And here I was thinking that my afternooon was going to be simple ...

I know what the cause is. It's really strange - NumPy didn't report this in the changelogs, so it looks like an old feature. In fact, I think it's very very new. We'll need to add a backwards compatible shim until then.

also - thank you for making this an rc :-)

I think the breakage might also be in 2.4.7 — the RC was intended to catch a typetracer change. This wasn't caught in our CI — I'll need to check our lower bounds on NumPy.