Wow, fantastic testing!
I've been going though your sample files and have tried a few different methods. The best results I've gotten so far is using
mlv_dump_on_steroids (via
Switch) and adding a 2x2 chroma smooth. It looks like you are using
MLVFS and it should work the same there but I didn't test it because I broke it on my system, need to look into that.
I made a tiff from the dng file using dcraw. This is tends to show focus pixels more than Resolve, Adobe Camera Raw and other raw converters. dcraw is also showing some artifacts that can be smoothed out with the other apps but I wanted to show a worse case scenario.

Click on the image to see it full sized, zoom in, download, pixel peep, whatever.
Note that the test pattern that you are using is really challenging the pixel removal. Hard contrasty edges will often show the focus pixels even when running through the map file and that's why I added some chroma smoothing.
Ok--so we know that this worked with your 14-bit file but the problem is with the 12-bit file, right? Where's the 12-bit MLV? From your examples it looks like you found focus pixels that are outside of the area where we know there are focus pixels. Not surprising because there are settings that can show strange patterns all over the image even on cameras that don't normally show focus pixels.
mlv_dump -v
raw_info:
api_version 0x00000001
height 1188
width 1808
pitch 3164
frame_size 0x0039674C
--> bits_per_pixel 14
black_level 2047
white_level 5586
active_area.y1 28
active_area.x1 72
active_area.y2 1185
active_area.x2 1808
exposure_bias 0, 0
cfa_pattern 0x02010100
calibration_ill 1