This is what I get using your process (and disabling filmic, that is applied later in the pipe on top of the mask – so maybe that is the problem):
(For reference, not disabling filmic gives this:)
The mask does mostly what it is asked : splitting features (sky vs. ground). The overblurring of the dark areas is a side-effect of the guided filter, which is not exposure-invariant, and as such will blur more the lowlights than the highlights (that’s a math problem with the internal variance computation, algorithmic fix is on the way with @rawfiner and needs testing/refining).
Meanwhile, if you want less blurring, hence more details, you will have to tape the mask more closely to small features (using higher feathering value) and perhaps brighten it a bit:
Now, the problem is the sharper the mask is, the more local contrast you may remove from the picture, with contrast compression settings like:
Actually, having a blurry mask is the guaranty that local contrast will be preserved, since the exposure blobs will get the same exposure compensation.
For example, this is the result of the default mask feathering = 10 that gives the above blurry mask:
And now, if I push the feathering all the way to 100:
(here is the corresponding mask for reference:)
So, again, I’m not sure if this is understood : a blurry/blotchy mask is actually better preserving the local contrast of the picture than a sharp/detailed one. Having details on the mask is usually bad news, except if you plan on increasing contrast through the tone EQ, in which case it will also increase the local contrast aka acutance aka perceived sharpness – but you might consider ditching the guided filter (details preservation) at all in this case and use the simple 1D tone curve approach.
Then, having balanced the sky exposure vs the ground one, you can use filmic to restore perceptual contrast:
