Where does one event end and another begin?

Beginnings and ends are set arbitrarily. Causality is continuous and vastly multi-threaded. There are no discrete beginnings and ends in the sense that I think you mean. I wish causality were as simple as many people take it to be. Much of our lives is spent in search of proximate/proximal causes, and that's okay in one sense, because when things go wrong, we need to have an idea what to do differently next time. But..none of that renders the world as simple as our limited observations can detect.

Interesting question!  You really need an interdisciplinary team of neuroscientists, systems theorists, information scientists, linguists, evolutionary psychologists, and others to answer this, but I'll take a preliminary whack at it just for fun. At the basic level, we have the brain systems that control where we focus our attention.  These go back a long ways, to the first oceanic creatures with somewhat complex sensor suites and nervous systems.  If I had to guess, I would say it probably starts at about the point when visual sensors could distinguish different objects doing different things and the organism had to choose what to look at.  When you get that far, there's going to be a premium on developing some kind of circuit that determines when "something interesting starts happening" and also determines when that something is "done."  To be useful, it's going to have to be sensitive to both rate and context.  As the nervous system grows in complexity, so too will the discrimination for these things.  Memory and learning add to it.  If we know from experience that a certain kind of event tends to be a one-shot deal, we can look away as soon as there's a pause.  But if we've learned that another type of event is usually followed by more after short pauses, we may group a series of those things together and think of them as one episodic "event." Humans learn event classification as part of learning language.  People name events differently, depending on whether we care about the sub-events or the whole thing.  Some of this is subject to fundamental sensory circuits and wiring.  Extended periods of immobility and inaction are punctuation points that you'll recognize no matter what your language.  But whether you even perceive shorter, less distinct pauses may depend on your language and mindset.  A change of agent will usually, but not always, be seen as punctuation too.  If A is carrying a bucket and then B carries it, it may be seen as two events, but a bucket brigade will tend to be seen as a continuous process, not a sequence of discrete events UNLESS there are substantial pauses. However, the distinctions that memory, learning, language, and culture introduce can't be completely arbitrary, because a base level sense of causality is hardwired into our nervous systems.  If I see your hand, hear the click of the switch, and see the light come on - in that order, repeatedly - I'm going to assume, as a default, that there is some causal connection unless I have some really persuasive reason to believe a separate, invisible agent happens to be triggering those events in sequence.  (E.g., if someone repeatedly drums four fingers in sequence on a table, I don't assume that the motion of the first finger CAUSES the motion of the second, because prior experience and learning tells me that a person is causing the sequence.)  If a set of cultural assumptions about when events start or end and what the causal relationships are gives us a screwed up sense of causality, it is going to be under heavy negative selective pressure in ordinary human interaction.  We can see that on both a personal and a societal level.  E.g., people who wrongly attribute random events to the benevolence or malevolence of gods, spirits, mystic powers, or conspiracies tend to make bad decisions.  Similarly, people who under- or over-aggregate "events" (like believing in "streaks" in independent events) tend to make costly mistakes. In short, what matters is a functionality test:  does this way of chunking the observable behavior help us build better models of causality than some other way?  Does it improve our data collection, our understanding, and our predictions?  If so, it's a good way to define events.  If not, it's something to work on.

Events are actually perceptual constructions. We sense the changes that are always happening in the world around us, and the brain organizes those detected changes that are learned to be in some way important as discreet events, but there are no true boundaries between events. We detect constant ongoing change, and respond to the parts of the process that have special significance and ignore the rest.