The XMega board variant arrived and the prototype built. Converting the firmware over to the XMega architecture was a tedious exercise, but the result seems to work pretty well. At boot the board sets itself up to clock from the 32 MHz RC oscillator brought down to 30 MHz. The DFLL uses the internal 32 kHz RC oscillator to tune it. Once the FE lock pin is asserted, the PLL is set up to triple the 10 MHz input and switch over to using that. This neatly avoids the whole problem of the clock instability causing controller lock-ups and malfunctions prior to physics lock.
A side effect of this is that programming now no longer strictly requires the external oscillator be fed. PDI programming can take place regardless. Of course, that means having to use a PDI programmer, which is marginally less convenient for me than SPI programming.
If there's a real downside it's that I've discovered the chink in the XMega's armor: the ADC is terrible. I've had to configure it for differential mode with the negative pin connected to ground internally. The result is effectively an 11 bit ADC instead of 12 bits, and even then it's so noisy that throwing away another bit to make the output compatible with the ATMega ADC doesn't matter - it still jumps around 2-3 bits worth even then. For our purposes, it isn't too harmful, though, as the phase discriminator output gets averaged over the course of many seconds. It's bad enough, though, that I may wind up adding another conversion to the capture ISR to get a two-sample average each second.
I designed an OH300/ECS2522 board variant with the XMega too. It's not strictly necessary to convert that design, but the benefit would be the possibility of unifying the firmware across both platforms. The only real difference between them would be changing the serial oscillator interface into an SPI interface to drive the external DAC (the XMega has a 12 bit DAC, but that's not nearly enough granularity). Other than that, all that changes are some constant values relating to the EFC slope.