Ultrafast in Codex: incident analysis at OpenAI takes 15 minutes instead of an hour or two

The speedup comes from the model itself, not the search harness: in Ultrafast mode, Codex generates tokens up to 8 times faster than usual.