Paul Price, would the ScaleVarianceTask that gets run before detection on coadds have also solved this?

It might make it better, but it relies on empty sky, and there's no guarantee that you have any around the source footprint. But the fundamental problem is that you're losing variance every time you convolve, and if you're not tracking covariance then you have no idea what the true errors are.

I handle this issue in estimate GAaP flux errors, which is a more rigorous variant of `ConvolvedFlux`. Because the convolution kernel is always Gaussian in `ConvolvedFluxPlugin` it is a trivial case of the correction I implement in GAaP plugin. I can adopt my solution here, once I get my GAaP uncertainties reviewed and merged.