I can't say I disagree, but I don't think this is a serious impediment. Though Sass and CoffeeScript have recently pushed this sort of referential opacity into CSS and JavaScript, it's been an "issue" in HTML markup for decades: As soon as you introduce any kind of templating or conditional logic, you're handing the browser generated content that doesn't map directly to line numbers of files on disk.
In that vein, how is Haml different from ERB or Jinja2? The latter may superficially resemble HTML, but they aren't HTML. An even stronger parallel could be drawn from Less or Sass's scss syntax.
I mean, hear what you mean but I think the level of annoyance or love for these tools depends on your workflow. Personally, I use LESS along with CodeKit (I used to use LESS.app) and I usually don't have those problems. The times that I do run into a line number reference problem I'm usually able to find the offending code quite easily as the line in my abstraction language file (LESS, Coffeescript, whatever) is usually only a few lines lower than what the debugger tells me. In addition, if the debugging tools give me just a short snippet of the offending code then I can pretty easily remember where it is regardless of how large the file is.
In the end though I think this is all about a person's workflow. Some workflows are better suited to these abstraction languages and their compilers/debuggers than others. Personal preference, I guess, is what it comes down to.