feat: eval type distinction — KB quality vs retrieval quality

[rajnavakoti]

Description

The eval pipeline currently runs one flow (retrieval + synthesis) and scores everything as CLEAN/INCOMPLETE/MISSING. But there are two distinct quality dimensions: does the KB contain the knowledge (KB quality), and can the retrieval pipeline find and synthesize it (retrieval quality). Separating these would let users diagnose whether gaps are curation problems or retrieval problems.

Acceptance Criteria

• Eval results include an eval_type field: kb-quality | retrieval-quality
• KB quality evals test whether entities exist and cover the topic (could use direct entity search, not full RAG)
• Retrieval quality evals test the full pipeline (retrieval + synthesis + citation)
• Viewer Evals page shows eval type as a filter/tab
• A question can fail KB quality (knowledge doesn't exist) vs retrieval quality (knowledge exists but wasn't retrieved)

Out of Scope

• Custom eval types beyond these two
• A/B testing different retrieval configurations