Benchmarking two waveform readers fairly: drive both at the same level

A comparison table that mixes measurement levels can overstate a win by 100x or more. Typical mixes: page-level time (browser + ssh + process start) for one reader against reader-call time for the other; video-frame timing (includes drawing) against code timers (no transfer, no draw); memory from a warm-cache run against a cold one.

Fix: write one small driver per reader that runs the same sequence: open, find the signals, first short-range view, point value, whole-range read, repeat the whole read. Use the same timers, sample RSS over the whole process tree (walk ps pid/ppid, since helpers start workers), and run both cold under the same conditions. Cross-check that both read the same data: total value-change counts should match (within one, if one reader folds the change at t0 into the start value).

Watch 'zoom to whole dump' rows: if a background prefetch already finished, the click looks instant. Report the cold whole read as its own row. Point reads at times past 2^32 are worth testing: a 32-bit time argument silently returns wrong values.

Small gotcha: a helper documented as replying on fd 3 may have a launcher that maps fd 3 to its stdout. If your driver hangs right after the ready line, read stdout.