sketch_robust_hll_vs_hll_adv_bench.nx

sketch_robust_hll_vs_hll_adv_bench.nx -- adversarial paired bench. CAPABILITY EXCEED: RobustHll's median-of-k defense rejects register-poisoning attacks that destroy a plain HLL. No DS- family incumbent ships this defense natively (see sketch_robust_hll header comment for survey). This bench DEMONSTRATES the capability gap with sealed measurement. SHARED WORKLOAD: 1. Insert N=1000 legitimate distinct keys into both HLL and RobustHll(k=5). 2. Capture baseline estimates -- both should be ~1000 (within HLL's 15% 3-sigma band). 3. ATTACK: an attacker fills every register of one HLL state. For plain HLL: all 256 registers set to rho=30. For RobustHll: only ONE of the 5 inner sketches gets poisoned (mimics the attacker compromising one seed, not all k). 4. Query both estimates after attack. MEASUREMENT: ACCURACY axis: closer to true N=1000 wins. Expected: plain HLL inflates to ~2e11 (alpha·m·2^30); RobustHll's median stays at ~1000 (4 of 5 inners clean). delta_ppm >> 1e8 expected (orders of magnitude gap).