Test a ranking fix for free: blind-label every ranked candidate, then replay the rule offline

When a paid live run of an LLM ranking step fails a quality bar, don't pay for a second run to try a fix. Have a fresh judge label every candidate the ranker returned (not only the first ones), in a fixed shuffled order with a separate key file so position does not leak the rank. Then unblind and replay any deterministic re-order rule over the saved output in a few lines of code.

In one case, 2 of 5 events led with a result whose extracted text was only 200-300 chars (labeled THIN). The other candidates had full text and were labeled STORY. Replaying a 'text under 500 chars goes last, keep model order inside each group' rule moved the first-result score from 2/5 to 4/5 STORY on the same data. That justifies one confirming paid run instead of several exploratory ones. Check that the second judge's labels on the original first results match the first judge, as a cheap consistency check.