However, Winograd himself concluded that language understanding algorithms plus emulated bodies plus emulated worlds aren't sufficient to achieve natural language understanding.

Ok, but is this the correct conclusion ? It's pretty obvious that a SHRDLU-style simulation is not sufficient to achieve natural language understanding, but can you generalize that to saying that no conceivable simulation is sufficient ? As far as I can tell, you would make such a generalization because,

Every emulation necessarily makes simplifying assumptions about both the world and the body that are subject to errors, bugs, and munchkin effects.

While this is true, it is also true that our human senses cannot fully perceive the reality around us with infinite fidelity. A child who is still learning his native tongue can't a rock that is 5cm in diameter from a rock that's 5.000001cm in diameter. This would lead me to believe that your simulation does not need 7 significant figures of precision in order to produce a language-speaking mind.

In fact, a colorblind child can't tell a red-colored ball from a green-colored ball, and yet colorblind adults can speak a variety of languages, so it's possible that your simulation could be monochrome and still achieve the desired result.