OK, so I went ahead and tried some groupings, even though we probably don't have enough data yet. So while the data fit for all-equal-probabilities hypothesis is indeed pretty bad, if we assume that recipes fall into just 2 different categories, then it looks much, much better.
As it stands, if we assume that there are 2 pools of recipes, and each pool has the same probability of being chosen, and every recipe
in a particular pool has the same probability of being chosen, then it fits existing data pretty well. The pools don't have the same number of recipes - something like 24/56 split, meaning recipes in the smaller pool have higher frequency of popping up (slightly more than 2x more often).
This is all highly speculative so far, of course