Sadly many of those instruments are in .nki or .nk* format which is the Kontakt Player or Kontakt Something Fullversion format.
The wave samples are (often? by design?) there as plain files, but it is hard work to guess how they should be arranged and what is needed. As far as I know the kontakt format has more features, such as scripting, than sfz, which is currently the "Linux sample flagship". I hope I am wrong here.
No, you're correct. As i've written in my other mail to LAD, todays sample libraries are much more than samples and the respective mapping on keys / velocity layers. Scripting and all the advanced engine features is what makes the difference. Just extracting the samples and mapping information from a Kontakt library will end up in very disappointing results.
I know there are sample converter programs (for Windows) like Chicken Translator or the "W. Grabowski Extreme Sample Converter". I have used them and even simple conversions like sf2 to gig, or gig to sfz were always a bit odd or plain wrong. They have menu entries for Kontakt and EXS24 (the Apple Logic Sampler format, you see that quite often as well) but I don't believe that will actually produce accurate conversions.
Some of those tools are using a library which we'll provide to read and write the Kontakt format. Of course they can only convert the information supported by both the source and the destination format.
Is there someone who knows more about these formats? Even if it is not possible to write a sampler engine (it would not be the first [partly]binary, closed format loaded by open source software) maybe there is at least a way to get all the needed information to convert/correct them by hand or individual scripts.
Besides the technical problem there is also a legal problem involved here. To protect the IP of sample developers the instruments sold by 3rd party developers are encrypted.
My guess would be that using something like Autosampler to capture a certain snapshot of a library and convert that into something that your favorite can read is by far the most promising solution. But of course this depends on what you're looking for.