The nomic-embed-text space over the corpus is strongly anisotropic (mean
pairwise cosine ~0.55), which compresses cosine-based domain separation almost
to nothing so the design-doc s5 operators (distance/overlap/Wasserstein) cannot
discriminate. Subtracting the global mean of the normalized embeddings restores
isotropy (mean pairwise cosine ~0) and sharpens the operators.
- add GeoMeanCache (engram_geo_mean_build / _maybe_refresh / _vec / _free): a
store-derived centering offset over the embed-eligible set, cached and
refreshed on significant drift; lives in geometry.c, not the store.
- engram_geometry_descriptor gains an optional global_mean: when supplied the
centroid, per-member cosine distance, and co-registration run in centered
space (GM=zeros reproduces the legacy raw path exactly).
- co-registration choice (b): the ANN query stays in raw unit space (index
unchanged) since centering is a rigid translation that ~preserves neighborhood
membership; only the descriptor statistics move to the centered frame.
Covariance/axes/radius are translation-invariant and therefore unchanged.
- test: synthetic ground-truth suite stays green (PERF + ASan/UBSan), plus new
centered/raw/mean-cache assertions.
- add bench_discrimination.c (env-gated, read-only, skips in CI): on a copy of
the real store the two-domain overlap operator drops 1.13 -> 0.008 and
cross-centroid cosine 0.899 -> 0.003 after centering, Euclid distance
unchanged (translation-invariant control).
No change to activation/retrieval behavior; wiring geometry into retrieval is a
separate, behavior-changing cutover.
will.anderson
8cae0f94eb