A heavy library with hide-owned on drained the old 8-page sequential fill before a rail filled,
leaving rows of 4-7 cards. Now the filtered-fill path pages up to 20 deep in concurrent waves of
5 (TTLCache + TMDB client are thread-safe), stopping as soon as it hits ~20 kept items or TMDB
runs out — so digging deep costs ~4 page-latencies instead of 20. Refactored the per-page filter
into consume(); trending + the no-filter path are unchanged in behaviour.
09a1646a6a