When we introduced collation starred relations (compact syntax), I think we said we wanted to forbid characters that are not NFD-inert, so that there is no ambiguity when rule strings are decomposed or otherwise normalized. If we still have consensus on this, then we should document and enforce it.
For example, fa.txt contains a starred rule with decomposable characters: `<<أٲإٳؤ` In escaped form, it is `<<\u0623\u0672\u0625\u0673\u0624`. U+0623 ARABIC LETTER ALEF WITH HAMZA ABOVE and U+0625 ARABIC LETTER ALEF WITH HAMZA BELOW are composites.
TC agrees with Markus's assessment.
Filed : to add a test for this.
Milestone 25rc deleted