But it also makes me wonder why Nate Cohn so often treats his Times/Siena cross tabs as gospel if he understands this.
As he notes cross tabs aren't weighted in most polls. The topline of a poll is properly weighted, so trust the toplines, but cross-tabs can have all sorts of sampling issues.
This is why generally I defer to actual specific polling of subgroups. We have dedicated polls of young people (Harvard Youth Poll), Latino voters (Telemundo), black people, etc.
These dedicated polls will have large samples of the specific subgroup, weight it properly, and poll well (for example asking polling in Spanish for Latino voters).
Absent dedicated polls, you can aggregate the cross tabs of a bunch of polls, which can be useful for identifying trends.