Hi @Igor,
Thanks for confirming that SNAP 3.0 fixes the issue on your side.
As to the statistics computation, maybe you did not notice the “Rough statistics” warning in your first image.
Computing the exact histogram of a full image is time consuming, so when you open a product, only approximate statistics are computed. SNAP computes exact statistics only on-request (when running the Statistics or Histogram analysis tool). The exact statistics computed in these tool are then stored in the product and reflected in all tool windows which require statistics input. This includes the colour manipulation window.
QGIS (and GDAL under the hood) are doing more or less the same. This is why you can notice subtle differences in the min/max values. In QGIS you can also ask for generating the actual statistics instead of the approximated ones (in the layer properties, tab Style).
Yet another difference can come from the JPEG2000 implementations you are using. Depending on your installation, GDAL may still use Jasper as the JP2 codec (it shouldn’t, this is unmaintained software with known bugs).
From your numbers we see that the statistics from a Kakadu decompression and an OpenJPEG decompression are the exact same in SNAP 3.0 (min : 1662 - max : 14254). This is all we need.