Verify the KLL rank-error constant against asap_sketchlib's KLL #944

Problem

kll.rs sizes and certifies KLL with ε = 2.296 / k^0.9723. This is Apache DataSketches' empirical fit for its own KLL. The executor runs asap_sketchlib::Kll, and the constant may not hold for it. sketchlib's own test only checks a looser bound, 2.446 / k^0.9433. At k = 269 that bound is 1.25%, while the Planner certifies 1.00%.

Why it happens

The constant was carried over from DataSketches. sketchlib compacts on a different schedule: level 0 compacts as soon as it is over capacity, and odd-sized levels are halved whole.

Resolution

Measure sketchlib's KLL rank error at the k that kll_k picks, then keep or refit the constant.

Scope and acceptance

• Many seeds, several n and value distributions, and merged states.
• Results recorded next to the constant.