> ...it is not true in general.
Agreed. But it is true for your original argument applied to a non-equilibrium state at any single time instant, while not assuming any more on the past or the future states.
If we adopt further assumptions about the past states, but not about the future states, then we agree this problem of prediction/retrodiction is asymmetric between past and future and this may be the reason why the multiplicity argument does not have to result in past entropy being higher. In particular, there may be other facts on record about the past states and that would override/modify the multiplicity argument for the past.
> it is possible to change your scenario and collect data that allows reliable extrapolation into the past, for example, by looking at the coin before it is flipped and remembering (or writing down) what you saw. But for the future this is not possible. That is the essential asymmetry.
Yes, in any realistic scenario involving prediction and retrodiction at any time, we have some data on the past states (in some lucky cases, enough data for having certainty), but never data on the future states (we may have predictions, but those are not of the same quality as data). I agree.
But in addition, you seem to argue that this asymmetry of availability of data is an essential fact to derivations of the 2nd law.
Consider a scenario where we have an isolated system that is a mixture of miscible liquids that exists for ages but we have data on its state only for the last 2 years, and that data says its entropy was steadily rising the whole time.
Now, given this data, was entropy increasing before the collection of data started, or was it decreasing? For short enough times around the edges, we can extrapolate the entropy curve based on collected data, but for long enough times, we can't. Using the multiplicity argument, including the data, we get that entropy of very distant past states was much larger than entropy of the first recorded state. Which contradicts 2nd law.
Effect of that "essential asymmetry" has a time span that is not universal, but limited to times where collection of data was happening. Outside it vanishes, and using only that data and the multiplicity argument, the distant past becomes as cloudy(highly entropic) as the distant future is.