If things in Go aren't as clear-cut as the classic utilitarian example of "teleporting into the present situation" (which is typically the way Go programs are written, and they nevertheless lose to top human players), then maybe we can expect that they aren't clear-cut in complex life situations either.

That's more a fact about Go programs, I think; reading the Riis material recently on the Rybka case, I had the strong impression that modern top-tier chess programs do not do anything at all like building a model or examining the game history, but instead do very fine-tuned evaluations of individual board positions as they evaluate plys deep into the game tree. So you could teleport a copy of Shredder into a game against Kramnik played up to that point by Shredder, and expect the performance to be identical.

(If there were any research on sunk cost in Go, I'd expect it to follow the learning pattern: high initially followed by steady decline with feedback. I looked in Google Scholar for '("wei qi" OR "weiqi" OR "wei-chi" OR "igo" OR "baduk" OR "baeduk") "sunk cost" game' but didn't turn up anything. GS doesn't respect capitalization so "Go" is useless to search for.)