Every turn decoded and validated every transcript row, so a session with screenshots held all of them for the whole turn (20 MB for one session). - images live once in a content-addressed `images` table; rows keep the hash and metadata, and loads attach a reader instead of the bytes - request encoders emit placeholders; the transport streams the body and reads each image just before sending it, one at a time, with an exact content-length. Request bytes are unchanged, so prompt caching holds - rolling fingerprints hash images by content; a hash saved before this is checked once in its old form and rewritten, so summaries survive - a one-time startup migration moves existing inline images, after a VACUUM INTO backup under $ALBEDO_HOME/backups - deleting a session releases images nothing else references - the cell journal keeps images in the row shape