Ok did a bit more testing on this tonight and it seems like @padding 2 has having a massive impact on the accuracy here. To the point that it seems like there may be a bug or something (or a gigantic hole in my understanding of the @padding modes). I’ll try and isolate it and post a code example tomorrow on a clear head.
Basically with @padding 0 I get good results across the board with the samples. The time-centroid version is better, perceptually-speaking, but the analysis on the full sample still sounds “correct”.
It seems like there’s either junk data being included in @padding 1/2 or the inclusion of zero-added stuff in sinusoidal analysis severely weighs down the averaging further down the road.