I'm a bit confused, as the documentation states that the fused type with greatest precision is used. Which of double and int64 is the type with greatest precision? We want double, but the warning says int64.

I'm a bit confused, as the documentation states that the fused type with greatest precision is used.

What documentation are you referring to?

Fused types allow you to write code that can be called with any one of an enumerated set of types. Precision doesn't come into place in my use here.

In this PR, call sites that pass in ints will use the int version and call sites that use doubles will call the double version. There are two different functions for each case in the generated C code.

Maybe the docs you're reading are talking about a fused type like cython.integral that can be any of the integer types? In that case precision does matter.

My impression was that what is wanted is a cast of int64 to double, with maybe a check that the value is preserved.

I’m implementing @rkern’s suggestion from here.

But I think it could be worth robustifying both, using np.arange() in the first instance and adding an integer check_constraint() variant in the second.

This is exactly that suggestion. I could have simply casted to double, but that could overflow and IMO this is better.

LGTM also, thanks all!