Think this :
in step 0 you could decide, in a given area, to commit some errors, all with respect to a given
colour because you know that you will be able to place a single sprite of that colour in that area

If you look at the single steps taken individually, you could tend to convert to SCR2 minimizing the
number of wrong pixels, where maybe committing errors on one single color in a 16x16 area could
lead to better overall results after sprite masking.

Do you see the point?

This is the reason why steps 0-3 should be solved jointly

MSX1 sprites can cover an area of 64x128 = 8192 pixels

maybe assuming some restrictions something could be done
very easily

e.g.
1) no sprite overlap
2) BMP of small size ( 64x128 would be excellent to simplify the algorithm, but also 96 x 128 can lead to easy results )
3) BMP with 16 coulors already matching MSX1 palette
...

Any idea?