chore(rlm): ReflectionResult records no provenance — no scoring model id, no evidence refs #16114

Parent: #16108

Gap

ReflectionResult (autobot-backend/rlm/types.py:37-63) carries critique and refinement_hint
but records neither what evidence produced them nor which model scored the response. A bad
refinement cannot be traced to its cause after the fact, and to_dict (:58-63) carries the same
blind spot into anything that persists it.

Proposal

Record the scoring model id and the evidence keys on ReflectionResult and its to_dict.

Acceptance criteria

• ReflectionResult carries the scoring model id and evidence references
• to_dict round-trips both
• Existing consumers (chat_workflow/graph.py, knowledge/pipeline/cognifiers/recursive_summarizer.py,
  eval/runner.py, advanced_rag_optimizer.py) still construct valid results

Trivial effort. Filed separately because it makes the other reflection children debuggable.