kll-percall (lib) reports a nominal 4·k footprint, not asap_sketchlib's allocation #191

kll-percall on the lib library reports a nominal footprint, kll_footprint = 4 · k · sizeof(T) (sketch-bench/src/wrappers/kll/mod.rs:19): 1600 B at k = 50, 6400 B at k = 200, 25600 B at k = 800. It is the same at every item count (approxbench reports 1600 B at 1e3, 1e6 and 1e8 items for k = 50).

It is not what asap_sketchlib::KLL allocates. init_internal (asap_sketchlib 0.3.0, kll.rs:300-326) boxes compute_max_capacity(k, m) item slots once and never grows them, plus a level index of MAX_LEVELS + 1 usizes and a merge buffer of capacity k:

aqpbm-planeval (handle.rs:40-59, 995-1006) and the hydra-kll wrapper (hydra_kll/mod.rs:33) already replicate the library's capacity; the kll-percall lib rows don't.

Impact. The optimizer's cost table takes KLL's mem_bytes_per_instance from these rows (study_saturation.py --phase optimizer-cost), so KLL memory is understated about 3.5× at k = 50. Memory-weighted results that chose KLL (the Fargate weight setting of the AutoSketch-vs-ASAP eval merged in #138) need a new cost table and a rerun. The CPU-only results are unaffected.

Fix. Report the allocation for the lib rows, sharing the replica with hydra-kll. The oxide rows keep a nominal size: sketch_oxide's KLL grows its buffers as it fills.