The problem is the generator isn't a linear source once the non-linear load is hooked up. If the load was a resistor taking AC power, you'd be correct. But the load is a bridge/cap and diode string. The generator doesn't see any load (runs open circuit) until the bridge diodes forward conduct Then once the generator's output voltage forward bias the bridge it's clamped at the output voltage and dumps as much current as it can. When the output cap isn't charged up enough to forward bias the LEDs, then the load is a short circuit. Once the output cap exceeds the turn-on voltage of the LEDs the load is just a clamp. None of the lends itself to the compromised model that Norton/Thevin equivalence requires.

I've measured this in the lab but I just confirmed Martin's measurements:
http://pilom.com/BicycleElectronics/DynamoCircuits.htm

You can see that saturation at max power in many of the curves. confirmed in several German language websites:
http://fahrradzukunft.de/14/neue-nabendynamos-im-test/

The limiting technique of adding in an extra winding is how they do it with 50/60Hz AC transformer off-line battery chargers, I assume that's how they do it in bike generators.