When a fast-format reader looks slow end to end, time the helper after the decode

A waveform viewer helper read a compressed columnar format with a multi-worker reader. End to end it looked about 4x slower than an older format's single-threaded reader on the same N=1000-signal window (24 s vs 6.6 s). Timing the stages showed decode was only 3.5 s and the workers were done by about 5.5 s. About 18 s went after the workers: per-row worker-to-coordinator messages, building one big reply, and writing about 100 MB of JSON text to stdout. Before concluding a format is slow, put timers on worker done, transfer, assembly and output write. In Node pipelines, send results as one transferable buffer instead of many messages, build the reply in one pass, and write a binary frame or large chunks rather than JSON text; any wire-format change needs the client and its tests changed at the same time. Report the decode-only number separately, so a format comparison isn't really a comparison of delivery paths.