v2.17.2

Added

• Hermes llama.cpp embeddings. Explicit embedder: "llamacpp" uses llama-server's local /health and /v1/embeddings endpoints, matching core's existing llama.cpp provider. JSON llamacpp_url / embedding_model options override LTM_LLAMA_CPP_URL / LTM_EMBED_MODEL; defaults are loopback port 8080 and BGE-M3. Stored vectors retain the actual dimension and selected model identity in either embedding layout. No cloud fallback; HTTP failures and malformed, empty, nonfinite or non-float32-storable vectors preserve FTS recall and are never cached.
• Real HTTP-contract regression tests cover configuration wiring, failure recovery, caching and model identity after core's embedding-table migration. Existing provider defaults are unchanged.