Well, it's another voltage drop and if Kerry runs this in a peaky configuration that yields 1A of LED current at a specific speed, it's more than the forward voltage of a rectifier diode so that it makes sense to shunt the resistor with a diode. Still, it induces the drop of one diode that I'd rather avoid.
The low pass filter is already there, it consists of the 10uF capacitor and the impedance represented by the reference voltage network on the inverting input of the OpAmp. This voltage references to the positive supply rail, a bit uncommon. More on the LPF further down.
At the point where the full wave mode current is higher than the doubler current, the doubler current is not increasing much any more. Go just a little further, add a little effect of temperature (dynamo getting warm = less current), you will not reach the switching point any more. Then, different models of dynamos differ sufficiently for the design not to work any more with some of them. I followed this approach first but learned that it's not reliable.
My next approach was switching a little earlier and using some hysteresis to prevent coming back instantly. It works, but the switching is obvious to the eye as the hysteresis couldn't be too small.
It took me quite a while to find a way out of this dilemma. This is how the circuit works today:
Around the cross-over point, there's rapid changing between the two modes. First it's just a very short time of full-wave rectifier mode, but then its duty cycle increases until its full-wave rectifier mode all the way. There is NO hysteresis in the circuit. There is only ONE low pass filter (before the OpAmp and it intentionally lets the ripple current pass through, eliminates higher frequency oscillation only), the signal path to the MOSFET is not artificially band limited. Setting the cross-over point is not very critical. It's also not completely uncritical, though. If set too high, it will never go to the bridge rectifier mode. So it can't work properly when the tuning capacitance is chosen for a peaky response and there's a valley between the peaks. This is exactly what Kerry wants to do.
I went back to the frequency detection approach because it's overall less critical. It doesn't have such a nice transition b/w the two modes but will reliably switch at a predefined frequency. The user can set this frequency without constraints and independently play with the tuning and smoothing capacitance. It also gets around rapid switching of the MOSFETs and eventually has no sense resistor to drop anything.
