To respond to the post you posted together, I think your intuition about the scoring is on the right track. It is not possible to get raw scores (with signs) using score_variant because a variant scorer will always be applied to the score output, as you referenced.
I think you can get the raw score with predict_variant or predict_sequence, and then process it however you like. If you find this cumbersome, I think the best way to evaluate whether a variant increases or decreases a splicing event is to look at changes in RNA-seq signals, since splice junctions, sites, and site usage are fundamentally derived from RNA-seq. In that case, use GeneMaskLFCScorer or GeneMaskActiveScorer to get the fold-change of the RNA-seq signal only in exons.
As you pointed out, the variant scorer intrinsically masks exons based on GENCODE V46. You might use GENCODE V49 when visualizing the MANE transcript with the predicted track, as shown here. (after downloading gtf from GENCODE - Human Release 49 and gunzip it)
import pyranges
gtf = pyranges.read_gtf(“gencode.v49.annotation.gtf”, as_df=True, duplicate_attr=True)
I believe you’ll need to implement it yourself to use a custom exon masking and score calculation. So, stick with V46 if it’s not really different from V49.