LZMA's range coding presents an unforeseen complication. Previously I had thought that the only state the decoder required was its position in the bitstream and knowledge of previously decoded bytes, and that if I could find a way to supply both to a decoder, I could start it decoding from an arbitrary point.
The range coding -- the Markov part of the Lempel-Ziv-Markov Algorithm -- means that the meaning of each bit in the bitstream is dependent on the values of all previous bits.
If I want to make a version of LZMA that can start decoding a stream in the middle, I'm probably going to have to completely get rid of range coding. That's really NOT going to help the compress ratio. At this point I have to wonder if I can even get better than 1:1.