It seems I got a bit further.
With one Raspberry Zero W ("machine 2", above), I had observed that sometimes one particular sensor, and sometimes the entire bus, disappeared from the system. The workaround described in this thread did well every time!
As mentioned earlier, each of the errors (except the very few single-reading errors) occurred exactly when another load had been switched off from the power grid. This load, neighbouring the power supply of the raspberry in the grid, was in one constellation the aquarium heater, in another constellation the icebox compressor, in a third constellation the gas boiler.
Now I observed additionally that with switching the gas boiler off, sometimes (frequently but not every time) one GPIO relay switched "uncleanly", with a hearable sort of double-click. I went further into that, and the reason turned out to be a large, but extremely short (few microseconds) voltage peak measurable everywhere in the system consisting of the raspberry and its peripheral electronics. I have to say that my measurement was not well shielded, so that quantification and localization failed. The reason for this voltage peak was (most likely) that the gas boiler has a transformer inside, which is a considerable inductive load switched off each time together with the gas boiler (also, the pump is an inductive load, but the gas boiler was never switched off with the pump running).
I did not find out whether the voltage peaks came over to the raspberry system via power grid or via electromagnetic radiation. I did no tests to differentiate that.
was to put an R-C combination ("snubber") in parallel to the switched load (not across the swich, not across the unswitched grid supplying the Zero, but after the switch in parallel to the switched load). I used 47 nF and 47 Ohm in series, with an additional 1M5 resistor in parallel to the capacitor.
Since that, all relays switch as they are supposed to. Moreover, most striking, the problem of disappearing sensors, or the entire bus disappearing, has vanished completely (if this can be said after only 3 weeks).
Of course, this is a solution working in my particular constellation. Yet, switched inductive loads are ubiquituous, and the may be one reason for the unreliable 1-wire bus of the pi. If so, a causal solution by a protective circuit, as described here in its simplest form, should be possible in some cases.