I remember why I wanted to keep that last document. It ties in with things that Larry and Raymond (can't find any references right now) and of course John Robbins have said, and also a thread on the CodeProject Lounge about gotos in the MFC source code.
I believe that, if we're going to deal with leaky abstractions, we have to know and understand what's going on under the abstraction. If you don't, when the system does something crazy - and it will - you won't be able to fix it. While our virtual machines are leaky, and especially while they're unreliable, you have to understand both the VM and the real machine beneath. This is particularly true in the edge cases where you're trying to interface between the VM and some native code.
Java tells you to wrap up your native code in a nice interface using the Java Native Interface. Fine, but that requires you can write C++ code (and I think it ties you to a particular JVM implementation - not sure on this). .NET tells you to marshal using [DllImport] attributes (or Declare statements, for VBites) which are a bit neater, but you suffer the leaking abstraction when the type you're trying to marshal has many incompatible options for implementation on the native side. I've answered (or tried to answer) a number of questions about P/Invoke interop marshalling, and people are always trying to do something crazy with strings. There are many ways to marshal a string, and some are better than others (hint: on the way in, pass an LPCTSTR; if you need to pass a string out, declare two parameters, an LPTSTR and a buffer size, and use a StringBuilder on the managed side).
Anyway, if you need to crack a tough problem, it helps if you have all the tools in your toolbox - you just don't have to use the dangerous ones. Don't get hung up on the taboos - if a tool is useful and appropriate for the situation, use it.