Think this through a little. If you load up a week's worth of the short units ( I am speculating for your machine that would be a total of about 105 WU per core per day or 735 WU's for a week per core. That gives about 12 minutes per WU. OET work units vary between about 5 minutes to 2 hours. If you loaded up 735 WU's per core and they went to only 1 hour, you would have 735 hours of work for that core. Since there is only 168 hours in a week you would be severely over cached. Hence the limiting mechanism. It gives BOINC a little time to adjust should the WU's become longer. Even with these very short WU's you still should be getting a steady supply so your machine is not idle.
Hope this helps.
Cheers

35 wu max and 10 days to complete only has problems if completion time exceeds 6.85 hours.

I would have thought this setting might be project specific

OC - You could simulate more cores than you have while at the same time limit the number of concurrent wus to run to your actual number of cores / threads. While this works pretty well when you are only crunching one sub-project, if you want to go multiple sub projects you can start looking at changing config to require the appropriate multiple of CPU per WU so you can stay balanced. I've only ever done this back in the Cancer GPU days but the theory still holds. Hit me up over on XS if you want to walk the details.

I leave my work pc unattended and offline during weekends.
that makes 52 Hour long offline period which makes 25 WU's per core limit risk of drying up my backlog.

Here is the solution I come up with; I also work on some longer WU projects like cancer or solar thing. that way 25 Wu per core does not affect my rig.
if you are in such a case give a hand to other sisiter projects a hand as well :)