Per my comment in https://phab.mercurial-scm.org/D4342 https://phab.mercurial-scm.org/D4342, I think we should consider recording the timing from here so times aren't skewed due to the presence of pipe I/O. That will introduce some overhead in the traced process. But such is the price to be paid for accurate time reporting, I suppose.

My thinking on that was to avoid the syscall overhead for getting timestamps in the main hg process, which we’ve seen be a huge drag on performance. I couldn’t think of anything cheaper than a single string interpolation and a write to a named pipe, so that’s what we’ve got for now - we can always pursue more in the future if we want, but I want this tracing to be cheap enough that I could plausibly turn it on for 1% of invocations )or whatever) and not have users notice.