Subject to re-adding the comment. It is quite important that dv > 0 since it is used as a denominator and log is computed on it. So while it will still be > 0, I think it is a worthwhile comment for later reference.

(I might consider micro-optimizing the code to change if abs(meanv) / dv < threshold: to if abs(meanv) < threshold * dv: though, but that is just me...

And possibly add a comment that // 1 is floor. Had to think a bit there...)

This seems OK, but the code is still actually quite confusing to me. I am not actually sure whether the filtering of large_steps actually matters, or if it is just a small optimization for the for step in steps... loop below? i.e. is it OK if we just loop through all steps (backwards) without prefiltering out large steps?
If not it may still be clearer to move the filtering into the loop, i.e.

for step in steps[::-1]:
    if (the_condition_above_without_vectorization): continue # skip large steps
    do_the_usual_thing

Note that the steps array is small anyways so there isn't much to gain from the vectorization.