Quote from the spec.,

Automatic hyphenation opportunities within a word must be ignored if the word contains a conditional hyphen (­ or U+00AD), in favor of the conditional hyphen(s). However, if, even after breaking at such opportunities, a portion of that word is is still too long to fit on one line, an automatic hyphenation opportunity may be used.

However, I think the idea is to let automatic hyphenations always respect the manual ones, which include both conditional hyphens (­ or U+00AD) and unconditional hyphens (U+2010 or U+002D). So, I'd like to propose that we address this in the specification.