But yes, Cluster was run on the cheap, hence the use of the Ariane 5 test flight, and didn't have insurance.
The board argues that there was a bias towards believing the software does not have an error. Thus, any out-of-range value is interpreted as a hardware error, which means the CPU should shut down.
There was a decision to not include Ariane 5 trajectory data in the SRI requirements and specification. Thus, while tests were rigorous at the equipment ("unit") level, and there were system tests, they didn't test that case. This is test design failure.
In addition, the board says "the review process was a contributory factor in the failure."
I can see how those can be aspects of "over-reliance on unit testing", but it doesn't explain, for example, how some of the variables from Ariane 4 were protected from overflow exceptions but others were not.
Lots of things had to go wrong to cause the Ariane 5 failure - including bad handling of overflow, as you mention. But to my mind, the universal last line of defence against any kind of mistake is an integration test: put all of the parts of the system together, feed them real input, and verify that you get correct output. Arianespace did not do that.
Well, until they actually launched it. It was a test flight, right? It proved to be an essential and very effective test.
Everything is incredibly obvious in hindsight, of course, but making things obvious is largely what hindsight is for.
And once you've finished reading that, go look up the Therac-25...
