In the following SQL, the second query - the one after the CREATE INDEX -
gets an incorrect answer:
CREATE TABLE t1(a INTEGER PRIMARY KEY, b);
INSERT INTO t1 VALUES(1,'coffee'),(2,'COFFEE'),(3,'stress'),(4,'STRESS');
SELECT '1:', a FROM t1 WHERE substr(b,4)='ess' COLLATE nocase;
CREATE INDEX t1b ON t1(substr(b,4));
SELECT '2:', a FROM t1 WHERE substr(b,4)='ess' COLLATE nocase;
The problem appears to be that the index is being used even though it uses
a different collation.
This problem was discovered during internal analysis and has never been
observed in the wild - probably because nobody has ever though to create
an index on an expression with non-BINARY collation.