I have encountered numerous times that when my tasks are only minutes overdue they will be discarded by the project.

As for some reasons I cannot run BOINC continuosly on many of my devices this happens frequently. The outcome is that a significant percentage of computing time is lost on these devices. Other projects give grace time on workunits that are still in work. E.g. Primegrid adds gracetime when the client reports back frequently automatically. I even had workunits there reporting back properly while having exceeded the deadline by the factor of three, saving my computing effort and the CO2 emitted for the electrical power.

Is it intended to have that sharp deadline and its side effects here?

Hello!

If your result arrives on the server earlier than the result for task that was released to replace your result, then your result will be considered successful. Usually you have about 1 or 2 "bonus" days. If your result is received faster than result # 2 begins compute, duplicate cancelled.