#798 Stage 1 epic: structured output through the LLM gateway + Anthropic prompt caching

Agentic plan Stage 1: (a) response_format JSON-schema param through call_llm to all providers (openai_compat response_format, ollama format, anthropic forced-tool), jsonschema validation + one retry in-gateway; (b) run_goal(output_schema) applied on the final-answer call only, agent_result stores raw+parsed; (c) MCP run_goal(output_schema); (d) anthropic cache_control on system+tools. Gates: golden demos + harness incl. new structured_output case.