> I wonder, is there a point where mimicry can fail? Can predators at some point start to eat the mimics?
The same sources of variation which provide the variety for evolution to work on to evolve avoid eating things with this appearance behavior will also provide variation that evolution can work on to evolve back to do eat things with this appearance behavior; the frequency with which eating causes death vs. the degree to which not-eating results in insufficient nutrition will decide which wins.
That's all true, but I should point out that there are finer grains of behavior, and some of them asymmetrically make it easier to learn that "actually it's OK to eat those". Predators could evolve the ability to tell the difference between mimics and originals (which is generally rare but extant; consider how human children are taught to tell the difference between coral snakes and king snakes, or how certain birds can reject eggs that they didn't personally lay, but most can't), or they could evolve the behavior of "don't eat things that look poisonous, unless you're really hungry" (which is not rare, and which doesn't work well in a low-mimic environment, but does work very well in a high-mimic environment).
As you note, the behavior you end up with is determined by how much stress the mimics place on you.