If you're not shy of a few swear words, Nat's What I Reckon has some great recipes that are straightforward. Here's link to his bolognese - https://www.youtube.com/watch?v=Sw_Ze9zIafM
Agreed, not knowing measure theory never stopped me from computing a conditional expectation. Some courses and books overemphasize rigor in probability and, while it obviously has its place, I've seen newcomers to the field become obsessed with doing everything via measure theory. Further to your point, volume two of Feller is pretty light on measure theory IIRC.
I do think that there's some merit in sticking with probability on discrete spaces for a while. Once you start dealing with continuous spaces, soon you're talking measure theory and you can wade deep into the technical details and miss some understanding of what's going on. I go back and forth on this as I think it's largely down to the reader to figure out what works for them, but I think probability is one of those fields where developing intuition early on is a must if you want to go further.
> Is Figure 8 an unconditional empirical CDF of inter-arrival times?
My understanding is that it's the inter-arrival times after some cleaning and resampling. If I've understood correctly, when they resampled the data, they did so uniformly between the neighbours of the points they omitted, which would actually make the data appear more like an exponential distribution.
> Especially considering its purpose. Maybe it would have been more accurate to say "there's a mixture of two Poissons: the bulk and the network disruption".
Could be. Could also follow a power law or a phase type distribution.
> But this isn't physics. We want to know how useful the approximation is, and whether there is a similarly tractable one with better predictive power.
It's worse, it's math :-) I take your point though, it all comes down to what you're trying to do. If inter-arrival times did follow an exponential distribution with parameter $\lambda$, then we'd have finite variance and I'd be pretty confident that I could build a performant predictive model. The presence of a heavy right tail makes me think otherwise.
It's definitely not a homogenous Poisson process, mainly because of the random changes to mining difficulty and propagation delays. There's a good paper here looking at block arrival times and fitting some different models - https://arxiv.org/pdf/1801.07447.pdf
To the programmer, developer or casual visitor looking at this and wondering whether it's worth the time and effort to dig into this, it is. Most of what's covered here can be understood with undergrad calculus, and will give you a solid basis for understanding and modelling random phenomena you may encounter in your studies, work or hobby.
Fun fact to get you started, Nakamoto suggested in the original Bitcoin paper that blocks would be added to the Bitcoin blockchain according to a homogenous Poisson process (spoiler alert: it's definitely not).