You mention "the margin of error on these sub-samples is pretty large" but do not provide the margins of error. That would be useful knowledge and highly relevant for interpreting the data.
You could also clarify the value of these results by providing the n, the subset of N respondents (N = total sample size) for each cell. This is related to the issue above and to another comment in your post, that other polls often don't present such crosstabs. One reason they don't is that margins of error quickly become huge in the cells, and thus few statisticians feel quite comfortable attributing meaning to the numbers. The smaller the number of respondents in any given cell, the more likely that margins of error are swamping any apparent variation across sample subsets. (and vice versa)
Very astute comment, but IMO doesn't go far enough. What if one thinks the country is headed in the wrong direction because of pork-barrel politics, politicization of Supreme Court nominations, and lack of will to protect the country from further terrorist attacks? According to the interpretation presented here, and common elsewhere, a person who answers "wrong track" for those reasons is counted as "unfavorable" with respect to Bush. Most of those positions could be either bipartisan or blaming one party more than the other, and certainly the third would be less likely to come from a person who would prefer to see Ted Kennedy in charge.
While this poll may be in the mainstream in its choice of interpretive framework, it is precisely the use of the data to attempt to shape public opinion that leads over time to invalid data. Validity in this context means measuring what you intend to measure. When the data is collected in this fashion and overinterpreted, self-aware respondents will begin to answers pollsters strategically: They will not provide their true opinion if they know their response will be used to argue that "people" have some other opinion. They will answer in terms of the way they expect the data to be interpreted. Accurate measurement becomes impossible.
There was a heyday of powerful predictiveness in US opinion polls, but that day I believe may be past.