The problem has finally been pinpointed, thanks Zaxo, tye and broquaint for their suggestions.

It turned out that the problem was in the library's C++ code. Internally, exceptions are caught using

  try {...} catch(...) {...}
[download]

In the catch, an error code is set, which is subsequently tested in the calling code (via a macro expansion) so that an exception is thrown (just don't ask, it's not my code). The problem is that in the original catch some logging is done which... yeah, right, generates an error and hence some exception is thrown that only got caught in the catch(...) of my XS code, completely bypassing the intended exception throwing/catching mechanism.

Lessons learned:

• exceptions are handled as one would expect in this setup (NT, ActiveState Perl, MSVC++),
• never ever trust someone who says he tested his code (his test may be succesful, but simply not adequate to find the real problem).

Thanks again to all who helped, -gjb-