I found that a carried-over eviction walk target of 1 makes pages_queued < target_pages / 2 always false, so a walk that queues nothing halves its walk period instead of increasing it and can eventually run on every handle-list round.

I reproduced this on develop cbfc07e3df: with 400 small tables, 23–27 stuck trees took 70–73% of walk passes; rounding the threshold up with (target_pages + 1) / 2 removed those stuck trees, although an eviction-throughput improvement has not been demonstrated.

Fix the backoff calculation and add regression coverage for a one-page remaining target with zero candidates, while preserving the existing behavior when the target is met.