Hello
> Does that mean that you could end up with wrong results (which would not
> be acceptable), or that you could end up with false positives that
> later get eliminated by the recheck (which would be fine)?
+ /*
+ * For non-C collations, extract the three bytes from each trigram
+ * and compare them using the collation's comparison function.
+ */
...
+ /* Use collation-aware comparison */
+ result = pg_strncoll(str_a, 3, str_b, 3, locale);
+ PG_RETURN_INT32(result);
For non-C collations, isn't the trigram likely a hash rather than a
proper string, where pg_strncoll won't work properly?