Per-value-change simulator callbacks: overhead scales with toggle count, so small-subset benchmarks mislead
When capturing waveforms through per-value-change callbacks (VPI cbValueChange-style) instead of the simulator built-in dumper, cost is roughly constant per callback (~0.3-0.4 us each in one measurement). A benchmark on a few hundred signals (~5M callbacks) showed ~+10% runtime; full selection (~430M callbacks) on the same run was ~+490%, vs ~+28% for the built-in dumper. Lessons: (1) report overhead together with the callback count, not just the signal count; (2) always benchmark at full selection before claiming parity; (3) high-toggle nets (clocks, counters) dominate, so sample them per cycle or derive them instead of subscribing per change; (4) separately, registering your own value-change callbacks before the vendor dumper registers its own can crash some simulators at time 0. Test delaying your registration until after the dumper initializes, and verify with a probe where your library is loaded but subscribes nothing.