(Side note: I'm not sure that's theoretically safe, because there could exist inputs for which the result isn't the same, but I've never actually found one of them.)

... and someone just pointed out that this rule is already there for reciprocalExact(c) values, so possibly I just happened to hit a case where it wasn't exact. I'm not going to close this in case there's a nuance I missed, but if that's already working-as-intended, feel free to close it.

Actually, thinking about it more: I believe there exist values such that reciprocalExact(c) is false, but where there's no input float32 value for which multiplication by the reciprocal produces a different result from division by the original value. Hmm. (... no wait, there almost certainly aren't. yay for math.)

Yes, I think this is unlikely to ever work. Closing

func main() {
	for _, d := range []float64{3, 5, 7, 9, 30} {
		r := 1 / d
		for _, x := range []float64{3, 5, 6, 7, 9} {
			if x*r != x/d {
				fmt.Printf("%v %v %v %v\n", d, x, x*r, x/d)
			}
		}
	}
}

Produces

3 5 1.6666666666666665 1.6666666666666667
3 7 2.333333333333333 2.3333333333333335
5 3 0.6000000000000001 0.6
5 6 1.2000000000000002 1.2
5 7 1.4000000000000001 1.4
7 5 0.7142857142857142 0.7142857142857143
7 9 1.2857142857142856 1.2857142857142858
9 7 0.7777777777777777 0.7777777777777778