sketch_kll.nx

sketch_kll.nx -- compactor-hierarchy quantile sketch. MRL99-style (Manku-Rajagopalan-Lindsay 1999) "compactor cascade": maintain levels 0, 1, 2, ... each holding up to k items. Items at level h carry weight 2^h. When level h fills, sort it, flip a coin, promote either even-indexed or odd-indexed items to level h+1 (the other half discarded). Karnin-Lang-Liberty 2016 (FOCS) tightens this via geometric per-level capacity decay; we use uniform capacity here for simpler code at slightly looser bound. Rank-error declared conservatively as ~1/sqrt(k) at conf 0.95. Versus Reservoir (sketch_reservoir.nx) HONEST COMPARISON (2026-05-20 audit, sketch_kll_vs_reservoir_bench.nx): - At MATCHED slot count, Reservoir wins on RANK-ERROR magnitude. KLL k=64 declared eps=0.12 vs Reservoir cap=1024 declared eps=0.0425 (both ~8KB). Empirically Reservoir wins on uniform 1..10000 stream too. - KLL's structural advantages live ELSEWHERE: (1) DETERMINISTIC mergeable across instances (Reservoir merge requires algorithmic choices that change semantics) (2) O(1) amortized update at any N (Reservoir's update is O(1/N) probability; quickly stops sampling at large N) (3) Cascade preserves rank info from FULL stream history (Reservoir's sample is a fixed random subset) - Earlier doc claim of "tighter rank-error at fixed memory" was a misframe (compared at same k, not same memory). Corrected. LOSSLESS-LANGUAGE DISCIPLINE (doc 20): Same NX_ENV_RANK_ERROR envelope as Reservoir; substrate doesn't double-count. Multiple primitives sharing an envelope kind is fine -- the discriminator is about the SHAPE of the error, not uniqueness.