I'd definitely go with throw - it's indicative of a logic bug, and I'd rather notice it during dev (or if not, catch in prod, complete with stack trace and what

I'd definitely go with throw - it's indicative of a logic bug, and I'd rather notice it during dev (or if not, catch in prod, complete with stack trace and whatnot), than to have it silently corrupting anything that relies on that calculation happening correctly, without me even being aware of it. Like, if I'm doing x + y, and y just accidentally happens to contain the wrong data because of a mistake in the code, a NaN means the code will proceed as normal, and nobody will notice anything is wrong. If it's not a noticeable bug, it might be a long time until it's spotted by someone. And when they do, who knows how much damage it will have caused.