Thank you for your reply! I will try that. When I can run sampling again 
I woke up to an output complaining of divergences so something like that eventually happened. There were none mentioned at ~99%, so what you mention may have been what found them. Arviz also complained about my pymc3 version (freshly installed, though maybe not via conda-forge) so I upgraded that to 3.11.2 and now I have new gotchas to poke around with.
Now I’m getting failed chains due to the “The derivative of RV … is zero”, which goes away if I specify init=“map”, which is discouraged (and much slower, it seems?). I’ll dig into my data and model and try to figure out what’s going on. This post seems helpful. Cookbook — Bayesian Modelling with PyMC3 | Eigenfoo
update: scaling my observed data (via sklearn.processing.scale) before giving it to pymc3 seems to have addressed the “the derivative of RV …” problem. Maybe it should be noted that my data consists of small floating point numbers with a bulk stdev of something like 1e-2 and a mean value of ~zero.