I think autopano-sift-c is maintainable, it doesn't have any awkward dependencies, and as far as I know still builds on all platforms despite abandonment. If you want to take it on then please dive-in.
I'll try and recall some history:
The C# implementation (which involved running two tools consecutively) was replaced by the C version that was contributed anonymously, this is the code that you can find on sourceforge.
Hugin has never shipped with an autopano-sift-c executable because of a SIFT algorithm patent (though Hugin has the capability to run autopano-sift-c for control-point finding if you have it).
Pablo wrote cpfind as a full replacement that doesn't infringe the SIFT patent, this *is* shipped with Hugin, and as far as I'm concerned cpfind is as capable and as fast at feature identification as autopano-sift-c. There hasn't been much interest in autopano-sift-c since.
I believe that the SIFT patent expired last year, so there is some value in looking at all this again - in particular because (as you have identified) there are some very mature GPU based SIFT implementations in libraries like opencl.
So, upgrading autopano-sift-c to use a modern SIFT implementation would be interesting, but there are two drawbacks: we don't actually know who wrote it, though they did attach a GPL licence, so it has an odd copyright status; it also reads a limited range of image file formats, whereas cpfind can read all the formats supported by Hugin.
Modifying cpfind in the Hugin project to use a modern SIFT library *could* produce a dramatic improvement in speed and quality (though nobody knows without trying), if so, and if it doesn't involve exotic library dependencies, it would be welcome as part of Hugin. This would be the best place to do this from my perspective.
--
Bruno