again, if your merge request contains 4 commits, each single commit must pass the test...
from your v4, Marvin fails with commit #1 and commit #1+2+3, but succeeds with commit #1+2 and with commit #1+2+3+4
your testbot results shows that commit #1+2+3+4 pass
you need to refactor your code so that each commit passes.
this is required when someone looks up for a regression. he/she knows that it worked at some point in the tree, but fails at some point later in the tree. so bisecting is your friend here, and having tests failing in the middle breaks the logic of bisecting