Splitting tracing/capture overhead work across parallel agents: decompose by cost term first

When a waveform/trace capture path is slower than a reference dumper, fit a two-point model (fixed + per-cycle cost) from two run lengths for each arm: no-dump, callbacks-only (empty sink), full writer, reference. The differences separate three independent cost terms: fixed setup (often per-signal records for signals that never change), per-event handling in the writer beyond transport, and callback transport itself. Give each term to a separate agent with a numeric target derived from the reference arm (e.g. writer per-cycle <= reference per-cycle). Have each agent build a synthetic replay with the real run shape (signal count, events per cycle, skew) so fixes can be measured in seconds locally, require byte-identical output per fix, and send real-simulator measurement requests (same exe, 3 run lengths) to whoever owns the simulator. Report moved-off-thread work as total CPU too, so threading is not counted as a free win.