First problem in what you are doing:
p2pool and the farm located remotely from each other
Bad idea. P2pool is extremely dependent upon latency due to the excessive restart rate.
Not really. The latency between the sites is up to 30ms, which causes only about 0.2% stales by itself. However, it would increase non-linearly with addition of units, since each of them receives several kilobytes (because of large coinbase) simultaneously at work restart.
Getwork will use way more bandwidth than stratum.
The trick I am trying to do is increasing rollntime to say 1000, which means 1 getwork request per 20s per unit. Getwork response fits within 1 packet, while stratum response from p2pool is several kB. Please correct me if I am wrong.
The Getwork in my S1 binary is the Getwork in cgminer - same thing - I'm part of the cgminer team, not a fork.
I may be wrong but it seems like it has something to do with the endian issue. It was fixed at some time before your last build. In the stock binary (I mean that old one, supplied as a part of S1 ROM), getwork sends wrong nonce, while in your last build, it sends empty responses (does not submit solutions). Maybe you just didn't test getwork?
Your setup is the fault.
Add a proxy if you wish to reduce bandwidth.
Moving away from stratum would only be due to you misunderstanding how the protocols work.
Which proxy do you suggest? The stratum-mining-proxy by slush?
Cross compiling is the same as for the avalon.
Read their info on how to do that ... or the abundant info provided by openwrt.
Thank you!