Worth looking at how the car tries to guess the battery capacity. There is no reliable way to measure this, short of discharging the battery to a particular state of charge (SoC) at a fixed discharge rate (because discharge rate has a big impact on losses) and fixed cell temperature (because temperature has a big impact) and then recharging the battery at a fixed rate (again because charge rate has a big impact on losses) and fixed cell temperature to another given SoC. This can't realistically be done, mainly because when driving the discharge rate varies massively from minute to minute, and also changes from discharge to charge fairly often, due to regen.
Measuring the battery voltage doesn't work too well, as apart from this being pretty flat over a wide SoC range in the centre of the discharge/charge curve, it will also vary with the charge/discharge current and with battery pack temperature, Add in that the cell balancing phase at the end of an AC charge significantly increases the charging losses, as does temperature and the SoC, and it's pretty clear that the SoC can only ever be a rough estimate of the true battery capacity remaining at any time.
Tesla seem to try and fine tune the way the system guesses the battery capacity/SoC, but even with all the corrections incorporated in the system I doubt they can get closer than about 10% to the true SoC at any time. It also seems as if the system tries to learn from the pattern of use of the car, to try to refine the information it reports, and the chances are that this feature alone may well account for an apparent steady change, when in reality the battery hasn't changed much at all. Then chuck in that it seems as if Tesla have altered the code that runs the battery SoC guesstimate at least once, via OTAs, and the chance of historical data being a true representation of the real state of the battery are probably pretty small.
