If cell ID recognition is solid, could there be something funky in recording of cell IDs? The reason I'm asking is that yesterday at the office when I created a new situation to try the solution posted by @nodevel, I started cell ID recording but it acted as if no cells were available (count stayed at zero). This would explain why in some cases when a situation switch has failed no new cells have been found when I restart recording. Either that, or the cell ID recognition (or the platform API, perhaps?) has a hidden bug somewhere

It's true that it can be quite inaccurate, and especially near the borders of the area it's very volatile. The dampening I mentioned helps with that, but in my experience it also helps a bit if you have a trace log of the cell IDs with timestamps from where you can assign the cells to different areas/locations and immediately see when the coverage at the location is stable enough (Llama has such a log). It also makes it easy to assign cells to areas afterwards, as a complement to recording. With dampening it would not matter if the phone occasionally switched to a cell further away for a couple of minutes, and so it might be possible to keep the effective physical area more compact.