I think the current implementation makes sense. Autoload as standard for the 99% of straightforward use cases where this is the best option, with the option of disabling this per cue and, using load cues, to precisely determine what is loaded and when, for more advanced use, and to interrupt autoloading of long chains of linked cues to prevent memory pressure, or interruption of currently playing material.
Unpicking the rest of your post,
Loading the first few frames of 12 videos should not really overtax any QLab system. The memory pressure should remain relatively unchanged and reasonably low, although it might take a few seconds, which shouldn't be a problem at the start of a 12 hour playback sequence
MP4 is a container. I assume you are using it with h264 codec. Because of the number of settable parameters in h264 it is possible to produce files that have excellent playback performance, but the main problem is that it is far more likely to produce files that don't play nicely , particularly if using keyframes with long GOP intervals. I don't really understand why the 50 quid it would cost for a 1TB HD to allow you to use ProRes files isn't a better option.
Having said all that, I just tried this in QLab 5 and it appears there might be some playback anomalies with sequences of very long files (which I have reported to Figure 53). Are you using QLab 4 or 5 for this?
Best
Mic