#921 Embeddings off Ollama: router embed(), memory/semantic/apps SDK repointed, agent_memory re-embedded

EARS SPEC: - The LLM router shall expose an embed operation backed by a hosted provider producing 768-dim vectors. - tools.memory.*, tools.semantic.step and apps SDK compute_embedding shall use the router embed operation. - When the provider switches, every agent_memory row shall be re-embedded from stored content before reads use the new space. - The agent default provider shall not be ollama. Parent epic: #913. Synthesis (technical problems / solution domains / alternatives) recorded as a comment when started.