FWIW this looks very similar to a bug recently fixed for radv.
Probably not so coincidentally, radeonsi has workarounds for 2/10/10/10 signed
formats, apparently gcn hw before vega (with the exception of stoney ridge) has
a bug interpreting the 2bit alpha as unsigned. The output from this test would
be consistent with that (alpha 1.0 as 2 bit snorm would be 0x01, whereas that
value interpreted as unorm would be 0.333).
So maybe the same bug could affect later evergreen hw as well.