As Unicode standard describes (e.g. the same 5.18 section mentioned above) case mapping depends on locale, so lowercasing the same string may have different results on different hosts, and so also the truthfulness of lowercase(x)==lowercase(y) is not universal and depends on the host locale.
Indeed, a fundamental to the problem is that most unicode text doesn't actually carry the relevant locale information... (Of course, one probably wouldn't want to rely on sender-specified locales for email adresses when deciding address equality -- that would open one up to all sorts of potential weird scenarios, i.e. a nightmare for security).
Presumably the case-insensitive version is also doing unicode normalization as well, which is what a byte-level comparison of tolower versions would miss
Or, more likely, by using a library like ICU.