I agree with you. But as with most prices, sometimes they are worth it. There are a lot of practical cases where you really want the performance (even outside scientific computing, which in itself is a pretty widely applied field).
> I'm sure there are some fields of academia where this is a reasonable tradeoff
I'm not sure where you got "academia" from; scientific computing is in no way restricted to research. Weather predictions, genome processing, and oil reserves exploration are just three of many examples for the commercial application of scientific computing.
> (if you're programming DSP algorithms running on a embedded battery powered device strapped to a dolphin, say...)
DSP applications are ubiquitous today; you'll find them, e.g., in your phone, or in your car.
> but your generic number crunching job rarely calls for it.
You're very wrong. Number crunching is the example for which you want lots of performance.
> As an aside, zero-overhead abstractions being unique to C++ doesn't sound right to me.
I didn't say they are. I said that no other language offers them to the degree that C++ does.
> so maybe it's just a name for C++'s tradeoffs?
No. It means abstractions which do not cause runtime overheads because the compiler can optimize them away.
(As an aside, zero-overhead abstractions being unique to C++ doesn't sound right to me. Then again I haven't heard the term used outside C++ circles and its meaning is nebulous, so maybe it's just a name for C++'s tradeoffs?)