fix: auto-recall passes embeddings model to backend (0.7.2)

[blackaxgit]

Bug report validated and fixed. Auto-recall (UserPromptSubmit hook) called ollama.embed(query, None) at recall.rs:157, which Ollama tolerates (falls back to its own configured model) but Azure rejects with DeploymentNotFound. Result: silently dark semantic recall on Azure-routed embeddings.

Fix

• Add embedding_model: Option<String> field + with_embedding_model(...) builder on RecallEngine.
• Forward it to embed() from try_semantic.
• Both production callers pass it: clx-hook from config.capability_route(Capability::Embeddings).model; clx-mcp from its existing embed_model field.

Tests

• embedding_model_builder_persists_value — locks the contract.
• embedding_model_default_is_none_for_back_compat — Ollama callers without explicit model still work.

Live-verified

• Triggered UserPromptSubmit with full-Azure config; grep 'Recall semantic embedding failed' ~/.clx/logs/clx.log returns nothing after the fix.

Affects

Anyone routing llm.embeddings to Azure (or any future backend without a baked-in default model). Hybrid configs (Ollama embeddings) are unaffected — that's why nobody noticed in 0.7.0.