sketch_observability_dashboard_bench.nx

sketch_observability_dashboard_bench.nx -- COMPOSITION integration bench. CLAIM TO VALIDATE: Substrate composes -- multiple sketches running on the SAME stream stay sub-linear in memory AND each track its metric accurately. This integration bench demonstrates the cardinal "composition over configuration": a realistic observability workload built from 4 sketches running side-by-side. WORKLOAD (simulated APM telemetry): For each of 5000 events: event.user_id -> HLL (track unique users) event.event_id -> CountSketch (estimate per-event frequencies, spotting top events) event.latency -> T-Digest (latency percentiles) event.latency -> AMS (F_2 = skew indicator) - user_ids range over 200 distinct (skewed to first 50) - event_ids range over 20 distinct (heavy tail at id=1) - latencies log-uniform in [1, 10000] microseconds MEMORY BUDGET: HLL lg_k=8: 256 + 32 = 288 B CountSketch d=5 w=256: 10288 B (counter grid 10240 + header) T-Digest delta=100: ~20 KB AMS d=5 s=128: 5120 + 80 = 5200 B TOTAL: ~36 KB. vs naive (5000 event records × ~40 B): 200 KB. SUB-LINEAR composition wins on memory. EACH METRIC VERIFIED INDEPENDENTLY: - HLL within 25% of 200 unique users - CountSketch correctly identifies event_id=1 as heaviest - T-Digest p99 within 30% of true p99 latency - AMS F_2 within 50% of true F_2 (skew indicator)