Lebesgue measure is a generalization of “segment length”. For example, the Lebesgue measure of the interval [2,28] is 28-2 = 26, which is just what you should expect, since that line segment is 26 units long. Since it’s a generalization of segment length, there are some pretty bizarre sets that you can also assign a Lebesgue measure to, but the basic idea is still segment length. Also, as ultrafilter, there are even some really bizarre sets that can’t be assigned Lebesgue measure (assuming the axiom of choice).
Anyway, while the definition of Lebesgue measure does involve limits, the actual Lebesgue measure of a set cannot “approach zero” or anything like that, the Lebesgue measure of a set is just some real number (or infinity), either zero or not.
It makes a lot of intuitive sense that the Lebesgue measure of a single point is just zero, period, since the “length” of a single point is just zero. That is why the with the probability function I gave above, the probability of picking a single point is just zero; not approaching zero, but just zero. And the probability of not picking a particular point is exactly one.
Also, speaking more generally here, it’s important to remember the context of what we’re talking about. On one level, probability is just a certain kind of mathematical function satisfying certain properties. There’s absolutely nothing in the axioms that say a probability of zero means it can’t happen, or a probability of 1 means it’s bound to happen.
What those probabilities mean and how to interpret them depends on the actual model you’re using the probability function for. Earlier I gave the example of an infinite sample space in which probability one does not mean it necessarily happens. What about a probability function with a finite sample space?
Building on yojimbguy’s comment, I see nothing wrong with the following probability function for rolling a single six sided die:
The sample space will be: {1, 2, 3, 4, 5, 6, something else happens}.
I’ll choose to define the probabilities of 1,2,…,6 as being 1/6 each, and the probability of “something else happens” as being zero (and extend it additively to assign the probability of other subsets of the sample space). I’ll set up my model that way because I’m only interested in the behavior of the die working the way it’s intended to. I don’t care about the times the die lands on its edge, or, if the die may be particularly fragile, cracks into two pieces when it lands. Either of these events can certainly happen, however; just because I assigned them probability zero doesn’t prohibit them from happening, it’s just the fact that I’m not interested in them.
So here I have a perfectly valid probability model; it may be a little unconventional, but it certainly satisfies all the probability axioms. It’s even an “accurate” probability model, in that it describes the probabilities of a fair die (working the way it’s intended to). However, the way I’ve chosen to assign the probabilities demonstrates that probability one does not mean “guaranteed to happen,” e.g., the probability that a 1, 2, 3, 4, 5, or 6 is rolled is one, but that doesn’t exclude the possibility that it may land on its edge, or break when it hits the table.
In other words, there doesn’t seem to be any mathematically objective part of probability theory defining what probability one or probability zero means; it seems to be all about how you set up your model and how you interpret it.