It’s funny how this discussion is rather closely following a 2.5-hour argument that occurred in the chat room about a year ago, where the leading contenders were @thought and Pan Piper. Thought also argued that the natural number for komi must be an integer, since go points other than the tie-breaker are integers. Of course, when trying to derive the number from a large data set, you will get a fraction. Someone said that analysis of a large number of pro games had yielded a number very close to 7 under Japanese rules. Since the 0.5 functions, in practical terms, as another full point in tie breaks, it has the effect in such circumstances of bringing the number up to 7. But as I understand it, the 0.5 does not derive from analysis, it is simply tagged on as a tie-breaker.
There is actually a small, OGS, data set that someone compiled, that shows odd variation in win rates by rank, which suggest that komi is a sliding scale depending on rank (see Win rate by players ranks in OGS). Like someone in the discussion above, I would like to see a large analysis of.wins by rank and color, which might establish a more realistic komi at different ranks (I talked about this here: New way of deciding Komi).