To gather the battery data, I ran the 3 PineTimes running stock InfiniTime 1.11.
I connected them to a Linux device with ITD.
I then set up a constant load on the PineTime, either with the heart rate monitor on and the torch on max, or with the heart rate monitor and screen both off.
I then used the itctl watch batt command to monitor when the battery percentage advertised by the watch changed.
I used the following formula to convert the percentages back to the voltage: 3200 + (reported_charge * 10 + 5) * (4180 - 3200) / 1000.
This is just a linear interpolation of reported_charge + 0.5 between 3200 mV and 4180 mV (min. and max. voltages respectively).
The reason for the + 0.5 is that the Bluetooth characteristic gets updated when the percentage changes, so the actual voltage will be in the upper range of the possible voltages for that percentage.
I also normalised the time between 0 and 100, and inverted that to get the "best theoretical charge level".
I then used some code to go through all possible values in a linear approximation using 6 points. I calculated the loss between the linear approximation and the best theoretical charge from all the tests, and chose the linear approximation with the lowest loss.
The results are shown in the CSV files below. The values where the watch reported 100% charge have been removed, as it is not possible to accurately calculate the voltage for those.