D079 - scripts/repeat.sh: an intermittent fault needs a denominator

Status: decided, before attempting BACKLOG §6c.

The churn crash was to be characterised by bisecting a threshold. The first attempt ran four builds at each of three values, found nothing, and nearly got written up as evidence that the knob does nothing. At a rate near one in seven, four runs miss the fault more often than they find it - twelve runs with zero events is not a result.

So the measurement gets a script rather than a procedure, and the script reports the denominator as loudly as the numerator and refuses to be quiet about a small sample: a run of fewer than twenty with no crash prints a note saying so.

It also names where each short run died, from the try with no res, because "crashed three times in twenty" and "died in the same call three times in twenty" are different findings and only the second is a bug report.

And the first question is whether the fault still exists. Recent sweeps complete cleanly with thread-churn passing at forty cycles, so bisecting a threshold could be characterising something already fixed. Establishing that a phenomenon still reproduces before spending an hour on its shape is cheaper than the alternative, and it is the step the earlier attempt skipped.