Via a dirty patch to LibAPRS that removes or emulates all Arduino-specific APIs, I was able to get packets decoding correctly.

The adc1_get_raw is the most straightforward way to sample from the ADC, but is not ideal:

• A task that repeatedly spins on the CPU until another 1/9600th of a second has passed (the way many Arduino sketches work) would monopolize the CPU too much to work well in the concurrent FreeRTOS environment on the ESP32.
• 9600 timer interrupts per second would be a lot of interrupts, and unless we dedicate one of the ESP32's two CPUs to the sampling task  we would not be guaranteed access to the CPU in time
• Reading from the ADC this way takes ~1/6000th of a second.

The datasheet mentions that the ADC can run at 200ksps when controlled by the RTC controller (I'm not sure what the DIG controller is). What gives?

Well, it turns out you can instruct the i2s peripheral to sample from the internal ADC, and write samples into a buffer using DMA.

In my patched LibAPRS, I disable the interrupt code that executes every 9600th of a second, and then in my application I call AFSK_adc_isr repeatedly, once we have a buffer full of audio.

This works well, though it does mean I don't process packets until my radio's squelch closes. I think I may be able to refactor this to process smaller buffers as they come in.