2025-12-17 – codex semantic memory cleanup

Tightened up tools/semantic_memory.py so it no longer pretends to use Azure embeddings and instead documents the reality: we rely on a local sentence-transformers model (all-MiniLM-L6-v2). Removed the unused Azure credential plumbing and simplified the function signatures, then confirmed stats still works cleanly against the existing index. This keeps the tool honest about its dependencies while preserving the current behavior.