Please bear with me and allow me some theoretical bloviation; or, if you like, you can just cut to the chase.
We saw that neither Aristotelian virtue-based ethics nor Kantian rule-based ethics give us an intuitively satisfying analysis of the issues of workplace ethics, and neither offers sufficient justification for acceptance contrary to our intuition. We need to abandon the whole idea that ethics in general is a matter of objective truth, and look at the issue in terms of subjective things like preferences, values and interests.
The first thing that jumps out is that there are actually four parties to these sorts of workplace romances: the supervisor, the subordinate, the company and the government. Each party brings a set of subjective interests to the issue, both generally and in specific cases.
The supervisor and the subordinate are both individual people. As such, they generally have some competing interests in this context: On the one hand, pretty much everyone likes sex. Many people would like to find a "soul mate", a sincere, deep and fulfilling personal relationship. On the other hand, people want to perform and keep their jobs. They want to choose their sexual and romantic partners; the subordinate especially would prefer for his supervisor not to use the power of her position to coerce an unwanted sexual relationship.
These subjective interests come into conflict because on the one hand, we're at work to work, not date. On the other hand, we spend half our waking life at work, and the workplace is in fact a pretty good context for evaluating potential sexual and romantic partners; many people have begun satisfying sexual and romantic relationships from their work.
The government has interests which are far too complicated to even summarize; it's sufficient to note that the United States government has passed laws assigning civil liability to companies which tolerate supervisor-subordinate sexual harassment.
The company, on the other hand, cares not a whit for its employees sexual or romantic issues; and a company (not being a real person) doesn't have sexual or romantic interests at all; a company doesn't even have a basis for empathy in this context. Companies do, however, have a strong interest in complying with the government's laws. But again, the company has a conflicting interest in keeping its employees happy.
In a subjective analysis, none of these conflicts are contradictions. Contradictions arise only when you assign mutually exclusive properties to the same thing; it's not a contradiction to observe that there are different things with different properties (i.e. a person can have two different subjective interests).
The contradiction problem crops up all over the place in objective theories of ethics: virtues are both objectively good and objectively bad; rules are rules and non-rules; good rules promote bad virtues and vice-versa. Subjectivism is incomplete by itself, but any subjective analysis at least gets off to a good start by simply assigning these conflicting properties to different subjective interests, rather than to the same "objective" virtue or rule.
The whole exercise, then, of analyzing the ethics of workplace romances becomes a relatively straightforward exercise of game theory. All the players have interests they seek to fulfill, and different moves which affect the fulfillment of those interests: enact this policy or that policy; approach or do not approach a supervisor romantically; accept or discourage a subordinate's approach, etc.
A side issue regarding the ethical nature of rules seem fairly apparent.
There are two kinds of rules: Rules which prohibit outcomes which people disapprove of directly, and rules which in effect prohibit outcomes which people don't disapprove of. Most everyone disapproves of stealing; the rule against stealing is just an extension of this disapproval. On the other hand, hardly anyone disapproves of supervisor-subordinate relationships per se.
The no supervisor-subordinate relationship rule really exists to prohibit sexual harassment, a side-effect of such relationships. In a sense it "over-legislates": Taken literally it prohibits both outcomes that people approve of as well as those people disapprove of. I think there are some good reasons for such over-legislation, but they're too complicated to get into right now.
In a subjectivist analysis, rules aren't "true" in any ethical sense. Rules are moves by both real people and abstract organizations to attempt to fulfill their own interests. In other words, we don't not steal because there's a rule against it; we create a rule against stealing to explicitly express our disapproval of stealing, usually through organizations like governments or companies. In this sense, the question, allowing exceptions--even arbitrary exceptions--isn't a contradiction (although it might be a poor strategy). A rule is just a move, which has counter-move of disobeying, which has its own counter-move of a punitive or tolerant response.
In all the theoretical
It depends on how much you value the possible outcomes (positively or negatively), and your estimate of the probability of each outcome occurring. If you think there's a high probability of getting something you value positively, such as "true love", and a low probability of getting something you value negatively, such as getting fired, then it's in your best interest to go for it. On the other hand, if the best outcome you expect is something you don't value very highly (perhaps one episode of mediocre sex) and the worst outcome is highly probable (he's going to complain when I dump him), then it's in your best interests to refrain.
But what about the rule? The rule doesn't have any ethical value per se. The company did not enact the rule because it's "ethically right"; it enacted the rule to protect its own interests, primarily its interests in making money and not being sued. The company's intent was not to protect the superior's interests, nor even the subordinate's interests; the intent was to protect its own interests.
(One can, however, apply the same sort of analysis to what sort of rules a company should create, and how it should enforce them: What are the values of the different outcomes? What is the probability they would occur? "Spending time and money creating and enforcing rules," is one of these outcomes; considerations of efficiency come into play.)
The rule is thus an indication of how the company will act and under what circumstances. You have to look at your own company's rules and policies to try and anticipate how they'll act under different circumstances. Some companies will invoke the rule only if there's a formal complaint; other companies will invoke the rule if there's even the slightest evidence of a relationship. The rule (and other rules and characteristic behavior) affects the analysis not by establishing (or even recognizing) any kind of ethical truth, but by determining the probability of different outcomes.
The advantage of the paradigm (meta-ethical subjective relativism) is that it captures quite a lot with game theory about how we actually behave in ways we tend to categorize as "ethical". Furthermore, it makes falsifiable predictions. Subjective relativist systems of government, such as democracy--which just allows people to negotiate how values are fulfilled, without making many a priori judgments about values per se--should prove more stable than either authoritarianism or some sort of process modeled on scientific institutions. Furthermore, we should be able to infer people's and organizations' interests from a game theoretical analysis of their behavior, and those values should match the values we infer from talking to them; if the values don't match, we should be able to directly attribute that mismatch to a substantial strategic value of lying.
Of course, the huge question that meta-ethical subjective relativism doesn't answer--at least not directly--is why we persist in what can be called under this paradigm the delusions of ethical objectivism. The jury is still out, but this question may well be answered by scientific psychology.
Aside from ignoring a whole class of meta-ethical intuitions, meta-ethical subjective relativism is on pretty solid philosophical ground. It depends only on the premises that people have subjective interests, they act to fulfill them, and that the fulfillment of different interests (between different interests within one person's mind, and between different people) can come into conflict; all of these premises have good scientific justification. The paradigm has good explanatory power, and the theories and hypotheses which comprise it are falsifiable. It matches many of our intuitions about specific ethical situations, and gives us good reasons for abandoning those intuitions which conflict.
 The reification of abstract, conceptual entities such as companies and governments and the fictional attribution of subjective values to these entities is benign in this case. We can reduce an abstract company or government to a collection of real people with real values. And since real people create these abstractions precisely for the purpose of explicitly privileging some set of values in a particular context, we have an authoritative reference for what values we ascribe to the abstractions. So we can discuss abstractions like companies and governments using the same sort of language we use to discuss real concrete people with real subjective values.
 Technically, probabilistic game theory; the outcome of many of the moves can't be definitely known in advance. One cannot know in advance if accepting a subordinate's approach would result in "true love", but people can and do estimate the probability of such an outcome.