It turned out that some code points occur in the default FCE table in denormalized form. As we always normalize given code points to NFD form, we completely ignore denormalized elements of the FCE table. If processing normalized and denormalized forms results in different collation elements, we get wrong collation order in the end.

This issue affects only one test for Japanese tailoring, but it's possible that we simply don't have enough tests to reveal a bigger impact of this problem.

More details in the gist.