Windows software is a similar case - a program works on one version of Windows due to accidental dependencies on Windows implementation details (e.g. memory management), and then fails on the next version of Windows when the underlying implementation changes. Raymond Chen has written about the huge difficulty for Windows to maintain backwards compatibility with "broken" programs. New versions of Windows provide special-case handling for old applications so they will still run. See http://www.joelonsoftware.com/articles/APIWar.html and scroll way down to "The Two Forces at Microsoft" for a long discussion of this, and how Apple is much stricter.
In the terminology of this article, Torvalds is firmly in the Raymond Chen camp as far as "The kernel is not allowed to break user software" is concerned. The difference between Windows and Linux (and especially between Windows 95 era Windows and Linux of the same vintage) is, apparently, that Linux didn't come from MS-DOS, and so never had to allow application software to get hooks into low-level parts of the kernel.
There was never an official version of Linux for hardware without memory protection, and there never will be. Scope is important.
No this is actually very different. Linus doesn't want breaking API changes to documented behavior, in the Raymond Chen case its not breaking applications that misbehave or abuse undocumented behavior.
The point I was making was that Linux didn't expose the same kind of deep, undocumented behavior because, as I said, it always had the ability to hide its inner workings.
