fix: fail fast when LLM_BACKEND=llamacpp but LlamaCppClient is unconfigured

Previously embed_one() silently fell back to Ollama embeddings,
which would load nomic-embed-text into VRAM alongside llama-swap —
wasting memory on an unintended model. Now returns an error with
an actionable message instead.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/ai/mod.rs

@@ -78,8 +78,9 @@ pub async fn embed_one(
                 .pop()
                 .ok_or_else(|| anyhow::anyhow!("llama-swap returned no embeddings"));
         }
-        log::warn!(
-            "LLM_BACKEND=llamacpp but LlamaCppClient is unconfigured; falling back to Ollama embeddings"
+        anyhow::bail!(
+            "LLM_BACKEND=llamacpp but LlamaCppClient is unconfigured — \
+             set LLAMA_SWAP_URL or switch to LLM_BACKEND=ollama"
         );
     }
     ollama.generate_embedding(text).await