I suspect the "warping" you're referring to is a measurement of where a correlated featured was found relative to the center of an AP. I don't believe it reflects an AP-scale repositioning of the feature separate from the rest of a frame. Your comment about changing one AP yielding a different result is not evidence one way or another. Whether the alignment algorithm operates on whole frames or by repositioning pieces of frames, changing any single AP will invariably yield different AP weighting, different frame selection, different alignment measurements within the altered AP, and different whole-frame least-error alignment solutions. There's no mystery here.. different input yields different results.
All that aside, I fully agree with and appreciate your intent: disarming a common misconception (particularly among beginners) that they must avoid rotation blur even at cost of frame counts. This discussion and Ittaku's analysis has been highly informative and I hope someone reading this will be convinced that it's okay to shoot more than a few minutes. 
Ryan, I'm confused as to why you don't think the APs are stacked independently, especially given the direct quotes from Emil in the video. But, Autostakkert is not open source. However, Rolf Hempel's PSS program is, and so this discussion has led me to look a bit more deeply into his documentation. Rolf knows what he is talking about, and has done an excellent job not only with his software but also with documentation, which sounds like it was actually his main objective, providing an open source format that we can all learn from (and potentially contribute to).
Please look at this document summarizing the algorithms that Rolf has used.
https://github.com/R...thm_summary.pdf
The document is not very long, but I have called attention to a few relevant quotes below. These are in chronological order, but have gaps between them, so definitely read the source document for full details.
-from p. 12 "Ranking of Alignment Points":
...After all APs have been set, for each frame and each AP the image quality is computed, based on the alignment box around the AP. This is a very compute intensive operation...
...The qualities are stored for all frames in a list. The list is stored in the AP dictionary as “alignment_point[‘frame_qualities’]. A list of the best frame indices (up to the specified percentage of frames to be stacked) is computed and stored in the AP dictionary as “alignment_point[‘best_frame_indices’]. Note that these lists in general are different at different APs because of local seeing...
-from p. 13 "Frame Stacking":
...First, for every AP an array with the size of the AP patch is computed. It is filled with “weights” between 0 and 1. Weights are 0 outside the patch rim and increase linearly to 1 at the AP center [ , ]...
...The weights for all points within the AP patch are stored with the AP at “alignment_point['weights_yx']”. In both coordinate directions the weights ramp up linearly from a small value on the lower patch boundary to 1 at the patch center, and from there ramp down again to a small value on the upper patch boundary...
...If “number_stacking_holes” is zero, no background image is needed in stacking. In this case everything is set for stacking. If it is greater than zero, the stacked image contains holes, so it has to be blended with a background image. The background is computed as the average of the best frames. Only global shifts are applied, no warping. This image must be blended gradually with the stacked image...
...Next, the total shift at the AP is computed as the sum of the global frame shift and the local warp shift “[shift_y, shift_x]”. Using these shift values, function “remap_rigid” shifts the AP patch around the AP in the current frame and adds it to the AP’s stacking buffer...
-from p. 17 "Merging alignment patches":
...So far stacking was performed locally on the AP patches. Now those patches are blended into the global “stacked_image_buffer”. This is done by method “merge_alignment_point_buffers”. It is crucial at this step to avoid sharp transitions between patches. After all, they have been rigidly shifted, most likely using different shift values. Therefore, overlapping patches must be blended with each other. The difficulty is, however, that the program so far has no notion of AP neighborhood. This problem is solved by multiplying the AP patches with weight functions which smoothly go to zero on the patch rim...
Edited by Tom Glenn, 21 October 2020 - 11:44 PM.