Hi Gianmario!
You are complete right, this is a quite unexpected and somewhat problematic area. Part of the problem is the use of floats as representations. You have already identified the recent previous discussion and the much improved v13 (imho). But going from datetime - absolutedate - datetime, the user should expect it to be identical. The opposite conversion chain (absolutedate - datetime -absolutedate) is not true as there is more precision in absolutedate.
The rounding issue was highlighted in Rounding error in datetime conversion . I thought there was a reason that we kept int, but cannot find any now, it seems better to use rounding, yes. But the best would probably be to avoid floats at all and use similar approach as in v13 to take the microseconds which are native python stored as ints.