We discussed the possible non-identifiabilities of the model at Change from global model fit to one-step-ahead prediction
I would still strongly suspect the non-identifiability is the main problem - especially since you mention you work with O and h on the log scale as they are very small and I would expect some of the non-identifiabilities to arise precisely when O is small.
I don’t want to pressure you into following my advice (it might in the end turn out to be wrong advice :-), but my experience is that attempts to work around modelling problems with some empirical tweaks to how the model is coded rarely help and usually only delay the necessity to delve in the math and understand the root of the problem. But obviously, you do you - it definitely seems you are learning a lot about Stan in the process and that’s also something :-)
EDIT: If you also could share a plot of the actual data, that might also help us see if the observed dynamics is rich enough to inform all the parameters…
Looking at the traceplots in Model fitting and sampling issue: Only 1 chain sampling properly makes me suspect even more that non-identifiability is an issue as there seems to be strong negative correlation between O and h
Best of luck!