Went for a spin last week in ~4°C ambient conditions and the light worked pretty well. Actual light output is probably bright enough (though of course I want more!), and I'm pleased to report that the thermal back-off worked. After about 5 mins on full it did indeed get hot enough to trigger the Atmel to switch to 'low' mode rather than catch fire...which was nice. But disappointed that it got that hot at all.

Have now built a model in LTSpice which indicates that reducing the switching rate should bring the MOSFET dissipation down without affecting too many other things, so I've increased Coff to 133pF (100pF added to the 33p on the board) and now it switches at ~50kHz - about a third of the original rate. Under the thermal camera things are a bit better: the MOSFET is now relatively cool, and it's the Schottky (a Vishay SS34) that gets hot first. I need to sort this before I reduce the sense resistor further to squeeze more current through the LEDs...

I've also added a 4u7F X5R ceramic across the output terminals. This ought to make the LEDs perform better, by reducing the peak current without changing the mean. (For example, according to the XM-L datasheet, running them at 1A continuous should produce 12% more light than running 2A at 50% duty). Without a light meter I doubt I'll be able to tell the difference though.