My $.02. To be fair, I've read pages 1-6 and 20-21.
There's some amazing work going on here, but I wanted to comment a bit (ok a lot).
Noise reduction: There is some serious degrain/denoise/fft3dfiltering going on. It seems like you are guilty of scrubbing all the film grain with the noise. Have you considered implementing an addgrain filter at the end?
I'm assuming you all are being careful not to lose too much detail with the noise. It can often be hard to tell from stills. The clips Dark Jedi posted don't seem to have lost too much image.
Interlace artifacts: When the GOUTs first came out and the first scripts for clean up were being discussed, filters for the interlace artifacts were heatedly debated. I don't see anything being used on them here. (Or did I miss it.)
Subtitles: I'm surprised no-one is talking forced softsubs. It gives you the ability to generate separate letterboxed and widescreen subs so positioning is optimum for either 16x9 or 4x3 TVs.
Burned in subs seem a bit dated is all.
Sharpening: Different sharpening algorithms can improve different features of an image. For any given filter you can only go so far before ringing or edge enhancement artifacts become an issue. I've found that MSharpen used mildly in conjunction with fft3dfilter's sharpen provides a nice balance. Here's some other sharpen filter ideas: http://www.aquilinestudios.org/avsfilters/sharpeners.html
Banding: Fft3dfilter use can result in color banding or posterization. I'm definitely seeing some of that in the stills and clips (look at the close up of Han's hand). The best answer is to either back off of fft3dfilter or use a deband filter. http://avisynth.org/mediawiki/External_filters#Debanding
Sample clip: In the still below, the silhouetted figure walking in the background leaves trails. Trippy and an artifact of agressive temporal noise filtering.

Other stuff: I don't know if it was answered, but DGIndex's iDCT algorithm decides how the mpeg video is decode for frame-serving. Simple MMX (XviD) is commonly recommended because it is relatively accurate and very fast.
In modern processors, I don't rally see any difference in decoding speed. For this project where quality is king, I would recommend using "IEEE-1180 Reference", because that's exactly what it is: the reference algorithm for decoding (at the expense of speed).
I really should read the rest of this thread so I can annoy you all more.