Faithful replica of _plan_cp_load_back_owner_lane_evictions' inner loop with the real
_cp_load_back_node_owner_page_counts bodies, over CUDA value tensors at the b300 hang's
scale (candidates~2000, deficit 7877, ~36 owner pages/node). Measured on a V100:
A. ORIGINAL (.item() x cp, no memo) 49.307s victims=219
B. FIXED (bincount + per-plan memo) 0.469s victims=219 (105x)
C. REDESIGN (counts-once + sorted pass) 0.146s victims=219 (338x)
selection identical: A==B, A==C
Confirms the committed memo+bincount fix (4e9b4d05c0) is ~105x and byte-identical; C is
the idealized single-owner ceiling (the real heap redesign must also handle leaf-up
eligibility + ancestor-unlock + multi-owner).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
leavelet
2e15cb257a