Replace the all-synthetic recall story with two real datasets, each
data-gated (fetch script + gitignored data/, clean no-op when absent):

- glove_concepts: top-50K GloVe-6B-50d vectors clustered into 5,000 concept
  boxes (a box = the bounding region of a cluster of related words, the
  box-embedding motivation) via clump. recall@10 92.1% (10x), 99.3% (50x) on
  real high-dim structure, the benchmark that actually stresses the
  center-proxy + region-distance rerank.
- geo_regions: 177 Natural Earth country bounding boxes, [lon, lat] queries.
  recall@3 92.9% over a world grid, and makes the surface-vs-center property
  visible (a South Pacific point resolves to Chile, far from any centroid).

Also bumps the segstore dep to 0.2 in lockstep with the store-cache change
already on main (Arc-keyed incremental per-segment rebuild). The on-disk store
format changed, so a 0.3.x store index is not read by 0.4.0.

Why not trained box embeddings: subsume's trained boxes don't encode WN18RR
hypernymy well (its own notes: hypernym MRR ~0.03; box methods are weak on full
WordNet), so a real trained-box dataset would be undertrained-looking. GloVe
concept boxes give real high-dim structure without that confound.