Cache the merged media catalog and trim card rendering

docs/media-scraper.md

@@ -190,6 +190,17 @@ Previews run after every walk:
   of 25 with a 60 s cooldown between them, because bursts of thumbnail fetches trigger Discord and
   provider rate limits. A batch counts rows, not successes.
 
+## Browsing performance
+
+Every grid request used to rescan every scrape database to merge URLs. `MediaService` now keeps a
+short-lived candidate cache: the key is a revision of the scrape metadata (id, state, update time,
+preview count) plus the normalized filter shape — page limits, offsets, and empty values are
+dropped, so every page of one filter set shares a single scan and later pages return in
+milliseconds. The 15 s TTL and the revision key cover new pages, finished scrapes, and freshly
+cached previews. The client only refetches stats when the filters change (or after a minute), and
+grid cards use lazy async decoding with CSS `content-visibility` so a long stream does not pay
+layout and paint for off-screen tiles.
+
 ## Rate limits
 
 Discord and the provider APIs throttle bursts. The engine detects limiting and holds all media