The root of the problem is that I2C has serious EMC immunity issues. It's well known and appreciated that its drive is weak and open-drain. The bus buffers and especially differential drivers can and do help there. (And will get you past a meter.) What's less well recognized is that a single glitch pulse on SCK knocks all the internal state machines out of whack and requires a bus reset to fix them. Hope you're doing that when your bus is idle or when you're getting anomalies! Most people don't. The Nexus 4 phones sure didn't; this is why their light sensors went dead or crazy or both after a while of uptime.
All of that gets easier to handle if there's a nice, big, low-impedance ground plane nearby, which is why you don't see so much trouble when it stays on the PCB.
I2C is terrible in so many ways. (Don't send it off-board. Just don't. Ever. Trust me on this one. I don't care what you read about the bus capacitance spec, do. not. do. this.) Sending out a clock on a 400pF on-card bus isn't too bad, but when you have a kilometer-long cable... yeah, we'd rather not send a clock down that, thank you. Hence the use of self-clocked or pray-we-have-similar-clock-frequencies protocols.
3.3V or 5V devices aren't RS-232 and people need to be careful about that. They're just UARTs with regular old logic levels. RS-232 or 422 or 485 are Serious Business levels to go out and do battle with the mean Real World and the only place they should ever land is at a dedicated transceiver. Full stop.