## Why

The RAG refactor needs a repeatable baseline before changing L0, metadata filtering, post-processing, or Spring AI retriever integration. Without fixed retrieval cases and measurable output, later changes can look cleaner architecturally while silently degrading recall or evidence quality.

## What Changes

- Add a retrieval evaluation baseline for RAG queries, separate from full diagnosis evaluation.
- Define golden retrieval cases covering Chat-style knowledge lookup and AIOps-style alert diagnosis retrieval.
- Add a lightweight offline evaluator that compares retrieved candidates against expected documents, breadcrumbs, and evidence keywords.
- Preserve baseline JSON and Markdown reports so future changes can compare retrieval behavior.
- No production retrieval behavior changes in this change.

## Capabilities

### New Capabilities

- `rag-retrieval-evaluation`: Defines fixed retrieval golden cases, deterministic retrieval evaluation, and baseline report preservation.

### Modified Capabilities

- None.

## Impact

- Adds retrieval evaluation fixtures, documentation, and scripts.
- May read existing retrieval/tool trace output or saved fixtures, but does not require live LLM calls.
- Does not change the `lookup_knowledge` runtime behavior, Milvus schema, document upload API, or Agent flow.