Given that, a computer which allows random audio API clients to destroy the computer's speakers is not correctly designed. Certainly, it may add cost, size, and weight, but if that's what you need for correctness, then that's what you need for correctness.
The brakes on a car add cost, size, and weight, but you can't just leave them out because you don't feel like accepting that penalty.
I don't find that controversial at all, but as time passes, it seems there are more ways that user activities can bring a computer down. The most obvious way is by sending out-of-bounds signals to storage devices that can erase or even physically damage them, but this example of destroying speakers, even though it should not be allowed to exist, seems plausible also.
Another common error is to allow a system power shutdown without parking the HDD R/W heads in a sacrificial area, so subsequent mechanical shocks won't cause the heads to collide with a legitimate data surface. This problem can be caused by insufficient power supply capacity -- a capacity that must detect the power shutdown and allow the head traversal to complete before the remaining power is used up.
Just examples, there are many similar ones.
> When the running average surpasses a value that indicates the speaker is reaching its limit, cut the volume. I take it that this doesn't work?
No, it might work, but it's really a question of cost, complexity and represents an added failure mode -- something else that can go wrong.