Build and break an experiment analysis

Everything here is a simulation with numpy and scipy; there are no model calls. The decision that matters most is to analyse every scenario many times, not once: a single 60-day run is one draw, and in this answer the very first no-difference run happens to come out "significant" at p = 0.008. Only the repeated runs show what each check actually does: how often the SRM check catches the fallback bug, how often the outcome test is fooled, and by how much.