All the arguments about profitability aside, how would one actually go about implementing it in a sensible way? what's the compatibility list? How much memory do we actually need to calculate the hash, and how do we send/receive it efficiently?
All these questions are good ones to answer in an arduino implementation, and could have efficiency consequences for custom boards/IC's afterward.
Looking at the numbers, I don't think anyone is going to find out anytime soon

The fact that the hardware breakeven is longer than the duration of life for the device, and that while interesting to get a C/lua/python/ASM implementation running in single-core or multi-core configurations, the lessons would provide too little for benefits on other platforms. My suggestion is to start with the public code in public repositories on all the miners published so far and attempt to rewrite parts of them for the arduino.
Might as well start directly on those other platforms. I'm going to watch for an android implementation now, surely that's more accessible and powerful than any arduino implementation might go.