I have years of history to the contrary. Until I moved over to the C-7/2.2.5 all of my scenes functioned reliably with not a single device being forgotten (or being sent the wrong value which to me is even more strange). So I have to expect that the issue lies elsewhere. In the past I had more ZWave devices as my entire HA environment was ZWave without any WiFi or ZigBee to share the load. So the fact that scenes worked on the older hub with older 300 series ZWave devices suggests the problem is something else. Whether it is with scenes or not I am not sure which is why I'm posting so much here.
This also does not explain why the same problem is seen in WiFi and ZigBee devices. Also if this were a ZWave mesh problem I should be suffering with other ZWave mesh problems, no? I'm mostly 500 series devices and I have a really well distributed mesh around my apartment. Also I am experiencing this issue with scenes that contain only 3–4 devices and scenes that contain a dozen. It does not matter the quantity. In fact a scene with only one ZWave device and one WiFi device exhibits this strange behavior with both of those devices being within 6' of the hub.
Lastly I'll add that I see in the logs that Hubitat doesn't send an instruction to the forgotten device(s) at all or it sends the wrong instruction. I do not see the Hub attempt to send a message but then that quietly fail with the device not responding. Instead I just see no message sent. That could be how scenes actions are logged or it could be how ZWave mesh overloads quietly fail. But again then why do I see the exact same forgotten/incorrect value with WiFi and ZigBee devices?
I tried making a rule that duplicates the effect of a big scene by sending lots of devices instructions all at once. The result is perfect performance. So there is a difference. I do not like this because I prefer to separate concerns: scenes record/recall state, rules express logic and function… However this might be my only solution since I have scenes failing that only contain two devices.
Perhaps there is an opportunity here to add to scenes a thin layer of smarts where scenes can optionally "confirm" that they're correctly completed or (and this is probably better as it is lighter weight) expose a way of partitioning scenes into separate chunks to be run with a 0.5s delay between for mesh reasons?
