To make a long story short and close this chapter, the project was a success!
All the goals and conditions were met (with the exception of spontaneous combustion and/or disappearance trough a hole in space/time), and this is how it all went:
AVI@TOR MK3 was build, incorporating all of the knowledge gained so far.
This included two tone florescent paint job (orange for visibility in high brightness, yellow for low light), to avoid getting lost in the snow, clouds and twilight conditions.
Complete with reinforced wings (no more broken off wings during flight due to overweighted airframe), all the extra components packed in to the canopy, smaller cuts for a smaller phone resulting in weight reduction from previous model and so on.
And it flew on it's own, controlled only by an Android phone (HTC Legend), using only it's on-board sensors (accelerometer, magnetometer and GPS; NO gyroscope).
Battling winds (between autopilot sessions there is a moment when the plane is standing still in the sky), overweighted airframe (not easy keeping it steady) and slow sensors (accelerometer may be accurate but it's slow compared to a gyroscope). But even so it manages to fly in to the general direction where it was pointed, steadying itself and correcting course.
Unfortunately human pilot error (many like shown above) and flying home made electronic in sub 0°C weather (which has caused solder connection to break) has shortened the useful academic life of the plane.
But not before some more fun has been had with it!
Aerial video of the prowling grounds, this time on a phone with a decent camera.
Aerial video with a "fish eye" lens (an attachment, which was later found out to mess with the internal sensors of the phone enough to cause some autopilot crashes).
Not to dwell too much on the autopilot, some time went in to enhancing the distortion caused by the lens. First attempt, a brute force implementation of the Barrel algorithm (which took 7 hours to render the clip above).
Second implementation, using openCV and it's build in features (which took only 15 minutes to render, 2 minutes of that time used by a script to generate a command with over 8000 parameters (picture names)).
Despite all the mentioned adventures, experiments and much more, the plane still flies. But no more for science (too much glue to fly straight any more), only for fun and experience. A fitting retirement for a successful project!