I was (re-)running fuzz tests again, and in addition to what was already
known (e.g. self-referencing structures) found two new issues:

- long as dict keys (https://github.com/kisielk/og-rek/issues/55), and
- encoding global with '\n' in module name (fixed in previous patch).

Update the corpus with points go-fuzz found, so that next runs could
restart with having all the previous interesting input vectors available.