Add advanced reranking pipeline: RRF, MMR, graph-distance

[rohitg00]

Problem

Hybrid search uses a fixed weighted sum of BM25 and vector scores. There is no diversity-aware reranking or graph-based relevance boosting.

Proposed Solution

1. Reciprocal Rank Fusion (RRF): Combine BM25 and vector scores using `1/(k + rank)` formula instead of simple weighted sum
2. Maximal Marginal Relevance (MMR): Add diversity penalty to avoid returning near-duplicate results
3. Graph-distance reranking: For queries that match graph nodes, boost results that are closer in the knowledge graph

Impact

Better search relevance, especially for ambiguous queries. Reduces redundancy in search results.

[rohitg00]

Resolved in v0.6.0 (PR #76). Triple-stream search implements RRF (k=60) fusion across BM25 + vector + graph streams. Session diversification (diversifyBySession, max 3 per session) provides MMR-like diversity. Graph-distance scoring via BFS path length in graph-retrieval.ts.