Issue:
For shared disk pages, the disk image is counted once in the connection-level cache total (correct, deduplicated) but is not added to the owning btree->bytes_inmem. This leaves each btree's in-memory byte count far smaller than its real footprint (it's missing the disk image, the bulk of every page).
Eviction apportions per-tree work proportional to btree->bytes_inmem (__wt_btree_bytes_evictable). The undercount makes the dominant trees look nearly empty, so they get a tiny eviction quota (target_pages). Eviction then under-samples exactly the trees holding the data, picks poor candidates, and sheds hot/reused pages—raising the miss rate and latency.

Fix:
Account disk image bytes for all the btrees that are sharing disk pages.