We all act on a variety of social scripts, from which we find it enormously psychologically difficult to deviate once they're activated. Often, this is a problem, causing people to follow orders when they shouldn't, or to stand by in a crisis when they ought to be rendering aid. Sometimes it has more benign effects. The loss of the shoe suddenly shunted the policeman from his cop-script to his chivalry-script. Something similar probably happened with the burglar.
The crucial move here is that he wasn't directly challenged within the terms of the burglar script: The guests didn't refuse outright to hand over their money, but only offered him some wine first, so there was no need to move on to Act II: further threats. But since he hadn't asked for wine, the offer was not in line with the compliant-hostage script either: It was a social courtesy. Once he'd accepted, he was reading from the party-guest script. And holding up one's host at a dinner party is simply not the done thing.
"Script-breaking" is a rather prominent part of feminist theory, for example, in Sharon Marcus' work on rape prevention. She argues that rape is "a series of steps and signals whose typical initial moments we can learn to recognize and whose final outcome we can learn to stave off." It's not as easy as it sounds--just as the burglar is operating within a script of criminality, the target will naturally want to lapse into the script of "victim", the series of moves that society recognizes as the proper way for victims to behave. However, the risk of script-breaking is that its unpredictable--since people don't do it that often (hence it being a break), we don't know what the result of our behavioral alchemy will be. As Sanchez indicates, the best hope might be trying to divert the narrative from one recognized situation (robbery) to another recognized situation (dinner party)--even if the recognition stems only from media, TV, or social mores rather than lived-experience. But this only mitigates, not eliminates, the danger. For example, if rape is primarily a function of the desire for power and domination, a case where the victim refuses to behave in a recognizably "victim-like" manner could disorient and befuddle her attacker, but it also could enrage and infuriate him, leading him to yet more sadistic violence. This problem would seem to be extant regardless of whether the alternate script is familiar or not (although familiarity does seem to be a precondition for successful diversion).
Diversion works because it presumes we all have access to a certain universal index of narrative experience that has roughly similar (or at least mutually intelligible) meanings across the board. Where this is not the case, script-breaking runs into problems. The script itself may not adequately describe the situation being faced (a "rape script" premised around a drunk guy at the bar may not be of help when the woman is facing rape from her husband or father), or the perpetrator of violence may not be familiar with the alternative script being proposed (if someone had no conception of what a dinner party meant, the guest's reaction might have felt like mockery or worse). However, conceptually speaking, looking at crime in terms of scripts, and seeking to derail it by noting chinks in the armor of the narrative, is a surprisingly effective way to combat and prevent crime. Who says post-structuralism never gave us any useful practices?