. Model program: selective thinking, forced math tool-routing, and the bge-m3 embedder upgrade.