I can’t reproduce my own error after restarting my environment. 
Hours of diagnostics later, the proper latent space model runs–first one I tried. SO I can’t reproduce the error. MAP still consistently estimates a z full of zeros, and every draw on a Metropolis trace is still 0, (I often use those for testing as I build models because they’re fast). But NUTS no longer throws the error it did before. I remember it being something about empty slices, but I can’t remember specifically. I’ll post again if it comes back.
I’ll also post a notebook with some version of the Krivitsky, Handcock, Raftery, & Hoff, (2009) latent space network model. I haven’t seen any pymc3 latent space examples, so hopefully it’ll be useful to someone else.
But thanks for your help, sorry it turned out to be nothing. For my part, at least I understand a lot more about Theano variables after all this.