That's... negative feedback, not positive feedback.
Just flip around which side it's on, or take feedback from after the inverter, it'll be doing alright then.
Note that hysteretic control has the downside of no frequency control; if under some combination of conditions, the comparator simply decides to go flat out ~MHz blasting, well, there's nothing your gate driver or transistor can do about it (well, until they melt, that is). Conversely, frequency drops sharply at the extremes of operation (here, low current; for a voltage converter, at very low or very high duty, i.e., Vout near 0 or Vin), which increases output ripple; it's harder to design an output filter (or input, for that matter) to cover that. (And again, not that that's necessarily a problem for the LED; more generally, about the control method for various applications, I mean.)
Indeed, frequency response is critical for the sensor; if you don't have enough, consider reducing loop bandwidth by using a PWM modulator and current error amp, instead of hysteresis. Such an
average current mode control can get away with lower ripple fraction, as well as loop bandwidth (roughly speaking, the two go hand in hand). So, it's good for high power applications in CCM, but probably not so practical for a little 5V (input) LED like shown here.
Tim