Due to relatively large number of inputs (All my switches are hooked to a DI on a Neuron device, the same goes for motion sensor. Then I have each light connected over a relay output or digital output for PWM control. Then there are blinds, each requiring two relay outputs. ...) I have three Neuron devices installed. 2 x L203 and 1 x M203.
I am running Neuron OpenSource OS image from May last year I think.
In the beginning I did intend to run my HA server on one of the Unipi devices, but have ultimately decided to rather use a separate NUC server that I had laying around. But I don't see any reason why it wouldn't work on the integrated RPi (especially if you plan to use only one device).
I also did start with some of the links that you are mentioning. But all of them eventually came to the questions how to lower the latency of getting a signal change state from Unipi device to HA. All solution are designed around polling (MODBUS registers) - and this is already being done by EVOK SW layer running on unipi device, so why introduce another layer on top of that that would do the same. So then it is just a question how to hook to EVOK. With number of standard interfaces EVOK luckily also exposes asynchronous one (websockets) and this is where the performance increase (compared other polling mechanisms) comes handy.
I am running HA with Neuron device for a while now and I haven't had any stability issues so far. So yes, I would definitely recommend trying it out.
I don't have any Zigbee devices installed but there are standard integrations available in HA so this should work out of the box, if you ask me.