Thanks for the PR. A couple of comments:

We should keep the "This reflects the value of the system FLT_ROUNDS macro at interpreter startup time" text: there's important information there. In C, FLT_ROUNDS is the one float.h value that's not necessarily constant, potentially reflecting the current rounding mode. But Python's value is constant, and won't change if the rounding mode is changed dynamically. (With hindsight, I suspect that sys.float_info.rounds isn't all that useful, and would have been better left out of the struct - it's the only piece of information that relates to the semantics of operations rather than simply to the storage format. But it's too late to change that now.)

The "for floating-point addition" language seems curiously specific (yes, I know it comes directly from the standard). I'd expect the rounding mode to apply to all arithmetic operations (possibly excluding sqrt); not just addition. The C++ standard appears to use the language "floating-point arithmetic operations" instead.

Thank you for the feedback, @mdickinson ! I have made a few changes to address your points.

Just let me know if you see anything that I might have overlooked, and I'll be glad to sort it out.