Can you please squash the commits.

In a few hours. ...can't you use the "squash and merge" option?

This is a bad idea to document, because you lose source tracking at errors. Your errors with not include in what line the error happened and it's going to be very painful to debug.

It seems like it was designed to also accept configuration sets and not just paths. That should be documented and not just be a hidden feature? I can see your point that it makes things harder to debug but should that not be the users choice in the end?
I personally find that feature quite useful for defining nixops deployments, where machines share some configurations but not so much that I would want to put in a separate file.

It's a slippery slope, you'll soon regret saving those 5s once you're deep debugging some coercion gone wrong, without having the line number where it happened.

I think we should at very least make this clear (in red) in the documentation.

I agree it should be mentioned in the documentation.
However, from a technical point it is not clear to me why or when nix loses track of the line numbers. I ran into this type of problem many times without using the above mentioned construct. Is that mentioned somewhere in the documentation?

You only run into this if you use nixos modules imports without filenames.

When you misspell an option in a nixos configuration file the error message does not contain the line number (it does tell the filename though). So, this statement would then only apply to syntactical errors (where nix usually shows the line numbers)?
Btw.: I did just try importing a module by function and it actually does tell me the line number of a syntax error. I could not see any difference in behavior.
Under what circumstance would that debugging problem, you refer to, occur?

Maybe @edolstra can explain precisely, I don't have time to really research this, I'm just warning this is going to create more pain than gain as it currently stands if documented.

Well, this feature is used in lots of places, so documenting it makes sense, even if it's not really a good idea to use.

For instance, I'm using it to generate NixOS systems for a number of machines, most of which differ only in their hostname.

Is this a bad idea? Should I do it some other way? I don't want to generate configuration.nix files for each of them; I'd rather have the machines all defined in a single file, one line per machine.