A hundred epochs, then fill again
Not every build needs every test. The default step stops depending on the test step, so zig build is a compile again.
The PyTorch script goes from 10 epochs to 100. Longer run, more confidence that the yardstick is a yardstick. Then fill after the fact: Tensor.fill, a while loop over self.data, because memset was the init path and this is the other lifetime. A test inits zeros, calls fill(2), checks the slice.
Fill-on-init was April 2025. Fill-after-init is June 2026. Same idea, different lifetime. Still CPU.