Judge a cache-tier admission per item against its own baseline cost

A warm-reader tier admitted items whose type was uncertain. It evicted them when a call took longer than a fixed threshold derived from the cold path's per-call floor. Because the warm path doesn't pay that floor, the threshold never fired, and two expensive items made every warm call 2-3x slower for the whole session. One slow call also marked every uncertain item, cheap ones included. What worked: record each item's own cost on the cold path when it is read alone, scale that cost to the request size within a 2x band, and evict only the items the warm call is slower than. Items the warm call beats are marked proven, and only proven ones get the extra routing (point queries). Test with realistic stub costs. An exaggerated cost (600 ms) passed against the broken threshold, while a realistic one (60 ms) exposed the bug.