Here's something to throw into the pot. If, as seems likely, we have to come up with a spelled-out list of guidelines for stuff not to say, why not distinguish between two grades of such things? Namely, things that would offend many/most on Stack Exchange, vs. things that did offend at least one person. (That might actually be about all we need to spell out, in fact.)

If this distinction is made, both sorts of messages should generally be removed when the flag is validated, but the consequences for the flaggee need not be identical. Someone who persists in targeting easily-offended users and saying borderline things should suffer the usual consequences (escalating suspensions), but someone who accidentally offends in unusual ways a few times need not be suspended, just warned that they offended someone.

This requires a judgement call to distinguish the two, but I don't think that's either avoidable or impractical; the only real challenge would seem to be keeping good track of a particular user's behavior, but just logging the number of dubious-flag-triggered warnings (and comparing their flaggers, for diamonds) would go far toward enabling that.

And, of course, there are still some things that have no reason to be offending anyone; a flagger shouldn't be blindly heeded if they're taking offense at ridiculously harmless stuff.

TL/DR: There's a difference between taking offense at something, taking offense at "nothing", and taking offense at *nothing at all*. Let's be sure to keep those differences in mind.