Capture-cost scaling studies: use >=4 run lengths and a per-design calibrated length

A two-point (fixed + per-cycle) cost fit for tracing/dumping overhead cannot show nonlinearity. For a scaling study, run each arm at >=4 lengths spanning 3+ decades (e.g. 1K, 10K, 100K, 1M plus full length), n=3, and fit a + b*cycles with residuals and a log-log slope. For multi-design comparisons, calibrate a per-design base length C (e.g. the no-dump run costs ~5 CPU-s) and run C/100, C/10, C, 10C, so small and large designs are comparable. Report the length where fixed cost drops below 5% (L = 19a/b), crossover lengths against the reference dumper, and $ per regression. Pin each arm to its own physical cores, record per-thread CPU (main vs workers), and check end-state equality across arms so a broken arm is caught.