I ran into this issue on both an old HD4890, as well as on a R9 290, Catalyst 14.12, Win7, so it has been around for some time.
Rendering into a multisampled FBO (tried both RGBA8 and RGBA fp16). GL_SAMPLE_ALPHA_TO_COVERAGE is enabled, the fragment shader writes the coverage value to alpha. After resolving the MSAA'ed FBO, the color channels behave as expected (the expected dithered patterns on the parts with fractional coverage values). The alpha channel is problematic. GL_SAMPLE_ALPHA_TO_ONE is enabled, and that should set the destination alpha to 1 regardless of the coverage alpha supplied by the shader. However it looks like this is completely ignored by the driver. The resulting alpha in the FBO is the coverage value, not 1. The same code works as expected on NVidia hardware (resulting alpha is always 1).
Is this a known issue ?