Remove all six module-level read-cache globals and their five clear_* functions from the entity domain so every accessor reads fresh from disk. These caches were harmless in one-shot CLI but were silent stale-read bugs under the long-lived convey server: an out-of-process write left the cache serving data that no longer matched disk. Removed caches: - _JOURNAL_ENTITY_CACHE (entities/journal.py) - _RELATIONSHIP_CACHE + _RELATIONSHIP_IDS_CACHE (entities/relationships.py) - _ENTITY_LOADING_CACHE (entities/loading.py) - _OBSERVATION_CACHE + _OBSERVATION_COUNT_CACHE (entities/observations.py) Replaced the cache-warm patterns in the two provably-hot all-facets loops with operation-scoped memos passed by argument (no module globals): - load_all_attached_entities threads a single journal_entities dict into _load_entities_from_relationships, computed once instead of per-facet. - get_journal_entities_data builds an all_relationships memo once and passes it into _build_facet_relationships, replacing the discarded pre-warm loop. Single-entity routes (get_journal_entity, voice tools) read fresh. merge: dropped _clear_merge_caches; caches_cleared now reports only ["discovery_clusters"] when that on-disk artifact is unlinked, else []. indexer no longer clears domain caches (also removes a latent L6 smell). Tests: removed the autouse cache-clearing fixtures and all clear_* call sites (test isolation is now automatic per tmp/fixture journal); converted the two cache-mechanism tests to freshness assertions; added same-process write->fresh-read coverage for journal entities, relationships, and observations. link/auth.py mtime-reload left untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>