I think you are right - that's a bug. As a consequence, overflowed would
be cleared less often than it is now, which in turn may cause a fatal
abort in case it fails to recover from a stack overflow (i.e. if it
consumes another 50 stack frames, instead of unwinding).

This entire machinery is there to guard against hard-to-analyze crashes
resulting from stack overflows, in particular when a dictionary lookup
happened to cause a stack overflow. The overflow would raise the Python
exception, which would then be cleared in the dictionary lookup, as if
nothing happened.