I do think your definition of a prior is a bit off though... like, the choices around priors wouldn't be whether to use "did she do it" vs the imbalances of evidence or such. We're trying to estimate the likelihood of "did she do it," so the question is what do we use to do that? The prior probability of "she did it" has to be the probability that some random member of the 20/female/american/student general public will commit the murder, because we don't have anything more than that (as you said). Then, we update our estimate by looking at the available evidence (such as, "was found standing by door holding bleach mop," which would the viewed in terms of the likelihood that that would be observed and the statement 'she did not do it' being true) to come to our posterior probability, which is the estimate informed by all of the available evidence.
As a side note, it's great when you can have a civil discussion on the internet. Maybe there's hope for HN after all.
But I strongly disagree on your understanding of the impact of priors. There are several ways to look at it, all equivalent. The easiest is the single prior: Did she do it? None of us are psychic, so that isn't helpful, 'cause we cant estimate it at all. We have to break that down into some of its components.
In fact, your argument (which I totally agree with, BTW) is based on a very specific prior: that in cases with this kind of relative imbalance of evidence, there is a high prob that one party is guilty and the other isn't (or some such, I may not be stating it as specifically as I should be).
All the evidence presented has it's own implied priors, and the right combination of evidence has it's own prior, etc. These all need to be combined with weights based on other priors. Since noone can compute all of these priors in the space of one hour, that makes it all a judgment call, not a rational decision.
And this is where I think the post's pontificating gets it terribly wrong. There's nothing rational at all about the process, other than you're attempting to blindly apply the Bayesian formula.