Hangups always happen in the smallest details. It took no time at all to build an entire new psychoacoustic engine from scratch (once I managed to track down the appropriate research papers), but something simple like mapping the new log scales onto the existing codebook abstraction turns out to suck time because I hadn't stopped to think about it long enough...
(How do you add a number to -Inf? You don't; you find a way to avoid doing so. There went wednesday night through Saturday)
The hard part is never getting something to work. The hard part is building an abstraction you won't be ashamed of two weeks later.... especially vital in the case where the abstraction could well last 20 years. Bad way to lose alot of time when chasing an unreasonable deadline....
Anyway, the major hangups are past I hope; decode is certainly clean and I can live with well documented hacks that are encoder only. Integration of the New Stuff (tm) with libvorbis is first-pass coding complete, but totally untested. I'll do that after sleep. After making sure it works, time to build codebooks--- after adding log quantization code to the trainer. Gaah.
At least the end rush will just be for the website.