Today I corrected the test of the ESP32 as I suggested yesterday. I added more ESP32 and I modified the code to emulate a busy CPU situation. I also paid close attention to the regularity of the timing between emission and reception of the UDP packets.
Here is a picture of the setup:
For the question of the stability of the timing between emission and reception, I found out that it is not stable enough. Here we can see that the timing can at least vary between 0,1 and 0,2s which is too much.
However, as shown below, the precision is fairly good for the reception. This picture is the one with the worst result I could find. So we can infer that the reception will be around this precise.
As a conclusion, we can use UDP broadcast for time reference but the synchronizing packets should not be sent by one of the participating ESP32. Rather, an ESP32 from one of the bottom_pcb would be more suitable.