> That people have scrupulously cleaned the board to get the clock working is crazy stuff.
The "watch" oscillator is an ultralow-consumption analog circuit, so you have to treat it as such. It's easy to create a robust high-power oscillator (well, the crystal would need to be built to withstand that, too; possible but there's no reason to do that); except that you wouldn't be able to run your RTC from a coin cell for months, and this is not what most users want.
In fact, you *CAN* run the RTC out of HSE on most of the STM32 families (I don't use the 'H7). And HSE, while still analog and still needs relatively careful treatment, is usually more robust than LSE; but logically there are no low-power facilities available for that.
> Putting a scope probe on the xtal doesn't seem a valid measurement with such a dodgy clock.
That is of course related. Typical passive oscilloscope probes represent typically 1MOhm and a couple of pF (or using 1:10 probe representing maybe around 10MOhm but few tens of pF), which is WAY too high a load to the typical LSE circuitry. There are active probes around which you can use to probe such circuts, but they are usually at the higher price range; and still it's a nontrivial thing to do without influencing the given circuit too much.
> without the CPU running it's difficult to know
You can output LSE to MCO in some families; again, I don't know about the availability of this on 'H7. Not in VBAT mode, though; but maybe in some low-power modes it may be available; check the RM.
> Unless you seal a board hermetically in its working environment it is going to get "dirty" thus potentially stopping the clock.
If you are concerned about [conductive] "dirt" (or, for that matter, humidity condensation or other environmental impact), consider using some form of conformal coating.
> I also couldn't make any sense out of the 512 Hz cal signal; it seemed to be running at 434 Hz.
I don't say the 'H7 are flawless, but I find it unlikely it wouldn't work at all.
> I also had trouble with a Discovery STM32H7B3I-DK.
That's disappointing indeed as those boards are supposed to be "known good", and I'd love to hear from ST in this regard (but won't hold my breath, ST almost never comments here on more complex technical issues); but again it's an overly complex board with too many sources for disturbance. Try maybe with a simple program with nothing else running (mainly no LCD and external memory) but the RTC and some suitable output pins. Or, if you have one at hand, try with a simpler Nucleo board.
> Even with "good" boards, the RTC calibrate frequency jumps around (with calibration turned off) by about 0.0001 Hz, making it difficult to know the exact calibration value to use.
The 'H7 is primarily a massive digital circuit built on a 28nm technology; I believe the analog performance on such an IC is a matter of heavy tradeoffs. Even with perfect 32kHz oscillations, I would expect variation in the input comparator threshold resulting in jitter in the order of say quarter of the 32kHz oscillation period, i.e. around 8us. That would exhibit itself as 8ppm jitter at 1Hz, whereas your finding is an order of magnitude higher, around 100ppm. Still assuming perfect 32kHz oscillation, maybe this is the contribution of the asynchronous part of the divider. I understand your frustration, but would you be willing to experiment further, maybe you could try to measure over a longer period.
JW
