Because the deadline for FAH2 WUs is just 24 hours. The others are generally ten days. BOINC will endeavour to ensure that WUs meet their deadline, so to do that it suspends execution of other WUs with a longer deadline, and runs the short deadline ones instead.

Why is the deadline so short? Because the techs and the scientists want it that way. If you don't like it, don't run them.

[I don't understand the other question. Maybe someone else can answer.]

Most of the sub-projects at WCG are discreet independent work units which answer a small independent question – usually does this small chemical molecule fit into and bind to a specific active site on a much larger molecule (usually a protein involved in a disease) and block its biological function.
FAH2 is different – the output of one small step of the process is used as input to the next small step. From memory the number of small consecutive steps required for one overall calculation is of the order of 1,000,000 or more.
When FAH2 was first introduced to WCG it had work units containing 100,000 small steps, had a duration of 10 to 15 hours on a fast machine with a standard deadline of 10 days. A sequence of 10 to 15 work units was required to complete one overall calculation i.e. anything up to 150 days! As the work units took a long time, a ‘trickle up’ mechanism was used which reported progress every 10,000 small steps. Progress was monitored by the server and if it was not fast enough to meet the deadline, the work unit was aborted with credit given for the work already done. This process was not always liked by the volunteers. There were also problems with the ‘trickle up’ mechanism which depended on continual access to the internet as the trickle results had to arrive in the correct order. This caused problems for those computers with intermittent access or after a longish scheduled maintenance slot at WCG.
When FAH2 was re-introduced, the ‘trickle up’ mechanism was abandoned in favour of smaller work units containing 10,000 or 50,000 small steps. The break between the shorter and longer work units is at batch 1770. In order to complete one overall calculation in a time agreeable between WCG and the scientists a 1 day deadline has been implemented which forces BOINC into its earliest deadline first mode from the normal first in first out mode. Even so in my BoincTasks History I found a long work unit of 50,000 small steps in its 40th iteration i.e. 2,000,000 small steps taking about 40 days so far! With the standard 10 day deadline this would not be practical.
I hope this gives some background to this project.
I run this as my only sub-project but restrict it to use only a maximum of 6 of my 8 threads. I also have ‘send other work if none is available for selected projects’. I use a 0.5 day buffer so most ‘other’ work is completed within 2 days.

I have all projects selected and haven't done any sort of configuration - if the tasks go up late, they go up late. AFAIK the FAQ says that if you send within 7 days of the deadline that it will still give you points. No need for such short timeframe.

No, one I submitted 45 minutes late was reported as valid, others accepted too.