Further check: in both versions:
- create a 200x200 image (8-bit gamma for 2.10)
- set up guides, and make a circle selection,
- bucket fill selection in a new transparent layer, export as InitialSelection-2.xx.png
- Select>None; Alpha-select the layer, and hide it
- Bucket fill selection in a new transparent layer, export as AlphaSelection-2.xx.png
For both versions, compare the initial selection with the alpha selection: load the two PNG as layers, set top to difference mode, new layer from visible and check the histogram, all pixels at 0, so alpha selection is exactly what it is supposed to be.
Now, compare the initial selections of both versions. No need for the histogram, in difference mode there is a visible circle, so the circle selections are not exactly identical.
Comparing the histograms of both versions:
+-----------+-------+-------+
| value | 2.08 | 2.10 |
+-----------+-------+-------+
| 0 | 19856 | 19856 |
+-----------+-------+-------+
| 1<=x<=128| 252 | 104 |
+-----------+-------+-------+
|128<=x<=255| 224 | 372 |
+-----------+-------+-------+
| 255 | 19668 | 19668 |
+-----------+-------+-------+
- Both version create exactly the same number of fully white or black pixels
- They also create the same number of grey/edge pixels
- However the 2.10 edge pixels are significantly more transparent than in 2.8
This extra transparency in 2.10 explains your problem I think. Since I would expect th edge pixels to be statistically centered around 128, I would think that this is a problem with 2.10.