A real run under a kept transcript compared each kept measure call by asking
the component again, though step 5 had just asked it the same call. The check
now takes the calls the leaf's run made: a kept call at the same proposal and
room that read no inputs compares with that answer, and only other calls are
asked again. The check reads the transcript's calls in place instead of
cloning them and their inputs per leaf, and a leaf hands its ideal and minimum
calls over in one leaf_call.