Full write up and table of stats on Mediawiki.

• 100% of the 25 non-small language samples (30+ queries) that had tokens affected by folding showed improvement in their zero-results rate, and almost all of small samples did, too.
• In the general sample (unweighted but filtered samples for each language Wikipedia), 84% showed improvements in ZRR, number of results, or top result, indicating that "foreign" diacritics and other variant characters are relatively common (≥ 0.1% of queries) in general search.

Side note on weighted vs unweighted ZRR for @EBernhardson and anyone else interested: The diffs between weighted and unweighted query count and ZRR was often 0 but usually less than 10% (for ZRR that's only a few percentage points.. e.g., 30% vs 33% ZRR). Vietnamese is a wild outlier, with a social media scandal post title repeated more than 12K times (plus four other similar queries repeated over 100 times) leading to 1K unweighted vs 15K weighted queries, and 19% unweighted vs 82% weighted ZRR. Better filtering or bot detection or something is probably helpful in such cases! (These are longish queries, too, but not long enough or junky enough to get caught in the "junky long query" filter I already use.) A few more details are in the writeup on-wiki.