FYI:
The vectorization and compilation-time issues in this thread–and what @gmarkall mentioned about typing–are a couple of the reasons why I (representing the Aesara group from which @aseyboldt and PyMC Labs forked) started looking into Numba’s compilation pipeline and–eventually–its typing system a while back. The typing/compilation-related dependencies mentioned above are also part of the impetus for recent discussions regarding the typing system’s external use and refactoring.
More specifically, some directly related concerns involving compilation and vectorization of Aesara-generated Numba code were covered in Numba performance doesn't scale as well as NumPy in vectorized max function - #11 by brandonwillard, where the topic of Numba-generated loop IR and its connection to vectorizability was covered and the exact same areas of the compilation pipeline were similarly altered.