Serial Experiments Lain challenge - revisited

[igro]

I got my hands on some DVDs of Lain from Japan (long story short - it pays to scare the hell out of your Japanese neighbours by greeting them in Japanese and then start talking some more ). I've heard that Lain was a bitch to encode (due to the crappy R1 source and what not), but at the same time it seemed like a perfect candidate for a vfr encode, so I wanted to give it a try.

The source seems pretty good to me (considering this is not a new show, etc) - no haloing/whatever, but I have some serious trouble getting to deinterlace it.

Here's a sample: http://www.mediafire.com/?mlxzmjx29qz

For most of the eps, everything seems to go ok: it detects film and video sections, everything is smooth, but look at the above sample from episode 11, especially these two parts:

(notice the obvious combing on the mouth)

(notice the not-so-obvious combing on the mouth )

No matter what I try I can't get these two parts right.
My script is pretty simple, based on the vfr examples from the tivtc documentation, NNEDI + TDeint.
I've tried many options of tfm, tdeint, but to no avail If someone could point me in the right direction, (even "RTFM for x filter" ) say if this can even be done, I'll be very grateful