You need to deliver Vbatt > Vload + some amount of voltage headroom. The headroom is at least 0.12V, and probably a bit more since real-world lights have some parasitic resistances in the circuit from various contact resistances. For an XM-L driven at full power (3A), you need to deliver at least 3.35V + 0.12V = 3.47V. Add in some reasonable amount of parasitic resistance, and Vbatt most likely ought to be closer to about 3.8V. For long run time in full regulation, you probably will need 4xNiMH in series. Unfortunately, you aren't going to get very good driver efficiency. I'd suggest sinking the driver to something to manage the waste heat that is going to be produced.
If you implement resistance fixes in your Mag and are lucky, maybe the parasitic resistance will be low enough that you might need a Vbatt of say 3.6V to stay in full regulation. Also, if you are lucky and you have an XM-L with a lower Vf than the datasheet spec (and Vf also drops as the LED heats up), maybe Vbatt can be 3.5V or even slightly lower. In that case, 3xNiMH could work reasonably well in terms of run time in regulatioin and better efficiency, and you could switch to a 3D Mag.
