Worker memory grows until killed

A Python (Gunicorn + Flask) recommendation API has workers that grow from 200MB to 2GB RSS over ~6 hours and get killed by the orchestrator, dropping in-flight requests. Each worker handles all routes. Dashboards show RSS climbing monotonically; tracemalloc top-stats taken an hour apart attribute the growth to a module-level dict used to memoize per-user feature vectors with a hand-rolled @cache decorator that has no eviction. Request volume is steady and the set of active users is bounded, but user IDs include a high-cardinality experiment bucket suffix appended last week. Triage and fix.