v1.21.0 - graph-retrieval stream into RRF (recall --graph-stream)

Added

• Graph-retrieval RRF stream. hippo recall --graph-stream fuses a third
  retrieval stream into the RRF ranking (beside BM25 and dense), re-ranking the
  in-pool candidates by graph proximity to the strong lexical seeds: a lexically-weak
  memory that is graph-adjacent to a strong hit gets lifted. Opt-in and default-off;
  the library exposes it as a graphStream option on hybridSearch (rrf scoring
  mode). Tune with --graph-hops (1..3, default 2) and --graph-seeds (anchor count,
  default 10). Read-only over the consolidated E3 graph; reuses the E3.2 traversal.
  Distinct from --hops (E3.2), which injects out-of-pool neighbours; this re-ranks
  within the candidate pool.

Notes

• The stream lives in the rrf fusion path, so it needs embeddings (it stays inert
  until hippo embed has run) and anchors on the top-seeds lexical hits, so on a
  pool with fewer than seeds candidates it degrades to the 2-list fusion. Local
  store only in the CLI for now.
• Validated by a pre-registered hippo-native mechanism ablation: a graph-adjacent,
  lexically-weak answer moves from rank 8/8 to 4/8 (into top-5) with no harm on
  controls. This is a mechanism result, not a population R@5 claim; the
  LongMemEval-oracle population ablation is deferred as L1-eval (it needs a hippo
  entity graph built over the LME corpus). See docs/evals/2026-06-02-l1-graph-stream-*.