LLM-based renderer (Task 06 option B) with fabrication guard

[tomas-samek]

Summary

Renderer is currently template-only. Task 06 option B is to allow an LLM-based rendering pass on top of retrieved memories (still no fabrication: content comes from memories, LLM just phrases it).

What to do

• Define a strict contract: LLM may reorder/reword retrieved content but must NOT introduce facts absent from the retrieved memories.
• Add a verification layer that rejects LLM output if it contains tokens/claims not traceable to a retrieved memory.
• Gate behind a config flag — template renderer remains the default, LLM renderer is opt-in.

Acceptance

• Opt-in LLM renderer exists with fabrication-detection guard.
• Tests show the guard rejects hallucinated output on a constructed example.

Links

• docs/design/honest_agent/tasks/06_renderer.md
• src/mcp/responder.rs