you just shouldn't trust anything they say.

Oh, trust me, I don't :-D

What we want are marginal costs, not average costs

The problem with marginal costs is that they are conditional. For example, the marginal benefit of your $1000 contribution depends on whether someone made a $1m contribution around the same time.

But we're ok with assigning credit along longish causal chains in many domains; why exclude charity?

I don't know about that -- I'm wary of assigning credit "along longish causal chains", charity is not an exception for me.