By looking at the genderless conditional probability ('somebody'), you're implying that women like Knox might have male-like murder levels, which is obviously wrong.

No, male-like murder levels would be higher than genderless murder levels. And, if I understand correctly, most of the excess male murder rate involves gang-related violence, which in this case was pretty clearly not involved.

Anyway, I agree that if you are doing pure Bayesian inference you have to condition on all kinds of available evidence, including gender, race, social class, nationality, etc. But we can't expect courts to consider this kind of evidence, for good reasons (avoid creating self-fulfilling prophecies and avoid incentivizing crime within certain demographics).