Exercises the pass-2 path end to end — key lookup, decrypt into tmpfs rather
than the on-disk cache, results written into tmpfs and returned as an encrypted
.hrb the server stores without a key. The .hek stays out of the repo.
400 rows / 200 patients, two samples each, so grouped splitting is actually
exercised. Signal is modest and noisy on purpose — a separable toy would score
1.0 with a broken model and prove nothing.
ryan