Hi Ian, thanks for taking an interest. This really only matters if source is dirty (but then again why else dedupe an overwrite). Here's the issue I had when I wrote this: a company I work with exports data from one of their software and pushes it in s3 as csv. Somehow they messed up and squashed two exports together, meaning one of my files was almost entirely made of deduplicates. There's an update datetime in these files, and they're even formatted correctly so they can be used as strings. But I couldn't use the update datetime because I couldn't select the cursor. I assume it would have used the source-defined cursor, which is the file last modified datetime and therefore wouldn't work. I ended up deleting the weird file and requesting a new export.
I only checked Postgres out of curiosity, and came to find out the option was completely unusable there, as there's no way to know which cursor will be used. It's hard to get dirty data in Postgres, but still, if the option exists at all it should be usable.