The addition of the "asymmetric" search (in the new section 8.2) exhibits in fact another need: users actually will want to have variable strength to match various parts of the search string. For now, the addition seems to assume that *only* the collation elements that have the lowest collation weights at a given level lower than or equal to the global search strength, are supposed to be meaning that they would match all the other collation elements with the same collation weights.
What is really needed is to create regular expressions where for example some parts would need to match with strength=third when other parts would match more loosely with a lower strength. Given the way the DUCET is built, as well as all standard tailorings for languages in the CLDR, it means that (for example with a search strength level 3) all lowercase letters in the search string would necessarily match loosely independantly of case, and all uppercase letters would need to match more exactly. Doesn't it mean that regular expressions could specify this strength explicitly? For example, searching for "résumé" matches all case combinations (and all combinations of accents or diacritics, except on "é"), but isn't there also the need to match the other letters more exactly, while also preserving the loose case matching ? We could say for example to search the initial "r" with loose case matching, and loose diacritics matching, then "é" with loose case matching but exact matching of the accent, and so on. Parts of the regexp would then specify the strength at which they will match. And it will be possible to have stricter matches for some or all lowercase letters, and loose case matching only for uppercase letters. The "asymmetric search" described in this update just describes an oversimplification of a more complete problem, and supposes that users of regular expressions know exactly which character is "marked" or "unmarked" at some level, something that is absolutely not obvious when you have very large collation tables (it is not evident for example that lowercase letters in search regexps mean loose matching and uppercase letters mean strict matching of case *and* diacritics). -- Philippe.
