I don't think there's an open-and-shut case that error values are "preferred or beneficial" in general with respect to exception handling through nonlocal control-flow transfers; there are reasonable arguments on both sides, and it seems likely that both are viable options in most circumstances, and which is better depends on those circumstances. (For example, lazy evaluation or catastrophic consequences attached to software crashes would seem to me to weigh heavily on the side of using error values, while interactive REPL-style usage would seem to me to weigh on the side of nonlocal control flow, perhaps with Common-Lisp-style restarts.)
My point was narrower: in this particular case, using error values analogous to NaNs, rather than shutting down the SRI, would have saved the rocket, without posing the risk of using garbage data that C's semantics pose. In this particular case, they would have been better. It's possible that, in other cases, perhaps even in other places in the SRI firmware, using error values would have been worse than unwinding the stack or shutting down the SRI. But in this case they would have been better.
(If you aren't convinced of that narrow point, you should probably read the report from the Lions board of inquiry.)
In the general case, the Ada behavior of detecting and signaling numeric errors seems to be the only robust choice to me. The alternative could only be acceptable if it's proven that the algorithms fail safe in the presence of tainted data.
However, default-ignore vs. default-explode isn't a difference between handling errors with IEEE-754-qNaN-style dataflow-tainting and CLU-style nonlocal control flow transfers, as you seem to think. In either case, algorithms that need data that could not be computed successfully will fail to get that data; in either case, you can write the code to either propagate that failure or to attempt to recover from it somehow. In either case, you will not be able to control the spacecraft if you needed that data to control it. The difference I was mentioning, which is not generally applicable, is that in unusual cases like the Ariane 5 disaster, where the erroneous data shouldn't have been computed in the first place because it wasn't needed, qNaN-style error handling clearly results in safer behavior.
But, in general, real-time control is not a place where there exists a "robust choice" when your software is broken. If your jet engine control software or your antilock braking software is broken, you're likely to die, and no error handling strategy can prevent that, at least not at the level we're talking about. Biological systems manage this situation by having a melange of somewhat-independent control loops using different algorithms that are constantly fighting for control of the organism; while this does seem to produce systems that are more resilient than engineered systems, it's less efficient and enormously more difficult to develop or understand, and it's the opposite direction from the rigorous certainty you and I favor.
Probably the fact that the engineers got the case analysis wrong suggests that the methods of reasoning they were using weren't very effective. Modern formal methods might or might not have helped; guaranteeing that a piece of code is statically free of exceptions is not a terribly advanced thing to do, but really you'd like to guarantee that the variable was actually big enough. Having less code to reason about certainly would have helped.
C’s choice is clearly unconscionable.