I was scoping around to see what caused the 2.5V supply modules to go bust.
Just as a reminder, they are pretty generic modules, with a 5V preregulator that steps down the 12V input. The 5V is then dropped through linear regulators to three supply rails of 1.8V, 2.5V and 3.3V. In this application only the 2.5V output is being used.
I added a dummy load (6W car bulb) to simulate a 1A resistive load.
A few seconds later, the output began to oscillate at 50-150kHz and the supplied current dropped to ~200mA while the supplied voltage oscillated between 2V and 3.8V. The LDO (AIC1084) got pretty hot as well. Even in stable operation the peak-to-peak noise was 400mV a 16% [de]regulation, well above the normal spec of 0.5-1.3%.
I touched some parts of the circuit with the bare hand and the oscillations stopped, only to start again. I suspected some bad soldering or faulty caps, so proceeded to check the traces.
Since I was using a resistive load the processing board load variations were ruled out.
In the picture below the output cap is at the bottom right, rated at 16V and 1000uF.
Well, it turned out the cap was wrongly placed on the 1.8V rail instead of the 2.5V one, it should actually be placed on the F5 spot.
Lo and behold, resoldering it to the correct spot yielded correct operation with less than 100mV noise. I left it running for 24h to see how it behaves. The 1A load should prove plenty of challenge instead of the nominal 0.7A drawn by the processing board.