Beneficial and neutral mutations are essentially left out of the equation - they would spread to 100% of the population, so in the steady state, the probability of mutation newly creating a beneficial mutation has to be 0% (since it's already present in every member).
The motivation for ignoring beneficial mutations (and back-mutations to beneficial states) is that they're extremely rare as compared to deleterious ones - most selection pressure in nature is aimed at merely preserving the functionality in the genome, weeding out new deleterious mutations rather than supporting new beneficial ones (though that is a critical role in the very long term, of course).
I still think that your definition of beneficial and deleterious is defined post-hoc as "survives/does not survive long enough to reach the stable state."
In my example earlier regarding extinct bird species on New Zealand, were there ever any beneficial mutations? After all, the genes no longer reproduce.
My point is that there is no stable state, so it's better to have a shorter-term definition of "beneficial" and "deleterious" based on relative reproduction fitness compared to others in the species population over a short time frame.
Additionally, beneficial and neutral mutations do not always spread to 100% of the species population. A Y-linked trait won't spread without a male lineage.
I still think that your definition of beneficial and deleterious is defined post-hoc as "survives/does not survive long enough to reach the stable state."
Yup, that's a pretty good description of the assumption/definition, for all the good and bad things it brings with it; I mostly agree with the rest of what you've said.
FWIW, I've mainly seen the one-mutation-one-death rule applied to arguments that attempted to place informational speed or capacity limits on the process of evolution, and in most cases it has turned out that these arguments fail when applied to the real world because of precisely the types of arguments that you've made against this rule (sexual recombination and the fact that evolution is never in a steady state tend to be the biggest problems).
An aside on the topic of defining beneficial/deleterious mutations: the problem is a very difficult one, because in reality the effect of a mutation is at best distributional and not measurable along a single axis of goodness/badness. I don't know much about the state of the art here, but if I was going to sit down and try to figure out a way to try to estimate "benificial-ness" of a mutation (even as a distribution), I'd say you've put your finger on exactly the right question to focus on, that there's a tension between what might be immediately beneficial and what would be beneficial in the longer term. For instance, a gene that made someone have babies like crazy by always sacrificing one's grandchildren for nourishment would be fantastic for 1-generation-out fitness, but terrible for 2-generations-out. Similarly, in an environment where pre-reproductive death was very common, an adaptation that decreased the likelihood to have children by a small factor but significantly increased the ability to keep them alive would be very beneficial at 2-generations but deleterious at 1. So pure-local effects measured 1 generation down the line won't necessarily tell us enough (though in the vast majority of cases, it's probably good enough).
On the other hand, a purely-global view is not right either, because as you've mentioned, environmental changes or genetic shifts within the species can turn previously helpful traits into "bad" ones. So measuring regret after the fact will not suffice, and further, it misses the fact that at the moment the mutation happens, there is some distribution that describes the likely outcomes of that mutation given the current state of the world, using no information from the future (even if we don't know it or can't feasibly calculate it). The extinct bird species definitely had beneficial mutations even though some freak event wiped them out later, and we need to account for that. So something more subtle is required.
In order to do better we'd probably have to make some assumptions to eventually cut off the familial dependencies, like perhaps that the presence or absence of a gene in an N-th grandparent can have no direct causal effect on the survival of the animal in question (i.e. direct nurture effects are limited in time - even this assumption is questionable in the face of things like family wealth). We'd also have to assume that we could quantify the expected changes in the environment into some sort of distribution, including distributional assumptions for expected changes in the rest of the population (luckily these changes should be rather slow, when measured in generations). Then we could in theory compare the expected size of a family tree branching from a member that possesses a new mutation after that N-th generation, to the base case, the family tree without that mutation. Of course, the "expected size" is a trivialization of the real distribution, which would better describe the possible effects of a mutation (and the details of that distribution might strongly effect the population dynamics).
It quickly gets messy, and requires a lot of assumptions. And even after all that, we wouldn't have a very clean number to work with, i.e. we couldn't easily continuous-ize the situation and write down an ODE that took the "beneficial-ness" of a mutation and showed us what would happen, because other details of the distribution would be important, as well.
