sketch_robust_hll_test.nx

sketch_robust_hll_test.nx -- adversarial-safety smoke for the median-of-k HLL wrapper. Two cases: 1. Honest input: stream N distinct keys, verify median estimate within +-15% of N (same tolerance as basic HLL). Verifies the wrapper doesn't regress accuracy. 2. Adversarial input: stream the same N keys, then explicitly saturate one inner sketch's registers (simulates an attacker who found a hash collision against that seed). Verify the median estimate STILL within 25% of N -- one poisoned sketch out of 5 doesn't move the median. 3. Typed envelope: query returns adv_safety = NX_ADV_ADVERSARIAL (distinguishes from basic HLL which is Honest).