sketch_hll_vs_materialized_set_bench.nx

sketch_hll_vs_materialized_set_bench.nx -- STRUCTURAL class comparison. HONEST SCOPE (do not overclaim): This bench proves HLL's memory is N-INDEPENDENT vs a naive O(N) i64 array, which is the algorithmic-class baseline that any "store every input then dedupe" approach must pay. Do NOT read this as "beats Mathematica 100x" -- we never actually run Mathematica here. For real head-to-head numbers see bench_python_vs_substrate_cardinality.sh which actually invokes Python's set() and measures sys.getsizeof. This bench remains useful as a STRUCTURAL gate: HLL's allocation does NOT grow with N, which is the bits-up property that lets the substrate run unboundedly on RAM-constrained devices. SUBSTRATE: HLL lg_k=12 uses 4128 bytes regardless of N. HARD-WIN GATE: At N=100k: HLL memory <= 5 KB Materialized memory >= 500 KB (>=100x larger) HLL cardinality estimate within 5% of truth=100000 Memory ratio >= 100x → structural EXCEED vs the standard incumbent approach. Note: the bench DOES materialize 100k keys in a deduped array to emulate the incumbent storage cost. That's not 'cheating' -- it's the EXACT cost any non-streaming tool must pay.