Hardest is a somewhat variable term.
From where I sit for the SNES I would probably want to look at Treasure of the Rudras/Rudra no Hihou/Rudora no Hiho.
The game used a Japanese linguistics based spell system. If I am going for easy translations I will take 50 different tables, table swap commands, compression, formatting/markup and some insane pointer system all day long before I want to tangle with having to refactor a linguistics based system as part of a translation.
No whether because I don't know Japanese that well, and indeed for the other languages I do know more well I am hardly going to be the toast of the evening should I attend a witty puns contest, linguistics would be where I look to find more challenging translations. While my misgivings might be a thing I would also have to look to how many games have had the text extracted but then "await translation or translators", and will be in 10 years time.
Having never tangled with Romancing Saga (or much SNES translation at all actually) I opted to read http://www.romhacking.net/documents/432/
Respectively they cover on map town names and the variable width font (though at the same time expanding a bit on the more typical in game text) for the first game. Impressive work on both counts, if someone wanted a rough and ready of example of how assembly hacking goes down in the real world they could do far worse than read that, and very much welcome for the end hack, however I am not sure (as in genuinely don't know, not just open pondering of whether it was a "needless" flourish) if a more bodge it and scarper method could not have still made the game playable.
The wiki did not have anything I could see and the VFW document for the second game was just that and no particular notes. To that end I don't know if the later games upped the ante in some way. The second game VFW looked like a fairly standard lookup and bounds checking affair though, though if you are subverting an entire system you might well be able to simplify things.