I'm a bit uneasy about the testcase here -- it assumes that hyphens:auto will yield the breaks "hy-phen-ation", which happens to match what current implementations produce (AFAIK all derived from the original Liang patterns for TeX) but is not the only possible correct result. A fuller set of breakpoints would be "hy-phen-a-tion" (see for example the title of https://www.tug.org/docs/liang/liang-thesis.pdf), but if an implementation found all three of these possible breakpoints, it would fail this test as written.

An implementation should not be penalized if it does a more thorough job of finding breakpoints than the existing patterns.

What should really be tested here is that the breakpoints found with surrounding punctuation are the same as those found without it -- whatever those may be -- not that one specific (and incomplete!) set of breakpoints is found. That'd be trickier to test, though, as it would no longer just involve a fixed reference.

Meanwhile, maybe it'd be good to revise the testcase to use a word where existing implementations more clearly find an exhaustive set of valid hyphenation points? E.g. I believe "ex-am-ple" is fully hyphenated by existing implementations.