Yes, it is indeed the MPL3115 and another look at the data sheet left me no wiser about interpretation of the "temperature compensation". However I found a really good app note at:
https://www.nxp.com/files-static/sensors/doc/app_note/AN4528.pdfwhich explains it really well. It seems the compensation is for the internal mechanics of the MEMS sensor, to ensure an accurate pressure measurement over a wide range of temperature, it's not to handle atmospheric deviation from ISA. The altitude readout is therefore an "indicated altitude" with reference to 1013.2mb and assuming 15 degC. Considering the intended applications for the device, it seems this raw value would be more useful than a compensated one in most cases (e.g. if you wanted to make an actual altimeter with it, you'd need to undo the temperature compensation to get the same indicated altitude that everybody else is using).
This interpretation also gives the closest match to what I saw in flight. If the readout was compensated (true) altitude then (after adjusting for QNH) it would be even further in disagreement with the GPS and (nearly identical) altimeter indicated altitude.
That still leaves the matter of the increasing error with height on my unit. It's possible to individually calibrate the sensors, overriding the factory values, but I doubt PAW does this, nor would I have thought it was necessary to do so to get much better accuracy than what I am seeing. It may yet turn out to be cabin pressure, I'll investigate further.
Alan