Poor anomaly detection for mean function with nearly all zero values

We have had customer report which shows there may be problems related to supplying all or nearly all zero values to mean function anomaly detection. My initial guess is it could be a numerical issue and possibly also related to our handling of variable count metric statistics, i.e. #1386.

We need to work on reproducing the behaviour first. If it is related to variable count handling a potential solution would be move away from the correction we apply as a a first step towards the full solution proposed in #1386. This would need thorough testing, but may bring more advantages than it costs.