Estimate an item's size from range reads you already did, before attempting a whole load

A reader with a per-item memory cap decided between 'load the whole item' and 'answer from bounded range reads' using a listing tool's count, which only covered listed items. Unlisted items had no estimate, so each one first attempted a whole load. On a 10 GB dump that took 18.5 s and grew the worker to 515 MB before the cap failed it, and only then fell back to range reads. The fix was nearly free: every range read the UI had already done (drawing a 1% view) returns the change count for its span. Keep the densest count/span per item, project it over the whole range, and route to range reads when the projection exceeds the cap. On the 10 GB dump the first query went from 18.5 s to 31 ms. Seed the range-read round sizing from the same density, so the first round is already sized to the per-round cap.