One idea here is
that there is no need for eassert (E) unless there's a genuine doubt
that E will be true (in some cases the removed eassert (E) calls were
ineffective anyway, due to preceding eassume (E) calls). The patch cuts
down the number of integer overflow checks to six in json.c, if I'm
counting correctly, and that should be good enough.
I don't understand why minimizing the number of checks and assertions should be a worthwhile goal. At the very least, the assertions document the assumptions that we make about the values, and as such they are valuable even if they never trigger.