Cultivation is not fully implemented because the current Healing simulation does not track the health of the raid, see documentation. This is a huge challenge of simulating healing, and one that is not easily solvable. I don't know what AMR's plans are for this, but I also hope that something is implemented to approximate real damage patterns. The other solution (that is mentioned in the documentation) is setting a pre-defined proc rate for Cultivation, but this would have to be configurable in order to accurately represent different fight types, and that setting would probably change our calculated stat weights significantly.
My current stat weights give much closer goals to what I expect, though versatility is aiming a bit lower than I would normally expect. I use the weights rather than machine learning, and alter them slightly. To generate those stats, I reduced the sim length to 120s, rather than the default 420s simulation. The long simulation significantly reduces the value of haste, since you hit OOM at around 100s or so, and I take mana management to be a in game decision more than a stat decision. 120 gave me results close to what I expect based on TwigIt and RDSW, which I then modify to slightly increase the value of haste since I view that as being more useful for raid blanketing large damage events.
My gearing strategy selects Tearstone over Sephuz, by a pretty huge margin. In this case I expect it's because on a long fight where the Sim OOMs, it stops casting WG, greatly devaluing Tearstone relative to Sephuz, which will still be increasing Rejuv and CC Regrowth HoTs. So with my shorter Sim time, Tearstone wins by a landslide.
I also have the strategy suggesting that i swap out Velen's (for Ekowraith in my case), which I have not quite figured out yet. I think it might be that in practice, we get more value from Velen's since we line it up with when healing is needed, and the overheal redistribution gives an added benefit. In the Sim, it seems to sync it with Tranq, then burns it on CD. Since all healing is effective healing in the sim, there's no burst of healing needed to sync the trinket (or tranq for that matter) with. That would reduce the relative value of the trinket effect when compared to the raw throughput of an additional 25ilvls on the chest, plus Ysera's Gift.
Overall, my suggestion for resto sims is to use a short fight length; 120s worked for me, but you can try shorter or longer and see what you get (the OOM point is around 100s with the Spread Healing rotation). After getting the results however, I suggest not trusting the result entirely, since the value of mana trinkets will be devalued significantly, and the effectiveness of various legs and trinks depends on fight and usage more than the sim accounts for. Supplement the results with other theorycrafting sources such as Twigit, RDSW, and the rdruid community for trinket/leg choices. The sim result for me aligns closely with what I expect, so the only alteration I have to make is manually locking in Velen's, and slightly raising the value of haste, as mentioned above. If I were to adhere to the community guidelines for stats (rdsw or twigit hpm) I wouldn't even need to change the haste value, but that's a personal call I made. I will note that the only difference in my final gear selection, the Machine Learning selection, and the unaltered stat weights from the Sim is enchants/gems, since all of our stats are so close (and the sim captures this result).